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

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128d          dummy_mask,cutoff_mask;
  96     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  97     __m128d          one     = _mm_set1_pd(1.0);
  98     __m128d          two     = _mm_set1_pd(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_pd(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112     krf              = _mm_set1_pd(fr->ic->k_rf);
 113     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 114     crf              = _mm_set1_pd(fr->ic->c_rf);
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     /* Setup water-specific parameters */
 120     inr              = nlist->iinr[0];
 121     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 122     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 123     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 124     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 125
 126     jq0              = _mm_set1_pd(charge[inr+0]);
 127     jq1              = _mm_set1_pd(charge[inr+1]);
 128     jq2              = _mm_set1_pd(charge[inr+2]);
 129     vdwjidx0A        = 2*vdwtype[inr+0];
 130     qq00             = _mm_mul_pd(iq0,jq0);
 131     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 132     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 133     qq01             = _mm_mul_pd(iq0,jq1);
 134     qq02             = _mm_mul_pd(iq0,jq2);
 135     qq10             = _mm_mul_pd(iq1,jq0);
 136     qq11             = _mm_mul_pd(iq1,jq1);
 137     qq12             = _mm_mul_pd(iq1,jq2);
 138     qq20             = _mm_mul_pd(iq2,jq0);
 139     qq21             = _mm_mul_pd(iq2,jq1);
 140     qq22             = _mm_mul_pd(iq2,jq2);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     /* Start outer loop over neighborlists */
 151     for(iidx=0; iidx<nri; iidx++)
 152     {
 153         /* Load shift vector for this list */
 154         i_shift_offset   = DIM*shiftidx[iidx];
 155
 156         /* Load limits for loop over neighbors */
 157         j_index_start    = jindex[iidx];
 158         j_index_end      = jindex[iidx+1];
 159
 160         /* Get outer coordinate index */
 161         inr              = iinr[iidx];
 162         i_coord_offset   = DIM*inr;
 163
 164         /* Load i particle coords and add shift vector */
 165         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 166                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 167
 168         fix0             = _mm_setzero_pd();
 169         fiy0             = _mm_setzero_pd();
 170         fiz0             = _mm_setzero_pd();
 171         fix1             = _mm_setzero_pd();
 172         fiy1             = _mm_setzero_pd();
 173         fiz1             = _mm_setzero_pd();
 174         fix2             = _mm_setzero_pd();
 175         fiy2             = _mm_setzero_pd();
 176         fiz2             = _mm_setzero_pd();
 177
 178         /* Reset potential sums */
 179         velecsum         = _mm_setzero_pd();
 180         vvdwsum          = _mm_setzero_pd();
 181
 182         /* Start inner kernel loop */
 183         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 184         {
 185
 186             /* Get j neighbor index, and coordinate index */
 187             jnrA             = jjnr[jidx];
 188             jnrB             = jjnr[jidx+1];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 194                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 195
 196             /* Calculate displacement vector */
 197             dx00             = _mm_sub_pd(ix0,jx0);
 198             dy00             = _mm_sub_pd(iy0,jy0);
 199             dz00             = _mm_sub_pd(iz0,jz0);
 200             dx01             = _mm_sub_pd(ix0,jx1);
 201             dy01             = _mm_sub_pd(iy0,jy1);
 202             dz01             = _mm_sub_pd(iz0,jz1);
 203             dx02             = _mm_sub_pd(ix0,jx2);
 204             dy02             = _mm_sub_pd(iy0,jy2);
 205             dz02             = _mm_sub_pd(iz0,jz2);
 206             dx10             = _mm_sub_pd(ix1,jx0);
 207             dy10             = _mm_sub_pd(iy1,jy0);
 208             dz10             = _mm_sub_pd(iz1,jz0);
 209             dx11             = _mm_sub_pd(ix1,jx1);
 210             dy11             = _mm_sub_pd(iy1,jy1);
 211             dz11             = _mm_sub_pd(iz1,jz1);
 212             dx12             = _mm_sub_pd(ix1,jx2);
 213             dy12             = _mm_sub_pd(iy1,jy2);
 214             dz12             = _mm_sub_pd(iz1,jz2);
 215             dx20             = _mm_sub_pd(ix2,jx0);
 216             dy20             = _mm_sub_pd(iy2,jy0);
 217             dz20             = _mm_sub_pd(iz2,jz0);
 218             dx21             = _mm_sub_pd(ix2,jx1);
 219             dy21             = _mm_sub_pd(iy2,jy1);
 220             dz21             = _mm_sub_pd(iz2,jz1);
 221             dx22             = _mm_sub_pd(ix2,jx2);
 222             dy22             = _mm_sub_pd(iy2,jy2);
 223             dz22             = _mm_sub_pd(iz2,jz2);
 224
 225             /* Calculate squared distance and things based on it */
 226             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 227             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 228             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 229             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 230             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 231             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 232             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 233             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 234             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 235
 236             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 237             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 238             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 239             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 240             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 241             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 242             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 243             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 244             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 245
 246             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 247             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 248             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 249             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 250             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 251             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 252             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 253             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 254             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 255
 256             fjx0             = _mm_setzero_pd();
 257             fjy0             = _mm_setzero_pd();
 258             fjz0             = _mm_setzero_pd();
 259             fjx1             = _mm_setzero_pd();
 260             fjy1             = _mm_setzero_pd();
 261             fjz1             = _mm_setzero_pd();
 262             fjx2             = _mm_setzero_pd();
 263             fjy2             = _mm_setzero_pd();
 264             fjz2             = _mm_setzero_pd();
 265
 266             /**************************
 267              * CALCULATE INTERACTIONS *
 268              **************************/
 269
 270             /* REACTION-FIELD ELECTROSTATICS */
 271             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 272             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 273
 274             /* LENNARD-JONES DISPERSION/REPULSION */
 275
 276             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 277             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 278             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 279             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 280             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 281
 282             /* Update potential sum for this i atom from the interaction with this j atom. */
 283             velecsum         = _mm_add_pd(velecsum,velec);
 284             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 285
 286             fscal            = _mm_add_pd(felec,fvdw);
 287
 288             /* Update vectorial force */
 289             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 290             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 291             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 292
 293             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 294             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 295             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             /* REACTION-FIELD ELECTROSTATICS */
 302             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 303             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velecsum         = _mm_add_pd(velecsum,velec);
 307
 308             fscal            = felec;
 309
 310             /* Update vectorial force */
 311             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 312             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 313             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 314
 315             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 316             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 317             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             /* REACTION-FIELD ELECTROSTATICS */
 324             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 325             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 326
 327             /* Update potential sum for this i atom from the interaction with this j atom. */
 328             velecsum         = _mm_add_pd(velecsum,velec);
 329
 330             fscal            = felec;
 331
 332             /* Update vectorial force */
 333             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 334             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 335             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 336
 337             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 338             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 339             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 340
 341             /**************************
 342              * CALCULATE INTERACTIONS *
 343              **************************/
 344
 345             /* REACTION-FIELD ELECTROSTATICS */
 346             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 347             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             velecsum         = _mm_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             /* Update vectorial force */
 355             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 356             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 357             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 358
 359             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 360             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 361             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             /* REACTION-FIELD ELECTROSTATICS */
 368             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 369             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velecsum         = _mm_add_pd(velecsum,velec);
 373
 374             fscal            = felec;
 375
 376             /* Update vectorial force */
 377             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 378             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 379             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 380
 381             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 382             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 383             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 384
 385             /**************************
 386              * CALCULATE INTERACTIONS *
 387              **************************/
 388
 389             /* REACTION-FIELD ELECTROSTATICS */
 390             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 391             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velecsum         = _mm_add_pd(velecsum,velec);
 395
 396             fscal            = felec;
 397
 398             /* Update vectorial force */
 399             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 400             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 401             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 402
 403             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 404             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 405             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             /* REACTION-FIELD ELECTROSTATICS */
 412             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 413             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 414
 415             /* Update potential sum for this i atom from the interaction with this j atom. */
 416             velecsum         = _mm_add_pd(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             /* Update vectorial force */
 421             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 422             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 423             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 424
 425             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 426             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 427             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             /* REACTION-FIELD ELECTROSTATICS */
 434             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 435             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 436
 437             /* Update potential sum for this i atom from the interaction with this j atom. */
 438             velecsum         = _mm_add_pd(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             /* Update vectorial force */
 443             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 444             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 445             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 446
 447             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 448             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 449             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             /* REACTION-FIELD ELECTROSTATICS */
 456             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 457             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             velecsum         = _mm_add_pd(velecsum,velec);
 461
 462             fscal            = felec;
 463
 464             /* Update vectorial force */
 465             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 466             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 467             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 468
 469             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 470             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 471             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 472
 473             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 474
 475             /* Inner loop uses 327 flops */
 476         }
 477
 478         if(jidx<j_index_end)
 479         {
 480
 481             jnrA             = jjnr[jidx];
 482             j_coord_offsetA  = DIM*jnrA;
 483
 484             /* load j atom coordinates */
 485             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 486                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 487
 488             /* Calculate displacement vector */
 489             dx00             = _mm_sub_pd(ix0,jx0);
 490             dy00             = _mm_sub_pd(iy0,jy0);
 491             dz00             = _mm_sub_pd(iz0,jz0);
 492             dx01             = _mm_sub_pd(ix0,jx1);
 493             dy01             = _mm_sub_pd(iy0,jy1);
 494             dz01             = _mm_sub_pd(iz0,jz1);
 495             dx02             = _mm_sub_pd(ix0,jx2);
 496             dy02             = _mm_sub_pd(iy0,jy2);
 497             dz02             = _mm_sub_pd(iz0,jz2);
 498             dx10             = _mm_sub_pd(ix1,jx0);
 499             dy10             = _mm_sub_pd(iy1,jy0);
 500             dz10             = _mm_sub_pd(iz1,jz0);
 501             dx11             = _mm_sub_pd(ix1,jx1);
 502             dy11             = _mm_sub_pd(iy1,jy1);
 503             dz11             = _mm_sub_pd(iz1,jz1);
 504             dx12             = _mm_sub_pd(ix1,jx2);
 505             dy12             = _mm_sub_pd(iy1,jy2);
 506             dz12             = _mm_sub_pd(iz1,jz2);
 507             dx20             = _mm_sub_pd(ix2,jx0);
 508             dy20             = _mm_sub_pd(iy2,jy0);
 509             dz20             = _mm_sub_pd(iz2,jz0);
 510             dx21             = _mm_sub_pd(ix2,jx1);
 511             dy21             = _mm_sub_pd(iy2,jy1);
 512             dz21             = _mm_sub_pd(iz2,jz1);
 513             dx22             = _mm_sub_pd(ix2,jx2);
 514             dy22             = _mm_sub_pd(iy2,jy2);
 515             dz22             = _mm_sub_pd(iz2,jz2);
 516
 517             /* Calculate squared distance and things based on it */
 518             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 519             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 520             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 521             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 522             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 523             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 524             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 525             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 526             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 527
 528             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 529             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 530             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 531             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 532             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 533             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 534             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 535             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 536             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 537
 538             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 539             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 540             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 541             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 542             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 543             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 544             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 545             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 546             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 547
 548             fjx0             = _mm_setzero_pd();
 549             fjy0             = _mm_setzero_pd();
 550             fjz0             = _mm_setzero_pd();
 551             fjx1             = _mm_setzero_pd();
 552             fjy1             = _mm_setzero_pd();
 553             fjz1             = _mm_setzero_pd();
 554             fjx2             = _mm_setzero_pd();
 555             fjy2             = _mm_setzero_pd();
 556             fjz2             = _mm_setzero_pd();
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             /* REACTION-FIELD ELECTROSTATICS */
 563             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 564             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 565
 566             /* LENNARD-JONES DISPERSION/REPULSION */
 567
 568             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 569             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 570             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 571             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 572             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 573
 574             /* Update potential sum for this i atom from the interaction with this j atom. */
 575             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 576             velecsum         = _mm_add_pd(velecsum,velec);
 577             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 578             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 579
 580             fscal            = _mm_add_pd(felec,fvdw);
 581
 582             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 583
 584             /* Update vectorial force */
 585             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 586             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 587             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 588
 589             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 590             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 591             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 592
 593             /**************************
 594              * CALCULATE INTERACTIONS *
 595              **************************/
 596
 597             /* REACTION-FIELD ELECTROSTATICS */
 598             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 599             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 600
 601             /* Update potential sum for this i atom from the interaction with this j atom. */
 602             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 603             velecsum         = _mm_add_pd(velecsum,velec);
 604
 605             fscal            = felec;
 606
 607             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 608
 609             /* Update vectorial force */
 610             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 611             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 612             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 613
 614             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 615             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 616             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 617
 618             /**************************
 619              * CALCULATE INTERACTIONS *
 620              **************************/
 621
 622             /* REACTION-FIELD ELECTROSTATICS */
 623             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 624             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 625
 626             /* Update potential sum for this i atom from the interaction with this j atom. */
 627             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 628             velecsum         = _mm_add_pd(velecsum,velec);
 629
 630             fscal            = felec;
 631
 632             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 633
 634             /* Update vectorial force */
 635             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 636             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 637             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 638
 639             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 640             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 641             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 642
 643             /**************************
 644              * CALCULATE INTERACTIONS *
 645              **************************/
 646
 647             /* REACTION-FIELD ELECTROSTATICS */
 648             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 649             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 650
 651             /* Update potential sum for this i atom from the interaction with this j atom. */
 652             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 653             velecsum         = _mm_add_pd(velecsum,velec);
 654
 655             fscal            = felec;
 656
 657             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 658
 659             /* Update vectorial force */
 660             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 661             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 662             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 663
 664             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 665             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 666             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             /* REACTION-FIELD ELECTROSTATICS */
 673             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 674             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 675
 676             /* Update potential sum for this i atom from the interaction with this j atom. */
 677             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 678             velecsum         = _mm_add_pd(velecsum,velec);
 679
 680             fscal            = felec;
 681
 682             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 683
 684             /* Update vectorial force */
 685             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 686             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 687             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 688
 689             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 690             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 691             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 692
 693             /**************************
 694              * CALCULATE INTERACTIONS *
 695              **************************/
 696
 697             /* REACTION-FIELD ELECTROSTATICS */
 698             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 699             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 700
 701             /* Update potential sum for this i atom from the interaction with this j atom. */
 702             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 703             velecsum         = _mm_add_pd(velecsum,velec);
 704
 705             fscal            = felec;
 706
 707             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 708
 709             /* Update vectorial force */
 710             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 711             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 712             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 713
 714             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 715             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 716             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 717
 718             /**************************
 719              * CALCULATE INTERACTIONS *
 720              **************************/
 721
 722             /* REACTION-FIELD ELECTROSTATICS */
 723             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 724             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 725
 726             /* Update potential sum for this i atom from the interaction with this j atom. */
 727             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 728             velecsum         = _mm_add_pd(velecsum,velec);
 729
 730             fscal            = felec;
 731
 732             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 733
 734             /* Update vectorial force */
 735             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 736             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 737             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 738
 739             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 740             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 741             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 742
 743             /**************************
 744              * CALCULATE INTERACTIONS *
 745              **************************/
 746
 747             /* REACTION-FIELD ELECTROSTATICS */
 748             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 749             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 750
 751             /* Update potential sum for this i atom from the interaction with this j atom. */
 752             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 753             velecsum         = _mm_add_pd(velecsum,velec);
 754
 755             fscal            = felec;
 756
 757             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 758
 759             /* Update vectorial force */
 760             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 761             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 762             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 763
 764             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 765             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 766             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* REACTION-FIELD ELECTROSTATICS */
 773             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 774             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 778             velecsum         = _mm_add_pd(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 783
 784             /* Update vectorial force */
 785             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 786             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 787             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 788
 789             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 790             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 791             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 792
 793             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 794
 795             /* Inner loop uses 327 flops */
 796         }
 797
 798         /* End of innermost loop */
 799
 800         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 801                                               f+i_coord_offset,fshift+i_shift_offset);
 802
 803         ggid                        = gid[iidx];
 804         /* Update potential energies */
 805         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 806         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 807
 808         /* Increment number of inner iterations */
 809         inneriter                  += j_index_end - j_index_start;
 810
 811         /* Outer loop uses 20 flops */
 812     }
 813
 814     /* Increment number of outer iterations */
 815     outeriter        += nri;
 816
 817     /* Update outer/inner flops */
 818
 819     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*327);
 820 }
 821 /*
 822  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_double
 823  * Electrostatics interaction: ReactionField
 824  * VdW interaction:            LennardJones
 825  * Geometry:                   Water3-Water3
 826  * Calculate force/pot:        Force
 827  */
 828 void
 829 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_avx_128_fma_double
 830                     (t_nblist * gmx_restrict                nlist,
 831                      rvec * gmx_restrict                    xx,
 832                      rvec * gmx_restrict                    ff,
 833                      t_forcerec * gmx_restrict              fr,
 834                      t_mdatoms * gmx_restrict               mdatoms,
 835                      nb_kernel_data_t * gmx_restrict        kernel_data,
 836                      t_nrnb * gmx_restrict                  nrnb)
 837 {
 838     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 839      * just 0 for non-waters.
 840      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 841      * jnr indices corresponding to data put in the four positions in the SIMD register.
 842      */
 843     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 844     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 845     int              jnrA,jnrB;
 846     int              j_coord_offsetA,j_coord_offsetB;
 847     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 848     real             rcutoff_scalar;
 849     real             *shiftvec,*fshift,*x,*f;
 850     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 851     int              vdwioffset0;
 852     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 853     int              vdwioffset1;
 854     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 855     int              vdwioffset2;
 856     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 857     int              vdwjidx0A,vdwjidx0B;
 858     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 859     int              vdwjidx1A,vdwjidx1B;
 860     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 861     int              vdwjidx2A,vdwjidx2B;
 862     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 863     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 864     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 865     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 866     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 867     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 868     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 869     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 870     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 871     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 872     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 873     real             *charge;
 874     int              nvdwtype;
 875     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 876     int              *vdwtype;
 877     real             *vdwparam;
 878     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 879     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 880     __m128d          dummy_mask,cutoff_mask;
 881     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 882     __m128d          one     = _mm_set1_pd(1.0);
 883     __m128d          two     = _mm_set1_pd(2.0);
 884     x                = xx[0];
 885     f                = ff[0];
 886
 887     nri              = nlist->nri;
 888     iinr             = nlist->iinr;
 889     jindex           = nlist->jindex;
 890     jjnr             = nlist->jjnr;
 891     shiftidx         = nlist->shift;
 892     gid              = nlist->gid;
 893     shiftvec         = fr->shift_vec[0];
 894     fshift           = fr->fshift[0];
 895     facel            = _mm_set1_pd(fr->epsfac);
 896     charge           = mdatoms->chargeA;
 897     krf              = _mm_set1_pd(fr->ic->k_rf);
 898     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 899     crf              = _mm_set1_pd(fr->ic->c_rf);
 900     nvdwtype         = fr->ntype;
 901     vdwparam         = fr->nbfp;
 902     vdwtype          = mdatoms->typeA;
 903
 904     /* Setup water-specific parameters */
 905     inr              = nlist->iinr[0];
 906     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 907     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 908     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 909     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 910
 911     jq0              = _mm_set1_pd(charge[inr+0]);
 912     jq1              = _mm_set1_pd(charge[inr+1]);
 913     jq2              = _mm_set1_pd(charge[inr+2]);
 914     vdwjidx0A        = 2*vdwtype[inr+0];
 915     qq00             = _mm_mul_pd(iq0,jq0);
 916     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 917     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 918     qq01             = _mm_mul_pd(iq0,jq1);
 919     qq02             = _mm_mul_pd(iq0,jq2);
 920     qq10             = _mm_mul_pd(iq1,jq0);
 921     qq11             = _mm_mul_pd(iq1,jq1);
 922     qq12             = _mm_mul_pd(iq1,jq2);
 923     qq20             = _mm_mul_pd(iq2,jq0);
 924     qq21             = _mm_mul_pd(iq2,jq1);
 925     qq22             = _mm_mul_pd(iq2,jq2);
 926
 927     /* Avoid stupid compiler warnings */
 928     jnrA = jnrB = 0;
 929     j_coord_offsetA = 0;
 930     j_coord_offsetB = 0;
 931
 932     outeriter        = 0;
 933     inneriter        = 0;
 934
 935     /* Start outer loop over neighborlists */
 936     for(iidx=0; iidx<nri; iidx++)
 937     {
 938         /* Load shift vector for this list */
 939         i_shift_offset   = DIM*shiftidx[iidx];
 940
 941         /* Load limits for loop over neighbors */
 942         j_index_start    = jindex[iidx];
 943         j_index_end      = jindex[iidx+1];
 944
 945         /* Get outer coordinate index */
 946         inr              = iinr[iidx];
 947         i_coord_offset   = DIM*inr;
 948
 949         /* Load i particle coords and add shift vector */
 950         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 951                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 952
 953         fix0             = _mm_setzero_pd();
 954         fiy0             = _mm_setzero_pd();
 955         fiz0             = _mm_setzero_pd();
 956         fix1             = _mm_setzero_pd();
 957         fiy1             = _mm_setzero_pd();
 958         fiz1             = _mm_setzero_pd();
 959         fix2             = _mm_setzero_pd();
 960         fiy2             = _mm_setzero_pd();
 961         fiz2             = _mm_setzero_pd();
 962
 963         /* Start inner kernel loop */
 964         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 965         {
 966
 967             /* Get j neighbor index, and coordinate index */
 968             jnrA             = jjnr[jidx];
 969             jnrB             = jjnr[jidx+1];
 970             j_coord_offsetA  = DIM*jnrA;
 971             j_coord_offsetB  = DIM*jnrB;
 972
 973             /* load j atom coordinates */
 974             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 975                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 976
 977             /* Calculate displacement vector */
 978             dx00             = _mm_sub_pd(ix0,jx0);
 979             dy00             = _mm_sub_pd(iy0,jy0);
 980             dz00             = _mm_sub_pd(iz0,jz0);
 981             dx01             = _mm_sub_pd(ix0,jx1);
 982             dy01             = _mm_sub_pd(iy0,jy1);
 983             dz01             = _mm_sub_pd(iz0,jz1);
 984             dx02             = _mm_sub_pd(ix0,jx2);
 985             dy02             = _mm_sub_pd(iy0,jy2);
 986             dz02             = _mm_sub_pd(iz0,jz2);
 987             dx10             = _mm_sub_pd(ix1,jx0);
 988             dy10             = _mm_sub_pd(iy1,jy0);
 989             dz10             = _mm_sub_pd(iz1,jz0);
 990             dx11             = _mm_sub_pd(ix1,jx1);
 991             dy11             = _mm_sub_pd(iy1,jy1);
 992             dz11             = _mm_sub_pd(iz1,jz1);
 993             dx12             = _mm_sub_pd(ix1,jx2);
 994             dy12             = _mm_sub_pd(iy1,jy2);
 995             dz12             = _mm_sub_pd(iz1,jz2);
 996             dx20             = _mm_sub_pd(ix2,jx0);
 997             dy20             = _mm_sub_pd(iy2,jy0);
 998             dz20             = _mm_sub_pd(iz2,jz0);
 999             dx21             = _mm_sub_pd(ix2,jx1);
1000             dy21             = _mm_sub_pd(iy2,jy1);
1001             dz21             = _mm_sub_pd(iz2,jz1);
1002             dx22             = _mm_sub_pd(ix2,jx2);
1003             dy22             = _mm_sub_pd(iy2,jy2);
1004             dz22             = _mm_sub_pd(iz2,jz2);
1005
1006             /* Calculate squared distance and things based on it */
1007             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1008             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1009             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1010             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1011             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1012             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1013             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1014             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1015             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1016
1017             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1018             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1019             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1020             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1021             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1022             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1023             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1024             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1025             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1026
1027             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1028             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1029             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1030             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1031             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1032             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1033             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1034             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1035             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1036
1037             fjx0             = _mm_setzero_pd();
1038             fjy0             = _mm_setzero_pd();
1039             fjz0             = _mm_setzero_pd();
1040             fjx1             = _mm_setzero_pd();
1041             fjy1             = _mm_setzero_pd();
1042             fjz1             = _mm_setzero_pd();
1043             fjx2             = _mm_setzero_pd();
1044             fjy2             = _mm_setzero_pd();
1045             fjz2             = _mm_setzero_pd();
1046
1047             /**************************
1048              * CALCULATE INTERACTIONS *
1049              **************************/
1050
1051             /* REACTION-FIELD ELECTROSTATICS */
1052             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1053
1054             /* LENNARD-JONES DISPERSION/REPULSION */
1055
1056             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1057             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1058
1059             fscal            = _mm_add_pd(felec,fvdw);
1060
1061             /* Update vectorial force */
1062             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1063             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1064             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1065
1066             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1067             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1068             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1069
1070             /**************************
1071              * CALCULATE INTERACTIONS *
1072              **************************/
1073
1074             /* REACTION-FIELD ELECTROSTATICS */
1075             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1076
1077             fscal            = felec;
1078
1079             /* Update vectorial force */
1080             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1081             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1082             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1083
1084             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1085             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1086             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1087
1088             /**************************
1089              * CALCULATE INTERACTIONS *
1090              **************************/
1091
1092             /* REACTION-FIELD ELECTROSTATICS */
1093             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1094
1095             fscal            = felec;
1096
1097             /* Update vectorial force */
1098             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1099             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1100             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1101
1102             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1103             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1104             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1105
1106             /**************************
1107              * CALCULATE INTERACTIONS *
1108              **************************/
1109
1110             /* REACTION-FIELD ELECTROSTATICS */
1111             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1112
1113             fscal            = felec;
1114
1115             /* Update vectorial force */
1116             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1117             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1118             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1119
1120             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1121             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1122             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1123
1124             /**************************
1125              * CALCULATE INTERACTIONS *
1126              **************************/
1127
1128             /* REACTION-FIELD ELECTROSTATICS */
1129             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1130
1131             fscal            = felec;
1132
1133             /* Update vectorial force */
1134             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1135             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1136             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1137
1138             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1139             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1140             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1141
1142             /**************************
1143              * CALCULATE INTERACTIONS *
1144              **************************/
1145
1146             /* REACTION-FIELD ELECTROSTATICS */
1147             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1148
1149             fscal            = felec;
1150
1151             /* Update vectorial force */
1152             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1153             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1154             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1155
1156             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1157             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1158             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             /* REACTION-FIELD ELECTROSTATICS */
1165             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1166
1167             fscal            = felec;
1168
1169             /* Update vectorial force */
1170             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1171             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1172             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1173
1174             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1175             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1176             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1177
1178             /**************************
1179              * CALCULATE INTERACTIONS *
1180              **************************/
1181
1182             /* REACTION-FIELD ELECTROSTATICS */
1183             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1184
1185             fscal            = felec;
1186
1187             /* Update vectorial force */
1188             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1189             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1190             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1191
1192             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1193             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1194             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1195
1196             /**************************
1197              * CALCULATE INTERACTIONS *
1198              **************************/
1199
1200             /* REACTION-FIELD ELECTROSTATICS */
1201             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1202
1203             fscal            = felec;
1204
1205             /* Update vectorial force */
1206             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1207             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1208             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1209
1210             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1211             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1212             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1213
1214             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1215
1216             /* Inner loop uses 277 flops */
1217         }
1218
1219         if(jidx<j_index_end)
1220         {
1221
1222             jnrA             = jjnr[jidx];
1223             j_coord_offsetA  = DIM*jnrA;
1224
1225             /* load j atom coordinates */
1226             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1227                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1228
1229             /* Calculate displacement vector */
1230             dx00             = _mm_sub_pd(ix0,jx0);
1231             dy00             = _mm_sub_pd(iy0,jy0);
1232             dz00             = _mm_sub_pd(iz0,jz0);
1233             dx01             = _mm_sub_pd(ix0,jx1);
1234             dy01             = _mm_sub_pd(iy0,jy1);
1235             dz01             = _mm_sub_pd(iz0,jz1);
1236             dx02             = _mm_sub_pd(ix0,jx2);
1237             dy02             = _mm_sub_pd(iy0,jy2);
1238             dz02             = _mm_sub_pd(iz0,jz2);
1239             dx10             = _mm_sub_pd(ix1,jx0);
1240             dy10             = _mm_sub_pd(iy1,jy0);
1241             dz10             = _mm_sub_pd(iz1,jz0);
1242             dx11             = _mm_sub_pd(ix1,jx1);
1243             dy11             = _mm_sub_pd(iy1,jy1);
1244             dz11             = _mm_sub_pd(iz1,jz1);
1245             dx12             = _mm_sub_pd(ix1,jx2);
1246             dy12             = _mm_sub_pd(iy1,jy2);
1247             dz12             = _mm_sub_pd(iz1,jz2);
1248             dx20             = _mm_sub_pd(ix2,jx0);
1249             dy20             = _mm_sub_pd(iy2,jy0);
1250             dz20             = _mm_sub_pd(iz2,jz0);
1251             dx21             = _mm_sub_pd(ix2,jx1);
1252             dy21             = _mm_sub_pd(iy2,jy1);
1253             dz21             = _mm_sub_pd(iz2,jz1);
1254             dx22             = _mm_sub_pd(ix2,jx2);
1255             dy22             = _mm_sub_pd(iy2,jy2);
1256             dz22             = _mm_sub_pd(iz2,jz2);
1257
1258             /* Calculate squared distance and things based on it */
1259             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1260             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1261             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1262             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1263             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1264             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1265             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1266             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1267             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1268
1269             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1270             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1271             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1272             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1273             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1274             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1275             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1276             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1277             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1278
1279             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1280             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1281             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1282             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1283             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1284             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1285             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1286             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1287             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1288
1289             fjx0             = _mm_setzero_pd();
1290             fjy0             = _mm_setzero_pd();
1291             fjz0             = _mm_setzero_pd();
1292             fjx1             = _mm_setzero_pd();
1293             fjy1             = _mm_setzero_pd();
1294             fjz1             = _mm_setzero_pd();
1295             fjx2             = _mm_setzero_pd();
1296             fjy2             = _mm_setzero_pd();
1297             fjz2             = _mm_setzero_pd();
1298
1299             /**************************
1300              * CALCULATE INTERACTIONS *
1301              **************************/
1302
1303             /* REACTION-FIELD ELECTROSTATICS */
1304             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1305
1306             /* LENNARD-JONES DISPERSION/REPULSION */
1307
1308             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1309             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1310
1311             fscal            = _mm_add_pd(felec,fvdw);
1312
1313             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1314
1315             /* Update vectorial force */
1316             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1317             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1318             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1319
1320             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1321             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1322             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1323
1324             /**************************
1325              * CALCULATE INTERACTIONS *
1326              **************************/
1327
1328             /* REACTION-FIELD ELECTROSTATICS */
1329             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1330
1331             fscal            = felec;
1332
1333             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1334
1335             /* Update vectorial force */
1336             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1337             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1338             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1339
1340             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1341             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1342             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             /* REACTION-FIELD ELECTROSTATICS */
1349             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1350
1351             fscal            = felec;
1352
1353             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1354
1355             /* Update vectorial force */
1356             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1357             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1358             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1359
1360             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1361             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1362             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1363
1364             /**************************
1365              * CALCULATE INTERACTIONS *
1366              **************************/
1367
1368             /* REACTION-FIELD ELECTROSTATICS */
1369             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1370
1371             fscal            = felec;
1372
1373             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1374
1375             /* Update vectorial force */
1376             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1377             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1378             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1379
1380             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1381             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1382             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             /* REACTION-FIELD ELECTROSTATICS */
1389             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1390
1391             fscal            = felec;
1392
1393             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1394
1395             /* Update vectorial force */
1396             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1397             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1398             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1399
1400             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1401             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1402             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* REACTION-FIELD ELECTROSTATICS */
1409             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1410
1411             fscal            = felec;
1412
1413             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1414
1415             /* Update vectorial force */
1416             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1417             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1418             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1419
1420             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1421             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1422             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1423
1424             /**************************
1425              * CALCULATE INTERACTIONS *
1426              **************************/
1427
1428             /* REACTION-FIELD ELECTROSTATICS */
1429             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1430
1431             fscal            = felec;
1432
1433             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1434
1435             /* Update vectorial force */
1436             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1437             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1438             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1439
1440             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1441             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1442             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1443
1444             /**************************
1445              * CALCULATE INTERACTIONS *
1446              **************************/
1447
1448             /* REACTION-FIELD ELECTROSTATICS */
1449             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1450
1451             fscal            = felec;
1452
1453             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1454
1455             /* Update vectorial force */
1456             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1457             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1458             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1459
1460             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1461             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1462             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             /* REACTION-FIELD ELECTROSTATICS */
1469             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1470
1471             fscal            = felec;
1472
1473             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1474
1475             /* Update vectorial force */
1476             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1477             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1478             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1479
1480             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1481             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1482             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1483
1484             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1485
1486             /* Inner loop uses 277 flops */
1487         }
1488
1489         /* End of innermost loop */
1490
1491         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1492                                               f+i_coord_offset,fshift+i_shift_offset);
1493
1494         /* Increment number of inner iterations */
1495         inneriter                  += j_index_end - j_index_start;
1496
1497         /* Outer loop uses 18 flops */
1498     }
1499
1500     /* Increment number of outer iterations */
1501     outeriter        += nri;
1502
1503     /* Update outer/inner flops */
1504
1505     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1506 }