src/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRF_VdwLJ_GeomW3W3_sse2_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_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_sse2_double.h"
  34 #include "kernelutil_x86_sse2_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_VF_sse2_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_sse2_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_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 272             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_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_sub_pd( _mm_mul_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             /* Calculate temporary vectorial force */
 289             tx               = _mm_mul_pd(fscal,dx00);
 290             ty               = _mm_mul_pd(fscal,dy00);
 291             tz               = _mm_mul_pd(fscal,dz00);
 292
 293             /* Update vectorial force */
 294             fix0             = _mm_add_pd(fix0,tx);
 295             fiy0             = _mm_add_pd(fiy0,ty);
 296             fiz0             = _mm_add_pd(fiz0,tz);
 297
 298             fjx0             = _mm_add_pd(fjx0,tx);
 299             fjy0             = _mm_add_pd(fjy0,ty);
 300             fjz0             = _mm_add_pd(fjz0,tz);
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             /* REACTION-FIELD ELECTROSTATICS */
 307             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 308             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_pd(velecsum,velec);
 312
 313             fscal            = felec;
 314
 315             /* Calculate temporary vectorial force */
 316             tx               = _mm_mul_pd(fscal,dx01);
 317             ty               = _mm_mul_pd(fscal,dy01);
 318             tz               = _mm_mul_pd(fscal,dz01);
 319
 320             /* Update vectorial force */
 321             fix0             = _mm_add_pd(fix0,tx);
 322             fiy0             = _mm_add_pd(fiy0,ty);
 323             fiz0             = _mm_add_pd(fiz0,tz);
 324
 325             fjx1             = _mm_add_pd(fjx1,tx);
 326             fjy1             = _mm_add_pd(fjy1,ty);
 327             fjz1             = _mm_add_pd(fjz1,tz);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             /* REACTION-FIELD ELECTROSTATICS */
 334             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 335             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 336
 337             /* Update potential sum for this i atom from the interaction with this j atom. */
 338             velecsum         = _mm_add_pd(velecsum,velec);
 339
 340             fscal            = felec;
 341
 342             /* Calculate temporary vectorial force */
 343             tx               = _mm_mul_pd(fscal,dx02);
 344             ty               = _mm_mul_pd(fscal,dy02);
 345             tz               = _mm_mul_pd(fscal,dz02);
 346
 347             /* Update vectorial force */
 348             fix0             = _mm_add_pd(fix0,tx);
 349             fiy0             = _mm_add_pd(fiy0,ty);
 350             fiz0             = _mm_add_pd(fiz0,tz);
 351
 352             fjx2             = _mm_add_pd(fjx2,tx);
 353             fjy2             = _mm_add_pd(fjy2,ty);
 354             fjz2             = _mm_add_pd(fjz2,tz);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 362             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velecsum         = _mm_add_pd(velecsum,velec);
 366
 367             fscal            = felec;
 368
 369             /* Calculate temporary vectorial force */
 370             tx               = _mm_mul_pd(fscal,dx10);
 371             ty               = _mm_mul_pd(fscal,dy10);
 372             tz               = _mm_mul_pd(fscal,dz10);
 373
 374             /* Update vectorial force */
 375             fix1             = _mm_add_pd(fix1,tx);
 376             fiy1             = _mm_add_pd(fiy1,ty);
 377             fiz1             = _mm_add_pd(fiz1,tz);
 378
 379             fjx0             = _mm_add_pd(fjx0,tx);
 380             fjy0             = _mm_add_pd(fjy0,ty);
 381             fjz0             = _mm_add_pd(fjz0,tz);
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             /* REACTION-FIELD ELECTROSTATICS */
 388             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 389             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 390
 391             /* Update potential sum for this i atom from the interaction with this j atom. */
 392             velecsum         = _mm_add_pd(velecsum,velec);
 393
 394             fscal            = felec;
 395
 396             /* Calculate temporary vectorial force */
 397             tx               = _mm_mul_pd(fscal,dx11);
 398             ty               = _mm_mul_pd(fscal,dy11);
 399             tz               = _mm_mul_pd(fscal,dz11);
 400
 401             /* Update vectorial force */
 402             fix1             = _mm_add_pd(fix1,tx);
 403             fiy1             = _mm_add_pd(fiy1,ty);
 404             fiz1             = _mm_add_pd(fiz1,tz);
 405
 406             fjx1             = _mm_add_pd(fjx1,tx);
 407             fjy1             = _mm_add_pd(fjy1,ty);
 408             fjz1             = _mm_add_pd(fjz1,tz);
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 416             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_pd(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             /* Calculate temporary vectorial force */
 424             tx               = _mm_mul_pd(fscal,dx12);
 425             ty               = _mm_mul_pd(fscal,dy12);
 426             tz               = _mm_mul_pd(fscal,dz12);
 427
 428             /* Update vectorial force */
 429             fix1             = _mm_add_pd(fix1,tx);
 430             fiy1             = _mm_add_pd(fiy1,ty);
 431             fiz1             = _mm_add_pd(fiz1,tz);
 432
 433             fjx2             = _mm_add_pd(fjx2,tx);
 434             fjy2             = _mm_add_pd(fjy2,ty);
 435             fjz2             = _mm_add_pd(fjz2,tz);
 436
 437             /**************************
 438              * CALCULATE INTERACTIONS *
 439              **************************/
 440
 441             /* REACTION-FIELD ELECTROSTATICS */
 442             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 443             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velecsum         = _mm_add_pd(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             /* Calculate temporary vectorial force */
 451             tx               = _mm_mul_pd(fscal,dx20);
 452             ty               = _mm_mul_pd(fscal,dy20);
 453             tz               = _mm_mul_pd(fscal,dz20);
 454
 455             /* Update vectorial force */
 456             fix2             = _mm_add_pd(fix2,tx);
 457             fiy2             = _mm_add_pd(fiy2,ty);
 458             fiz2             = _mm_add_pd(fiz2,tz);
 459
 460             fjx0             = _mm_add_pd(fjx0,tx);
 461             fjy0             = _mm_add_pd(fjy0,ty);
 462             fjz0             = _mm_add_pd(fjz0,tz);
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 470             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velecsum         = _mm_add_pd(velecsum,velec);
 474
 475             fscal            = felec;
 476
 477             /* Calculate temporary vectorial force */
 478             tx               = _mm_mul_pd(fscal,dx21);
 479             ty               = _mm_mul_pd(fscal,dy21);
 480             tz               = _mm_mul_pd(fscal,dz21);
 481
 482             /* Update vectorial force */
 483             fix2             = _mm_add_pd(fix2,tx);
 484             fiy2             = _mm_add_pd(fiy2,ty);
 485             fiz2             = _mm_add_pd(fiz2,tz);
 486
 487             fjx1             = _mm_add_pd(fjx1,tx);
 488             fjy1             = _mm_add_pd(fjy1,ty);
 489             fjz1             = _mm_add_pd(fjz1,tz);
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             /* REACTION-FIELD ELECTROSTATICS */
 496             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 497             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 498
 499             /* Update potential sum for this i atom from the interaction with this j atom. */
 500             velecsum         = _mm_add_pd(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             /* Calculate temporary vectorial force */
 505             tx               = _mm_mul_pd(fscal,dx22);
 506             ty               = _mm_mul_pd(fscal,dy22);
 507             tz               = _mm_mul_pd(fscal,dz22);
 508
 509             /* Update vectorial force */
 510             fix2             = _mm_add_pd(fix2,tx);
 511             fiy2             = _mm_add_pd(fiy2,ty);
 512             fiz2             = _mm_add_pd(fiz2,tz);
 513
 514             fjx2             = _mm_add_pd(fjx2,tx);
 515             fjy2             = _mm_add_pd(fjy2,ty);
 516             fjz2             = _mm_add_pd(fjz2,tz);
 517
 518             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 519
 520             /* Inner loop uses 300 flops */
 521         }
 522
 523         if(jidx<j_index_end)
 524         {
 525
 526             jnrA             = jjnr[jidx];
 527             j_coord_offsetA  = DIM*jnrA;
 528
 529             /* load j atom coordinates */
 530             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 531                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 532
 533             /* Calculate displacement vector */
 534             dx00             = _mm_sub_pd(ix0,jx0);
 535             dy00             = _mm_sub_pd(iy0,jy0);
 536             dz00             = _mm_sub_pd(iz0,jz0);
 537             dx01             = _mm_sub_pd(ix0,jx1);
 538             dy01             = _mm_sub_pd(iy0,jy1);
 539             dz01             = _mm_sub_pd(iz0,jz1);
 540             dx02             = _mm_sub_pd(ix0,jx2);
 541             dy02             = _mm_sub_pd(iy0,jy2);
 542             dz02             = _mm_sub_pd(iz0,jz2);
 543             dx10             = _mm_sub_pd(ix1,jx0);
 544             dy10             = _mm_sub_pd(iy1,jy0);
 545             dz10             = _mm_sub_pd(iz1,jz0);
 546             dx11             = _mm_sub_pd(ix1,jx1);
 547             dy11             = _mm_sub_pd(iy1,jy1);
 548             dz11             = _mm_sub_pd(iz1,jz1);
 549             dx12             = _mm_sub_pd(ix1,jx2);
 550             dy12             = _mm_sub_pd(iy1,jy2);
 551             dz12             = _mm_sub_pd(iz1,jz2);
 552             dx20             = _mm_sub_pd(ix2,jx0);
 553             dy20             = _mm_sub_pd(iy2,jy0);
 554             dz20             = _mm_sub_pd(iz2,jz0);
 555             dx21             = _mm_sub_pd(ix2,jx1);
 556             dy21             = _mm_sub_pd(iy2,jy1);
 557             dz21             = _mm_sub_pd(iz2,jz1);
 558             dx22             = _mm_sub_pd(ix2,jx2);
 559             dy22             = _mm_sub_pd(iy2,jy2);
 560             dz22             = _mm_sub_pd(iz2,jz2);
 561
 562             /* Calculate squared distance and things based on it */
 563             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 564             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 565             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 566             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 567             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 568             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 569             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 570             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 571             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 572
 573             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 574             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 575             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 576             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 577             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 578             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 579             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 580             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 581             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 582
 583             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 584             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 585             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 586             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 587             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 588             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 589             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 590             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 591             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 592
 593             fjx0             = _mm_setzero_pd();
 594             fjy0             = _mm_setzero_pd();
 595             fjz0             = _mm_setzero_pd();
 596             fjx1             = _mm_setzero_pd();
 597             fjy1             = _mm_setzero_pd();
 598             fjz1             = _mm_setzero_pd();
 599             fjx2             = _mm_setzero_pd();
 600             fjy2             = _mm_setzero_pd();
 601             fjz2             = _mm_setzero_pd();
 602
 603             /**************************
 604              * CALCULATE INTERACTIONS *
 605              **************************/
 606
 607             /* REACTION-FIELD ELECTROSTATICS */
 608             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 609             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 610
 611             /* LENNARD-JONES DISPERSION/REPULSION */
 612
 613             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 614             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 615             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 616             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 617             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 618
 619             /* Update potential sum for this i atom from the interaction with this j atom. */
 620             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 621             velecsum         = _mm_add_pd(velecsum,velec);
 622             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 623             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 624
 625             fscal            = _mm_add_pd(felec,fvdw);
 626
 627             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 628
 629             /* Calculate temporary vectorial force */
 630             tx               = _mm_mul_pd(fscal,dx00);
 631             ty               = _mm_mul_pd(fscal,dy00);
 632             tz               = _mm_mul_pd(fscal,dz00);
 633
 634             /* Update vectorial force */
 635             fix0             = _mm_add_pd(fix0,tx);
 636             fiy0             = _mm_add_pd(fiy0,ty);
 637             fiz0             = _mm_add_pd(fiz0,tz);
 638
 639             fjx0             = _mm_add_pd(fjx0,tx);
 640             fjy0             = _mm_add_pd(fjy0,ty);
 641             fjz0             = _mm_add_pd(fjz0,tz);
 642
 643             /**************************
 644              * CALCULATE INTERACTIONS *
 645              **************************/
 646
 647             /* REACTION-FIELD ELECTROSTATICS */
 648             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 649             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),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             /* Calculate temporary vectorial force */
 660             tx               = _mm_mul_pd(fscal,dx01);
 661             ty               = _mm_mul_pd(fscal,dy01);
 662             tz               = _mm_mul_pd(fscal,dz01);
 663
 664             /* Update vectorial force */
 665             fix0             = _mm_add_pd(fix0,tx);
 666             fiy0             = _mm_add_pd(fiy0,ty);
 667             fiz0             = _mm_add_pd(fiz0,tz);
 668
 669             fjx1             = _mm_add_pd(fjx1,tx);
 670             fjy1             = _mm_add_pd(fjy1,ty);
 671             fjz1             = _mm_add_pd(fjz1,tz);
 672
 673             /**************************
 674              * CALCULATE INTERACTIONS *
 675              **************************/
 676
 677             /* REACTION-FIELD ELECTROSTATICS */
 678             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 679             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 680
 681             /* Update potential sum for this i atom from the interaction with this j atom. */
 682             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 683             velecsum         = _mm_add_pd(velecsum,velec);
 684
 685             fscal            = felec;
 686
 687             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 688
 689             /* Calculate temporary vectorial force */
 690             tx               = _mm_mul_pd(fscal,dx02);
 691             ty               = _mm_mul_pd(fscal,dy02);
 692             tz               = _mm_mul_pd(fscal,dz02);
 693
 694             /* Update vectorial force */
 695             fix0             = _mm_add_pd(fix0,tx);
 696             fiy0             = _mm_add_pd(fiy0,ty);
 697             fiz0             = _mm_add_pd(fiz0,tz);
 698
 699             fjx2             = _mm_add_pd(fjx2,tx);
 700             fjy2             = _mm_add_pd(fjy2,ty);
 701             fjz2             = _mm_add_pd(fjz2,tz);
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             /* REACTION-FIELD ELECTROSTATICS */
 708             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 709             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 710
 711             /* Update potential sum for this i atom from the interaction with this j atom. */
 712             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 713             velecsum         = _mm_add_pd(velecsum,velec);
 714
 715             fscal            = felec;
 716
 717             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 718
 719             /* Calculate temporary vectorial force */
 720             tx               = _mm_mul_pd(fscal,dx10);
 721             ty               = _mm_mul_pd(fscal,dy10);
 722             tz               = _mm_mul_pd(fscal,dz10);
 723
 724             /* Update vectorial force */
 725             fix1             = _mm_add_pd(fix1,tx);
 726             fiy1             = _mm_add_pd(fiy1,ty);
 727             fiz1             = _mm_add_pd(fiz1,tz);
 728
 729             fjx0             = _mm_add_pd(fjx0,tx);
 730             fjy0             = _mm_add_pd(fjy0,ty);
 731             fjz0             = _mm_add_pd(fjz0,tz);
 732
 733             /**************************
 734              * CALCULATE INTERACTIONS *
 735              **************************/
 736
 737             /* REACTION-FIELD ELECTROSTATICS */
 738             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 739             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 740
 741             /* Update potential sum for this i atom from the interaction with this j atom. */
 742             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 743             velecsum         = _mm_add_pd(velecsum,velec);
 744
 745             fscal            = felec;
 746
 747             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 748
 749             /* Calculate temporary vectorial force */
 750             tx               = _mm_mul_pd(fscal,dx11);
 751             ty               = _mm_mul_pd(fscal,dy11);
 752             tz               = _mm_mul_pd(fscal,dz11);
 753
 754             /* Update vectorial force */
 755             fix1             = _mm_add_pd(fix1,tx);
 756             fiy1             = _mm_add_pd(fiy1,ty);
 757             fiz1             = _mm_add_pd(fiz1,tz);
 758
 759             fjx1             = _mm_add_pd(fjx1,tx);
 760             fjy1             = _mm_add_pd(fjy1,ty);
 761             fjz1             = _mm_add_pd(fjz1,tz);
 762
 763             /**************************
 764              * CALCULATE INTERACTIONS *
 765              **************************/
 766
 767             /* REACTION-FIELD ELECTROSTATICS */
 768             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 769             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 770
 771             /* Update potential sum for this i atom from the interaction with this j atom. */
 772             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 773             velecsum         = _mm_add_pd(velecsum,velec);
 774
 775             fscal            = felec;
 776
 777             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 778
 779             /* Calculate temporary vectorial force */
 780             tx               = _mm_mul_pd(fscal,dx12);
 781             ty               = _mm_mul_pd(fscal,dy12);
 782             tz               = _mm_mul_pd(fscal,dz12);
 783
 784             /* Update vectorial force */
 785             fix1             = _mm_add_pd(fix1,tx);
 786             fiy1             = _mm_add_pd(fiy1,ty);
 787             fiz1             = _mm_add_pd(fiz1,tz);
 788
 789             fjx2             = _mm_add_pd(fjx2,tx);
 790             fjy2             = _mm_add_pd(fjy2,ty);
 791             fjz2             = _mm_add_pd(fjz2,tz);
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             /* REACTION-FIELD ELECTROSTATICS */
 798             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 799             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 800
 801             /* Update potential sum for this i atom from the interaction with this j atom. */
 802             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 803             velecsum         = _mm_add_pd(velecsum,velec);
 804
 805             fscal            = felec;
 806
 807             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 808
 809             /* Calculate temporary vectorial force */
 810             tx               = _mm_mul_pd(fscal,dx20);
 811             ty               = _mm_mul_pd(fscal,dy20);
 812             tz               = _mm_mul_pd(fscal,dz20);
 813
 814             /* Update vectorial force */
 815             fix2             = _mm_add_pd(fix2,tx);
 816             fiy2             = _mm_add_pd(fiy2,ty);
 817             fiz2             = _mm_add_pd(fiz2,tz);
 818
 819             fjx0             = _mm_add_pd(fjx0,tx);
 820             fjy0             = _mm_add_pd(fjy0,ty);
 821             fjz0             = _mm_add_pd(fjz0,tz);
 822
 823             /**************************
 824              * CALCULATE INTERACTIONS *
 825              **************************/
 826
 827             /* REACTION-FIELD ELECTROSTATICS */
 828             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 829             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 830
 831             /* Update potential sum for this i atom from the interaction with this j atom. */
 832             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 833             velecsum         = _mm_add_pd(velecsum,velec);
 834
 835             fscal            = felec;
 836
 837             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 838
 839             /* Calculate temporary vectorial force */
 840             tx               = _mm_mul_pd(fscal,dx21);
 841             ty               = _mm_mul_pd(fscal,dy21);
 842             tz               = _mm_mul_pd(fscal,dz21);
 843
 844             /* Update vectorial force */
 845             fix2             = _mm_add_pd(fix2,tx);
 846             fiy2             = _mm_add_pd(fiy2,ty);
 847             fiz2             = _mm_add_pd(fiz2,tz);
 848
 849             fjx1             = _mm_add_pd(fjx1,tx);
 850             fjy1             = _mm_add_pd(fjy1,ty);
 851             fjz1             = _mm_add_pd(fjz1,tz);
 852
 853             /**************************
 854              * CALCULATE INTERACTIONS *
 855              **************************/
 856
 857             /* REACTION-FIELD ELECTROSTATICS */
 858             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 859             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 860
 861             /* Update potential sum for this i atom from the interaction with this j atom. */
 862             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 863             velecsum         = _mm_add_pd(velecsum,velec);
 864
 865             fscal            = felec;
 866
 867             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 868
 869             /* Calculate temporary vectorial force */
 870             tx               = _mm_mul_pd(fscal,dx22);
 871             ty               = _mm_mul_pd(fscal,dy22);
 872             tz               = _mm_mul_pd(fscal,dz22);
 873
 874             /* Update vectorial force */
 875             fix2             = _mm_add_pd(fix2,tx);
 876             fiy2             = _mm_add_pd(fiy2,ty);
 877             fiz2             = _mm_add_pd(fiz2,tz);
 878
 879             fjx2             = _mm_add_pd(fjx2,tx);
 880             fjy2             = _mm_add_pd(fjy2,ty);
 881             fjz2             = _mm_add_pd(fjz2,tz);
 882
 883             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 884
 885             /* Inner loop uses 300 flops */
 886         }
 887
 888         /* End of innermost loop */
 889
 890         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 891                                               f+i_coord_offset,fshift+i_shift_offset);
 892
 893         ggid                        = gid[iidx];
 894         /* Update potential energies */
 895         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 896         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 897
 898         /* Increment number of inner iterations */
 899         inneriter                  += j_index_end - j_index_start;
 900
 901         /* Outer loop uses 20 flops */
 902     }
 903
 904     /* Increment number of outer iterations */
 905     outeriter        += nri;
 906
 907     /* Update outer/inner flops */
 908
 909     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*300);
 910 }
 911 /*
 912  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse2_double
 913  * Electrostatics interaction: ReactionField
 914  * VdW interaction:            LennardJones
 915  * Geometry:                   Water3-Water3
 916  * Calculate force/pot:        Force
 917  */
 918 void
 919 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse2_double
 920                     (t_nblist * gmx_restrict                nlist,
 921                      rvec * gmx_restrict                    xx,
 922                      rvec * gmx_restrict                    ff,
 923                      t_forcerec * gmx_restrict              fr,
 924                      t_mdatoms * gmx_restrict               mdatoms,
 925                      nb_kernel_data_t * gmx_restrict        kernel_data,
 926                      t_nrnb * gmx_restrict                  nrnb)
 927 {
 928     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 929      * just 0 for non-waters.
 930      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 931      * jnr indices corresponding to data put in the four positions in the SIMD register.
 932      */
 933     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 934     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 935     int              jnrA,jnrB;
 936     int              j_coord_offsetA,j_coord_offsetB;
 937     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 938     real             rcutoff_scalar;
 939     real             *shiftvec,*fshift,*x,*f;
 940     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 941     int              vdwioffset0;
 942     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 943     int              vdwioffset1;
 944     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 945     int              vdwioffset2;
 946     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 947     int              vdwjidx0A,vdwjidx0B;
 948     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 949     int              vdwjidx1A,vdwjidx1B;
 950     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 951     int              vdwjidx2A,vdwjidx2B;
 952     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 953     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 954     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 955     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 956     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 957     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 958     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 959     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 960     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 961     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 962     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 963     real             *charge;
 964     int              nvdwtype;
 965     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 966     int              *vdwtype;
 967     real             *vdwparam;
 968     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 969     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 970     __m128d          dummy_mask,cutoff_mask;
 971     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 972     __m128d          one     = _mm_set1_pd(1.0);
 973     __m128d          two     = _mm_set1_pd(2.0);
 974     x                = xx[0];
 975     f                = ff[0];
 976
 977     nri              = nlist->nri;
 978     iinr             = nlist->iinr;
 979     jindex           = nlist->jindex;
 980     jjnr             = nlist->jjnr;
 981     shiftidx         = nlist->shift;
 982     gid              = nlist->gid;
 983     shiftvec         = fr->shift_vec[0];
 984     fshift           = fr->fshift[0];
 985     facel            = _mm_set1_pd(fr->epsfac);
 986     charge           = mdatoms->chargeA;
 987     krf              = _mm_set1_pd(fr->ic->k_rf);
 988     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 989     crf              = _mm_set1_pd(fr->ic->c_rf);
 990     nvdwtype         = fr->ntype;
 991     vdwparam         = fr->nbfp;
 992     vdwtype          = mdatoms->typeA;
 993
 994     /* Setup water-specific parameters */
 995     inr              = nlist->iinr[0];
 996     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 997     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 998     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 999     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1000
1001     jq0              = _mm_set1_pd(charge[inr+0]);
1002     jq1              = _mm_set1_pd(charge[inr+1]);
1003     jq2              = _mm_set1_pd(charge[inr+2]);
1004     vdwjidx0A        = 2*vdwtype[inr+0];
1005     qq00             = _mm_mul_pd(iq0,jq0);
1006     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1007     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1008     qq01             = _mm_mul_pd(iq0,jq1);
1009     qq02             = _mm_mul_pd(iq0,jq2);
1010     qq10             = _mm_mul_pd(iq1,jq0);
1011     qq11             = _mm_mul_pd(iq1,jq1);
1012     qq12             = _mm_mul_pd(iq1,jq2);
1013     qq20             = _mm_mul_pd(iq2,jq0);
1014     qq21             = _mm_mul_pd(iq2,jq1);
1015     qq22             = _mm_mul_pd(iq2,jq2);
1016
1017     /* Avoid stupid compiler warnings */
1018     jnrA = jnrB = 0;
1019     j_coord_offsetA = 0;
1020     j_coord_offsetB = 0;
1021
1022     outeriter        = 0;
1023     inneriter        = 0;
1024
1025     /* Start outer loop over neighborlists */
1026     for(iidx=0; iidx<nri; iidx++)
1027     {
1028         /* Load shift vector for this list */
1029         i_shift_offset   = DIM*shiftidx[iidx];
1030
1031         /* Load limits for loop over neighbors */
1032         j_index_start    = jindex[iidx];
1033         j_index_end      = jindex[iidx+1];
1034
1035         /* Get outer coordinate index */
1036         inr              = iinr[iidx];
1037         i_coord_offset   = DIM*inr;
1038
1039         /* Load i particle coords and add shift vector */
1040         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1041                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1042
1043         fix0             = _mm_setzero_pd();
1044         fiy0             = _mm_setzero_pd();
1045         fiz0             = _mm_setzero_pd();
1046         fix1             = _mm_setzero_pd();
1047         fiy1             = _mm_setzero_pd();
1048         fiz1             = _mm_setzero_pd();
1049         fix2             = _mm_setzero_pd();
1050         fiy2             = _mm_setzero_pd();
1051         fiz2             = _mm_setzero_pd();
1052
1053         /* Start inner kernel loop */
1054         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1055         {
1056
1057             /* Get j neighbor index, and coordinate index */
1058             jnrA             = jjnr[jidx];
1059             jnrB             = jjnr[jidx+1];
1060             j_coord_offsetA  = DIM*jnrA;
1061             j_coord_offsetB  = DIM*jnrB;
1062
1063             /* load j atom coordinates */
1064             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1065                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1066
1067             /* Calculate displacement vector */
1068             dx00             = _mm_sub_pd(ix0,jx0);
1069             dy00             = _mm_sub_pd(iy0,jy0);
1070             dz00             = _mm_sub_pd(iz0,jz0);
1071             dx01             = _mm_sub_pd(ix0,jx1);
1072             dy01             = _mm_sub_pd(iy0,jy1);
1073             dz01             = _mm_sub_pd(iz0,jz1);
1074             dx02             = _mm_sub_pd(ix0,jx2);
1075             dy02             = _mm_sub_pd(iy0,jy2);
1076             dz02             = _mm_sub_pd(iz0,jz2);
1077             dx10             = _mm_sub_pd(ix1,jx0);
1078             dy10             = _mm_sub_pd(iy1,jy0);
1079             dz10             = _mm_sub_pd(iz1,jz0);
1080             dx11             = _mm_sub_pd(ix1,jx1);
1081             dy11             = _mm_sub_pd(iy1,jy1);
1082             dz11             = _mm_sub_pd(iz1,jz1);
1083             dx12             = _mm_sub_pd(ix1,jx2);
1084             dy12             = _mm_sub_pd(iy1,jy2);
1085             dz12             = _mm_sub_pd(iz1,jz2);
1086             dx20             = _mm_sub_pd(ix2,jx0);
1087             dy20             = _mm_sub_pd(iy2,jy0);
1088             dz20             = _mm_sub_pd(iz2,jz0);
1089             dx21             = _mm_sub_pd(ix2,jx1);
1090             dy21             = _mm_sub_pd(iy2,jy1);
1091             dz21             = _mm_sub_pd(iz2,jz1);
1092             dx22             = _mm_sub_pd(ix2,jx2);
1093             dy22             = _mm_sub_pd(iy2,jy2);
1094             dz22             = _mm_sub_pd(iz2,jz2);
1095
1096             /* Calculate squared distance and things based on it */
1097             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1098             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1099             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1100             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1101             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1102             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1103             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1104             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1105             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1106
1107             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1108             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1109             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1110             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1111             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1112             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1113             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1114             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1115             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1116
1117             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1118             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1119             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1120             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1121             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1122             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1123             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1124             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1125             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1126
1127             fjx0             = _mm_setzero_pd();
1128             fjy0             = _mm_setzero_pd();
1129             fjz0             = _mm_setzero_pd();
1130             fjx1             = _mm_setzero_pd();
1131             fjy1             = _mm_setzero_pd();
1132             fjz1             = _mm_setzero_pd();
1133             fjx2             = _mm_setzero_pd();
1134             fjy2             = _mm_setzero_pd();
1135             fjz2             = _mm_setzero_pd();
1136
1137             /**************************
1138              * CALCULATE INTERACTIONS *
1139              **************************/
1140
1141             /* REACTION-FIELD ELECTROSTATICS */
1142             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1143
1144             /* LENNARD-JONES DISPERSION/REPULSION */
1145
1146             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1147             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1148
1149             fscal            = _mm_add_pd(felec,fvdw);
1150
1151             /* Calculate temporary vectorial force */
1152             tx               = _mm_mul_pd(fscal,dx00);
1153             ty               = _mm_mul_pd(fscal,dy00);
1154             tz               = _mm_mul_pd(fscal,dz00);
1155
1156             /* Update vectorial force */
1157             fix0             = _mm_add_pd(fix0,tx);
1158             fiy0             = _mm_add_pd(fiy0,ty);
1159             fiz0             = _mm_add_pd(fiz0,tz);
1160
1161             fjx0             = _mm_add_pd(fjx0,tx);
1162             fjy0             = _mm_add_pd(fjy0,ty);
1163             fjz0             = _mm_add_pd(fjz0,tz);
1164
1165             /**************************
1166              * CALCULATE INTERACTIONS *
1167              **************************/
1168
1169             /* REACTION-FIELD ELECTROSTATICS */
1170             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1171
1172             fscal            = felec;
1173
1174             /* Calculate temporary vectorial force */
1175             tx               = _mm_mul_pd(fscal,dx01);
1176             ty               = _mm_mul_pd(fscal,dy01);
1177             tz               = _mm_mul_pd(fscal,dz01);
1178
1179             /* Update vectorial force */
1180             fix0             = _mm_add_pd(fix0,tx);
1181             fiy0             = _mm_add_pd(fiy0,ty);
1182             fiz0             = _mm_add_pd(fiz0,tz);
1183
1184             fjx1             = _mm_add_pd(fjx1,tx);
1185             fjy1             = _mm_add_pd(fjy1,ty);
1186             fjz1             = _mm_add_pd(fjz1,tz);
1187
1188             /**************************
1189              * CALCULATE INTERACTIONS *
1190              **************************/
1191
1192             /* REACTION-FIELD ELECTROSTATICS */
1193             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1194
1195             fscal            = felec;
1196
1197             /* Calculate temporary vectorial force */
1198             tx               = _mm_mul_pd(fscal,dx02);
1199             ty               = _mm_mul_pd(fscal,dy02);
1200             tz               = _mm_mul_pd(fscal,dz02);
1201
1202             /* Update vectorial force */
1203             fix0             = _mm_add_pd(fix0,tx);
1204             fiy0             = _mm_add_pd(fiy0,ty);
1205             fiz0             = _mm_add_pd(fiz0,tz);
1206
1207             fjx2             = _mm_add_pd(fjx2,tx);
1208             fjy2             = _mm_add_pd(fjy2,ty);
1209             fjz2             = _mm_add_pd(fjz2,tz);
1210
1211             /**************************
1212              * CALCULATE INTERACTIONS *
1213              **************************/
1214
1215             /* REACTION-FIELD ELECTROSTATICS */
1216             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1217
1218             fscal            = felec;
1219
1220             /* Calculate temporary vectorial force */
1221             tx               = _mm_mul_pd(fscal,dx10);
1222             ty               = _mm_mul_pd(fscal,dy10);
1223             tz               = _mm_mul_pd(fscal,dz10);
1224
1225             /* Update vectorial force */
1226             fix1             = _mm_add_pd(fix1,tx);
1227             fiy1             = _mm_add_pd(fiy1,ty);
1228             fiz1             = _mm_add_pd(fiz1,tz);
1229
1230             fjx0             = _mm_add_pd(fjx0,tx);
1231             fjy0             = _mm_add_pd(fjy0,ty);
1232             fjz0             = _mm_add_pd(fjz0,tz);
1233
1234             /**************************
1235              * CALCULATE INTERACTIONS *
1236              **************************/
1237
1238             /* REACTION-FIELD ELECTROSTATICS */
1239             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1240
1241             fscal            = felec;
1242
1243             /* Calculate temporary vectorial force */
1244             tx               = _mm_mul_pd(fscal,dx11);
1245             ty               = _mm_mul_pd(fscal,dy11);
1246             tz               = _mm_mul_pd(fscal,dz11);
1247
1248             /* Update vectorial force */
1249             fix1             = _mm_add_pd(fix1,tx);
1250             fiy1             = _mm_add_pd(fiy1,ty);
1251             fiz1             = _mm_add_pd(fiz1,tz);
1252
1253             fjx1             = _mm_add_pd(fjx1,tx);
1254             fjy1             = _mm_add_pd(fjy1,ty);
1255             fjz1             = _mm_add_pd(fjz1,tz);
1256
1257             /**************************
1258              * CALCULATE INTERACTIONS *
1259              **************************/
1260
1261             /* REACTION-FIELD ELECTROSTATICS */
1262             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1263
1264             fscal            = felec;
1265
1266             /* Calculate temporary vectorial force */
1267             tx               = _mm_mul_pd(fscal,dx12);
1268             ty               = _mm_mul_pd(fscal,dy12);
1269             tz               = _mm_mul_pd(fscal,dz12);
1270
1271             /* Update vectorial force */
1272             fix1             = _mm_add_pd(fix1,tx);
1273             fiy1             = _mm_add_pd(fiy1,ty);
1274             fiz1             = _mm_add_pd(fiz1,tz);
1275
1276             fjx2             = _mm_add_pd(fjx2,tx);
1277             fjy2             = _mm_add_pd(fjy2,ty);
1278             fjz2             = _mm_add_pd(fjz2,tz);
1279
1280             /**************************
1281              * CALCULATE INTERACTIONS *
1282              **************************/
1283
1284             /* REACTION-FIELD ELECTROSTATICS */
1285             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1286
1287             fscal            = felec;
1288
1289             /* Calculate temporary vectorial force */
1290             tx               = _mm_mul_pd(fscal,dx20);
1291             ty               = _mm_mul_pd(fscal,dy20);
1292             tz               = _mm_mul_pd(fscal,dz20);
1293
1294             /* Update vectorial force */
1295             fix2             = _mm_add_pd(fix2,tx);
1296             fiy2             = _mm_add_pd(fiy2,ty);
1297             fiz2             = _mm_add_pd(fiz2,tz);
1298
1299             fjx0             = _mm_add_pd(fjx0,tx);
1300             fjy0             = _mm_add_pd(fjy0,ty);
1301             fjz0             = _mm_add_pd(fjz0,tz);
1302
1303             /**************************
1304              * CALCULATE INTERACTIONS *
1305              **************************/
1306
1307             /* REACTION-FIELD ELECTROSTATICS */
1308             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1309
1310             fscal            = felec;
1311
1312             /* Calculate temporary vectorial force */
1313             tx               = _mm_mul_pd(fscal,dx21);
1314             ty               = _mm_mul_pd(fscal,dy21);
1315             tz               = _mm_mul_pd(fscal,dz21);
1316
1317             /* Update vectorial force */
1318             fix2             = _mm_add_pd(fix2,tx);
1319             fiy2             = _mm_add_pd(fiy2,ty);
1320             fiz2             = _mm_add_pd(fiz2,tz);
1321
1322             fjx1             = _mm_add_pd(fjx1,tx);
1323             fjy1             = _mm_add_pd(fjy1,ty);
1324             fjz1             = _mm_add_pd(fjz1,tz);
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             /* REACTION-FIELD ELECTROSTATICS */
1331             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1332
1333             fscal            = felec;
1334
1335             /* Calculate temporary vectorial force */
1336             tx               = _mm_mul_pd(fscal,dx22);
1337             ty               = _mm_mul_pd(fscal,dy22);
1338             tz               = _mm_mul_pd(fscal,dz22);
1339
1340             /* Update vectorial force */
1341             fix2             = _mm_add_pd(fix2,tx);
1342             fiy2             = _mm_add_pd(fiy2,ty);
1343             fiz2             = _mm_add_pd(fiz2,tz);
1344
1345             fjx2             = _mm_add_pd(fjx2,tx);
1346             fjy2             = _mm_add_pd(fjy2,ty);
1347             fjz2             = _mm_add_pd(fjz2,tz);
1348
1349             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1350
1351             /* Inner loop uses 250 flops */
1352         }
1353
1354         if(jidx<j_index_end)
1355         {
1356
1357             jnrA             = jjnr[jidx];
1358             j_coord_offsetA  = DIM*jnrA;
1359
1360             /* load j atom coordinates */
1361             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1362                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1363
1364             /* Calculate displacement vector */
1365             dx00             = _mm_sub_pd(ix0,jx0);
1366             dy00             = _mm_sub_pd(iy0,jy0);
1367             dz00             = _mm_sub_pd(iz0,jz0);
1368             dx01             = _mm_sub_pd(ix0,jx1);
1369             dy01             = _mm_sub_pd(iy0,jy1);
1370             dz01             = _mm_sub_pd(iz0,jz1);
1371             dx02             = _mm_sub_pd(ix0,jx2);
1372             dy02             = _mm_sub_pd(iy0,jy2);
1373             dz02             = _mm_sub_pd(iz0,jz2);
1374             dx10             = _mm_sub_pd(ix1,jx0);
1375             dy10             = _mm_sub_pd(iy1,jy0);
1376             dz10             = _mm_sub_pd(iz1,jz0);
1377             dx11             = _mm_sub_pd(ix1,jx1);
1378             dy11             = _mm_sub_pd(iy1,jy1);
1379             dz11             = _mm_sub_pd(iz1,jz1);
1380             dx12             = _mm_sub_pd(ix1,jx2);
1381             dy12             = _mm_sub_pd(iy1,jy2);
1382             dz12             = _mm_sub_pd(iz1,jz2);
1383             dx20             = _mm_sub_pd(ix2,jx0);
1384             dy20             = _mm_sub_pd(iy2,jy0);
1385             dz20             = _mm_sub_pd(iz2,jz0);
1386             dx21             = _mm_sub_pd(ix2,jx1);
1387             dy21             = _mm_sub_pd(iy2,jy1);
1388             dz21             = _mm_sub_pd(iz2,jz1);
1389             dx22             = _mm_sub_pd(ix2,jx2);
1390             dy22             = _mm_sub_pd(iy2,jy2);
1391             dz22             = _mm_sub_pd(iz2,jz2);
1392
1393             /* Calculate squared distance and things based on it */
1394             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1395             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1396             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1397             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1398             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1399             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1400             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1401             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1402             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1403
1404             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1405             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1406             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1407             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1408             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1409             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1410             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1411             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1412             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1413
1414             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1415             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1416             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1417             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1418             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1419             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1420             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1421             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1422             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1423
1424             fjx0             = _mm_setzero_pd();
1425             fjy0             = _mm_setzero_pd();
1426             fjz0             = _mm_setzero_pd();
1427             fjx1             = _mm_setzero_pd();
1428             fjy1             = _mm_setzero_pd();
1429             fjz1             = _mm_setzero_pd();
1430             fjx2             = _mm_setzero_pd();
1431             fjy2             = _mm_setzero_pd();
1432             fjz2             = _mm_setzero_pd();
1433
1434             /**************************
1435              * CALCULATE INTERACTIONS *
1436              **************************/
1437
1438             /* REACTION-FIELD ELECTROSTATICS */
1439             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1440
1441             /* LENNARD-JONES DISPERSION/REPULSION */
1442
1443             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1444             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1445
1446             fscal            = _mm_add_pd(felec,fvdw);
1447
1448             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1449
1450             /* Calculate temporary vectorial force */
1451             tx               = _mm_mul_pd(fscal,dx00);
1452             ty               = _mm_mul_pd(fscal,dy00);
1453             tz               = _mm_mul_pd(fscal,dz00);
1454
1455             /* Update vectorial force */
1456             fix0             = _mm_add_pd(fix0,tx);
1457             fiy0             = _mm_add_pd(fiy0,ty);
1458             fiz0             = _mm_add_pd(fiz0,tz);
1459
1460             fjx0             = _mm_add_pd(fjx0,tx);
1461             fjy0             = _mm_add_pd(fjy0,ty);
1462             fjz0             = _mm_add_pd(fjz0,tz);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             /* REACTION-FIELD ELECTROSTATICS */
1469             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1470
1471             fscal            = felec;
1472
1473             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1474
1475             /* Calculate temporary vectorial force */
1476             tx               = _mm_mul_pd(fscal,dx01);
1477             ty               = _mm_mul_pd(fscal,dy01);
1478             tz               = _mm_mul_pd(fscal,dz01);
1479
1480             /* Update vectorial force */
1481             fix0             = _mm_add_pd(fix0,tx);
1482             fiy0             = _mm_add_pd(fiy0,ty);
1483             fiz0             = _mm_add_pd(fiz0,tz);
1484
1485             fjx1             = _mm_add_pd(fjx1,tx);
1486             fjy1             = _mm_add_pd(fjy1,ty);
1487             fjz1             = _mm_add_pd(fjz1,tz);
1488
1489             /**************************
1490              * CALCULATE INTERACTIONS *
1491              **************************/
1492
1493             /* REACTION-FIELD ELECTROSTATICS */
1494             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1495
1496             fscal            = felec;
1497
1498             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1499
1500             /* Calculate temporary vectorial force */
1501             tx               = _mm_mul_pd(fscal,dx02);
1502             ty               = _mm_mul_pd(fscal,dy02);
1503             tz               = _mm_mul_pd(fscal,dz02);
1504
1505             /* Update vectorial force */
1506             fix0             = _mm_add_pd(fix0,tx);
1507             fiy0             = _mm_add_pd(fiy0,ty);
1508             fiz0             = _mm_add_pd(fiz0,tz);
1509
1510             fjx2             = _mm_add_pd(fjx2,tx);
1511             fjy2             = _mm_add_pd(fjy2,ty);
1512             fjz2             = _mm_add_pd(fjz2,tz);
1513
1514             /**************************
1515              * CALCULATE INTERACTIONS *
1516              **************************/
1517
1518             /* REACTION-FIELD ELECTROSTATICS */
1519             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1520
1521             fscal            = felec;
1522
1523             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1524
1525             /* Calculate temporary vectorial force */
1526             tx               = _mm_mul_pd(fscal,dx10);
1527             ty               = _mm_mul_pd(fscal,dy10);
1528             tz               = _mm_mul_pd(fscal,dz10);
1529
1530             /* Update vectorial force */
1531             fix1             = _mm_add_pd(fix1,tx);
1532             fiy1             = _mm_add_pd(fiy1,ty);
1533             fiz1             = _mm_add_pd(fiz1,tz);
1534
1535             fjx0             = _mm_add_pd(fjx0,tx);
1536             fjy0             = _mm_add_pd(fjy0,ty);
1537             fjz0             = _mm_add_pd(fjz0,tz);
1538
1539             /**************************
1540              * CALCULATE INTERACTIONS *
1541              **************************/
1542
1543             /* REACTION-FIELD ELECTROSTATICS */
1544             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1549
1550             /* Calculate temporary vectorial force */
1551             tx               = _mm_mul_pd(fscal,dx11);
1552             ty               = _mm_mul_pd(fscal,dy11);
1553             tz               = _mm_mul_pd(fscal,dz11);
1554
1555             /* Update vectorial force */
1556             fix1             = _mm_add_pd(fix1,tx);
1557             fiy1             = _mm_add_pd(fiy1,ty);
1558             fiz1             = _mm_add_pd(fiz1,tz);
1559
1560             fjx1             = _mm_add_pd(fjx1,tx);
1561             fjy1             = _mm_add_pd(fjy1,ty);
1562             fjz1             = _mm_add_pd(fjz1,tz);
1563
1564             /**************************
1565              * CALCULATE INTERACTIONS *
1566              **************************/
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1570
1571             fscal            = felec;
1572
1573             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1574
1575             /* Calculate temporary vectorial force */
1576             tx               = _mm_mul_pd(fscal,dx12);
1577             ty               = _mm_mul_pd(fscal,dy12);
1578             tz               = _mm_mul_pd(fscal,dz12);
1579
1580             /* Update vectorial force */
1581             fix1             = _mm_add_pd(fix1,tx);
1582             fiy1             = _mm_add_pd(fiy1,ty);
1583             fiz1             = _mm_add_pd(fiz1,tz);
1584
1585             fjx2             = _mm_add_pd(fjx2,tx);
1586             fjy2             = _mm_add_pd(fjy2,ty);
1587             fjz2             = _mm_add_pd(fjz2,tz);
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             /* REACTION-FIELD ELECTROSTATICS */
1594             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1595
1596             fscal            = felec;
1597
1598             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1599
1600             /* Calculate temporary vectorial force */
1601             tx               = _mm_mul_pd(fscal,dx20);
1602             ty               = _mm_mul_pd(fscal,dy20);
1603             tz               = _mm_mul_pd(fscal,dz20);
1604
1605             /* Update vectorial force */
1606             fix2             = _mm_add_pd(fix2,tx);
1607             fiy2             = _mm_add_pd(fiy2,ty);
1608             fiz2             = _mm_add_pd(fiz2,tz);
1609
1610             fjx0             = _mm_add_pd(fjx0,tx);
1611             fjy0             = _mm_add_pd(fjy0,ty);
1612             fjz0             = _mm_add_pd(fjz0,tz);
1613
1614             /**************************
1615              * CALCULATE INTERACTIONS *
1616              **************************/
1617
1618             /* REACTION-FIELD ELECTROSTATICS */
1619             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1620
1621             fscal            = felec;
1622
1623             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1624
1625             /* Calculate temporary vectorial force */
1626             tx               = _mm_mul_pd(fscal,dx21);
1627             ty               = _mm_mul_pd(fscal,dy21);
1628             tz               = _mm_mul_pd(fscal,dz21);
1629
1630             /* Update vectorial force */
1631             fix2             = _mm_add_pd(fix2,tx);
1632             fiy2             = _mm_add_pd(fiy2,ty);
1633             fiz2             = _mm_add_pd(fiz2,tz);
1634
1635             fjx1             = _mm_add_pd(fjx1,tx);
1636             fjy1             = _mm_add_pd(fjy1,ty);
1637             fjz1             = _mm_add_pd(fjz1,tz);
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             /* REACTION-FIELD ELECTROSTATICS */
1644             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1645
1646             fscal            = felec;
1647
1648             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1649
1650             /* Calculate temporary vectorial force */
1651             tx               = _mm_mul_pd(fscal,dx22);
1652             ty               = _mm_mul_pd(fscal,dy22);
1653             tz               = _mm_mul_pd(fscal,dz22);
1654
1655             /* Update vectorial force */
1656             fix2             = _mm_add_pd(fix2,tx);
1657             fiy2             = _mm_add_pd(fiy2,ty);
1658             fiz2             = _mm_add_pd(fiz2,tz);
1659
1660             fjx2             = _mm_add_pd(fjx2,tx);
1661             fjy2             = _mm_add_pd(fjy2,ty);
1662             fjz2             = _mm_add_pd(fjz2,tz);
1663
1664             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1665
1666             /* Inner loop uses 250 flops */
1667         }
1668
1669         /* End of innermost loop */
1670
1671         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1672                                               f+i_coord_offset,fshift+i_shift_offset);
1673
1674         /* Increment number of inner iterations */
1675         inneriter                  += j_index_end - j_index_start;
1676
1677         /* Outer loop uses 18 flops */
1678     }
1679
1680     /* Increment number of outer iterations */
1681     outeriter        += nri;
1682
1683     /* Update outer/inner flops */
1684
1685     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1686 }