src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwNone_GeomW3W3_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_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_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sse4_1_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     __m128d          dummy_mask,cutoff_mask;
  90     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  91     __m128d          one     = _mm_set1_pd(1.0);
  92     __m128d          two     = _mm_set1_pd(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_pd(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106     krf              = _mm_set1_pd(fr->ic->k_rf);
 107     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 108     crf              = _mm_set1_pd(fr->ic->c_rf);
 109
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 113     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 114     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 115
 116     jq0              = _mm_set1_pd(charge[inr+0]);
 117     jq1              = _mm_set1_pd(charge[inr+1]);
 118     jq2              = _mm_set1_pd(charge[inr+2]);
 119     qq00             = _mm_mul_pd(iq0,jq0);
 120     qq01             = _mm_mul_pd(iq0,jq1);
 121     qq02             = _mm_mul_pd(iq0,jq2);
 122     qq10             = _mm_mul_pd(iq1,jq0);
 123     qq11             = _mm_mul_pd(iq1,jq1);
 124     qq12             = _mm_mul_pd(iq1,jq2);
 125     qq20             = _mm_mul_pd(iq2,jq0);
 126     qq21             = _mm_mul_pd(iq2,jq1);
 127     qq22             = _mm_mul_pd(iq2,jq2);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 153                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 154
 155         fix0             = _mm_setzero_pd();
 156         fiy0             = _mm_setzero_pd();
 157         fiz0             = _mm_setzero_pd();
 158         fix1             = _mm_setzero_pd();
 159         fiy1             = _mm_setzero_pd();
 160         fiz1             = _mm_setzero_pd();
 161         fix2             = _mm_setzero_pd();
 162         fiy2             = _mm_setzero_pd();
 163         fiz2             = _mm_setzero_pd();
 164
 165         /* Reset potential sums */
 166         velecsum         = _mm_setzero_pd();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177
 178             /* load j atom coordinates */
 179             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 181
 182             /* Calculate displacement vector */
 183             dx00             = _mm_sub_pd(ix0,jx0);
 184             dy00             = _mm_sub_pd(iy0,jy0);
 185             dz00             = _mm_sub_pd(iz0,jz0);
 186             dx01             = _mm_sub_pd(ix0,jx1);
 187             dy01             = _mm_sub_pd(iy0,jy1);
 188             dz01             = _mm_sub_pd(iz0,jz1);
 189             dx02             = _mm_sub_pd(ix0,jx2);
 190             dy02             = _mm_sub_pd(iy0,jy2);
 191             dz02             = _mm_sub_pd(iz0,jz2);
 192             dx10             = _mm_sub_pd(ix1,jx0);
 193             dy10             = _mm_sub_pd(iy1,jy0);
 194             dz10             = _mm_sub_pd(iz1,jz0);
 195             dx11             = _mm_sub_pd(ix1,jx1);
 196             dy11             = _mm_sub_pd(iy1,jy1);
 197             dz11             = _mm_sub_pd(iz1,jz1);
 198             dx12             = _mm_sub_pd(ix1,jx2);
 199             dy12             = _mm_sub_pd(iy1,jy2);
 200             dz12             = _mm_sub_pd(iz1,jz2);
 201             dx20             = _mm_sub_pd(ix2,jx0);
 202             dy20             = _mm_sub_pd(iy2,jy0);
 203             dz20             = _mm_sub_pd(iz2,jz0);
 204             dx21             = _mm_sub_pd(ix2,jx1);
 205             dy21             = _mm_sub_pd(iy2,jy1);
 206             dz21             = _mm_sub_pd(iz2,jz1);
 207             dx22             = _mm_sub_pd(ix2,jx2);
 208             dy22             = _mm_sub_pd(iy2,jy2);
 209             dz22             = _mm_sub_pd(iz2,jz2);
 210
 211             /* Calculate squared distance and things based on it */
 212             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 213             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 214             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 215             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 216             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 217             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 218             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 219             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 220             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 221
 222             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 223             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 224             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 225             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 226             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 227             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 228             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 229             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 230             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 231
 232             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 233             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 234             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 235             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 236             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 237             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 238             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 239             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 240             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 241
 242             fjx0             = _mm_setzero_pd();
 243             fjy0             = _mm_setzero_pd();
 244             fjz0             = _mm_setzero_pd();
 245             fjx1             = _mm_setzero_pd();
 246             fjy1             = _mm_setzero_pd();
 247             fjz1             = _mm_setzero_pd();
 248             fjx2             = _mm_setzero_pd();
 249             fjy2             = _mm_setzero_pd();
 250             fjz2             = _mm_setzero_pd();
 251
 252             /**************************
 253              * CALCULATE INTERACTIONS *
 254              **************************/
 255
 256             /* REACTION-FIELD ELECTROSTATICS */
 257             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 258             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 259
 260             /* Update potential sum for this i atom from the interaction with this j atom. */
 261             velecsum         = _mm_add_pd(velecsum,velec);
 262
 263             fscal            = felec;
 264
 265             /* Calculate temporary vectorial force */
 266             tx               = _mm_mul_pd(fscal,dx00);
 267             ty               = _mm_mul_pd(fscal,dy00);
 268             tz               = _mm_mul_pd(fscal,dz00);
 269
 270             /* Update vectorial force */
 271             fix0             = _mm_add_pd(fix0,tx);
 272             fiy0             = _mm_add_pd(fiy0,ty);
 273             fiz0             = _mm_add_pd(fiz0,tz);
 274
 275             fjx0             = _mm_add_pd(fjx0,tx);
 276             fjy0             = _mm_add_pd(fjy0,ty);
 277             fjz0             = _mm_add_pd(fjz0,tz);
 278
 279             /**************************
 280              * CALCULATE INTERACTIONS *
 281              **************************/
 282
 283             /* REACTION-FIELD ELECTROSTATICS */
 284             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 285             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 286
 287             /* Update potential sum for this i atom from the interaction with this j atom. */
 288             velecsum         = _mm_add_pd(velecsum,velec);
 289
 290             fscal            = felec;
 291
 292             /* Calculate temporary vectorial force */
 293             tx               = _mm_mul_pd(fscal,dx01);
 294             ty               = _mm_mul_pd(fscal,dy01);
 295             tz               = _mm_mul_pd(fscal,dz01);
 296
 297             /* Update vectorial force */
 298             fix0             = _mm_add_pd(fix0,tx);
 299             fiy0             = _mm_add_pd(fiy0,ty);
 300             fiz0             = _mm_add_pd(fiz0,tz);
 301
 302             fjx1             = _mm_add_pd(fjx1,tx);
 303             fjy1             = _mm_add_pd(fjy1,ty);
 304             fjz1             = _mm_add_pd(fjz1,tz);
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             /* REACTION-FIELD ELECTROSTATICS */
 311             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 312             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 313
 314             /* Update potential sum for this i atom from the interaction with this j atom. */
 315             velecsum         = _mm_add_pd(velecsum,velec);
 316
 317             fscal            = felec;
 318
 319             /* Calculate temporary vectorial force */
 320             tx               = _mm_mul_pd(fscal,dx02);
 321             ty               = _mm_mul_pd(fscal,dy02);
 322             tz               = _mm_mul_pd(fscal,dz02);
 323
 324             /* Update vectorial force */
 325             fix0             = _mm_add_pd(fix0,tx);
 326             fiy0             = _mm_add_pd(fiy0,ty);
 327             fiz0             = _mm_add_pd(fiz0,tz);
 328
 329             fjx2             = _mm_add_pd(fjx2,tx);
 330             fjy2             = _mm_add_pd(fjy2,ty);
 331             fjz2             = _mm_add_pd(fjz2,tz);
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             /* REACTION-FIELD ELECTROSTATICS */
 338             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 339             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 340
 341             /* Update potential sum for this i atom from the interaction with this j atom. */
 342             velecsum         = _mm_add_pd(velecsum,velec);
 343
 344             fscal            = felec;
 345
 346             /* Calculate temporary vectorial force */
 347             tx               = _mm_mul_pd(fscal,dx10);
 348             ty               = _mm_mul_pd(fscal,dy10);
 349             tz               = _mm_mul_pd(fscal,dz10);
 350
 351             /* Update vectorial force */
 352             fix1             = _mm_add_pd(fix1,tx);
 353             fiy1             = _mm_add_pd(fiy1,ty);
 354             fiz1             = _mm_add_pd(fiz1,tz);
 355
 356             fjx0             = _mm_add_pd(fjx0,tx);
 357             fjy0             = _mm_add_pd(fjy0,ty);
 358             fjz0             = _mm_add_pd(fjz0,tz);
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             /* REACTION-FIELD ELECTROSTATICS */
 365             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 366             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velecsum         = _mm_add_pd(velecsum,velec);
 370
 371             fscal            = felec;
 372
 373             /* Calculate temporary vectorial force */
 374             tx               = _mm_mul_pd(fscal,dx11);
 375             ty               = _mm_mul_pd(fscal,dy11);
 376             tz               = _mm_mul_pd(fscal,dz11);
 377
 378             /* Update vectorial force */
 379             fix1             = _mm_add_pd(fix1,tx);
 380             fiy1             = _mm_add_pd(fiy1,ty);
 381             fiz1             = _mm_add_pd(fiz1,tz);
 382
 383             fjx1             = _mm_add_pd(fjx1,tx);
 384             fjy1             = _mm_add_pd(fjy1,ty);
 385             fjz1             = _mm_add_pd(fjz1,tz);
 386
 387             /**************************
 388              * CALCULATE INTERACTIONS *
 389              **************************/
 390
 391             /* REACTION-FIELD ELECTROSTATICS */
 392             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 393             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velecsum         = _mm_add_pd(velecsum,velec);
 397
 398             fscal            = felec;
 399
 400             /* Calculate temporary vectorial force */
 401             tx               = _mm_mul_pd(fscal,dx12);
 402             ty               = _mm_mul_pd(fscal,dy12);
 403             tz               = _mm_mul_pd(fscal,dz12);
 404
 405             /* Update vectorial force */
 406             fix1             = _mm_add_pd(fix1,tx);
 407             fiy1             = _mm_add_pd(fiy1,ty);
 408             fiz1             = _mm_add_pd(fiz1,tz);
 409
 410             fjx2             = _mm_add_pd(fjx2,tx);
 411             fjy2             = _mm_add_pd(fjy2,ty);
 412             fjz2             = _mm_add_pd(fjz2,tz);
 413
 414             /**************************
 415              * CALCULATE INTERACTIONS *
 416              **************************/
 417
 418             /* REACTION-FIELD ELECTROSTATICS */
 419             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 420             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 421
 422             /* Update potential sum for this i atom from the interaction with this j atom. */
 423             velecsum         = _mm_add_pd(velecsum,velec);
 424
 425             fscal            = felec;
 426
 427             /* Calculate temporary vectorial force */
 428             tx               = _mm_mul_pd(fscal,dx20);
 429             ty               = _mm_mul_pd(fscal,dy20);
 430             tz               = _mm_mul_pd(fscal,dz20);
 431
 432             /* Update vectorial force */
 433             fix2             = _mm_add_pd(fix2,tx);
 434             fiy2             = _mm_add_pd(fiy2,ty);
 435             fiz2             = _mm_add_pd(fiz2,tz);
 436
 437             fjx0             = _mm_add_pd(fjx0,tx);
 438             fjy0             = _mm_add_pd(fjy0,ty);
 439             fjz0             = _mm_add_pd(fjz0,tz);
 440
 441             /**************************
 442              * CALCULATE INTERACTIONS *
 443              **************************/
 444
 445             /* REACTION-FIELD ELECTROSTATICS */
 446             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 447             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 448
 449             /* Update potential sum for this i atom from the interaction with this j atom. */
 450             velecsum         = _mm_add_pd(velecsum,velec);
 451
 452             fscal            = felec;
 453
 454             /* Calculate temporary vectorial force */
 455             tx               = _mm_mul_pd(fscal,dx21);
 456             ty               = _mm_mul_pd(fscal,dy21);
 457             tz               = _mm_mul_pd(fscal,dz21);
 458
 459             /* Update vectorial force */
 460             fix2             = _mm_add_pd(fix2,tx);
 461             fiy2             = _mm_add_pd(fiy2,ty);
 462             fiz2             = _mm_add_pd(fiz2,tz);
 463
 464             fjx1             = _mm_add_pd(fjx1,tx);
 465             fjy1             = _mm_add_pd(fjy1,ty);
 466             fjz1             = _mm_add_pd(fjz1,tz);
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             /* REACTION-FIELD ELECTROSTATICS */
 473             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 474             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 475
 476             /* Update potential sum for this i atom from the interaction with this j atom. */
 477             velecsum         = _mm_add_pd(velecsum,velec);
 478
 479             fscal            = felec;
 480
 481             /* Calculate temporary vectorial force */
 482             tx               = _mm_mul_pd(fscal,dx22);
 483             ty               = _mm_mul_pd(fscal,dy22);
 484             tz               = _mm_mul_pd(fscal,dz22);
 485
 486             /* Update vectorial force */
 487             fix2             = _mm_add_pd(fix2,tx);
 488             fiy2             = _mm_add_pd(fiy2,ty);
 489             fiz2             = _mm_add_pd(fiz2,tz);
 490
 491             fjx2             = _mm_add_pd(fjx2,tx);
 492             fjy2             = _mm_add_pd(fjy2,ty);
 493             fjz2             = _mm_add_pd(fjz2,tz);
 494
 495             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 496
 497             /* Inner loop uses 288 flops */
 498         }
 499
 500         if(jidx<j_index_end)
 501         {
 502
 503             jnrA             = jjnr[jidx];
 504             j_coord_offsetA  = DIM*jnrA;
 505
 506             /* load j atom coordinates */
 507             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 508                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 509
 510             /* Calculate displacement vector */
 511             dx00             = _mm_sub_pd(ix0,jx0);
 512             dy00             = _mm_sub_pd(iy0,jy0);
 513             dz00             = _mm_sub_pd(iz0,jz0);
 514             dx01             = _mm_sub_pd(ix0,jx1);
 515             dy01             = _mm_sub_pd(iy0,jy1);
 516             dz01             = _mm_sub_pd(iz0,jz1);
 517             dx02             = _mm_sub_pd(ix0,jx2);
 518             dy02             = _mm_sub_pd(iy0,jy2);
 519             dz02             = _mm_sub_pd(iz0,jz2);
 520             dx10             = _mm_sub_pd(ix1,jx0);
 521             dy10             = _mm_sub_pd(iy1,jy0);
 522             dz10             = _mm_sub_pd(iz1,jz0);
 523             dx11             = _mm_sub_pd(ix1,jx1);
 524             dy11             = _mm_sub_pd(iy1,jy1);
 525             dz11             = _mm_sub_pd(iz1,jz1);
 526             dx12             = _mm_sub_pd(ix1,jx2);
 527             dy12             = _mm_sub_pd(iy1,jy2);
 528             dz12             = _mm_sub_pd(iz1,jz2);
 529             dx20             = _mm_sub_pd(ix2,jx0);
 530             dy20             = _mm_sub_pd(iy2,jy0);
 531             dz20             = _mm_sub_pd(iz2,jz0);
 532             dx21             = _mm_sub_pd(ix2,jx1);
 533             dy21             = _mm_sub_pd(iy2,jy1);
 534             dz21             = _mm_sub_pd(iz2,jz1);
 535             dx22             = _mm_sub_pd(ix2,jx2);
 536             dy22             = _mm_sub_pd(iy2,jy2);
 537             dz22             = _mm_sub_pd(iz2,jz2);
 538
 539             /* Calculate squared distance and things based on it */
 540             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 541             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 542             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 543             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 544             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 545             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 546             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 547             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 548             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 549
 550             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 551             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 552             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 553             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 554             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 555             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 556             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 557             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 558             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 559
 560             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 561             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 562             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 563             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 564             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 565             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 566             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 567             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 568             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 569
 570             fjx0             = _mm_setzero_pd();
 571             fjy0             = _mm_setzero_pd();
 572             fjz0             = _mm_setzero_pd();
 573             fjx1             = _mm_setzero_pd();
 574             fjy1             = _mm_setzero_pd();
 575             fjz1             = _mm_setzero_pd();
 576             fjx2             = _mm_setzero_pd();
 577             fjy2             = _mm_setzero_pd();
 578             fjz2             = _mm_setzero_pd();
 579
 580             /**************************
 581              * CALCULATE INTERACTIONS *
 582              **************************/
 583
 584             /* REACTION-FIELD ELECTROSTATICS */
 585             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 586             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 587
 588             /* Update potential sum for this i atom from the interaction with this j atom. */
 589             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 590             velecsum         = _mm_add_pd(velecsum,velec);
 591
 592             fscal            = felec;
 593
 594             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 595
 596             /* Calculate temporary vectorial force */
 597             tx               = _mm_mul_pd(fscal,dx00);
 598             ty               = _mm_mul_pd(fscal,dy00);
 599             tz               = _mm_mul_pd(fscal,dz00);
 600
 601             /* Update vectorial force */
 602             fix0             = _mm_add_pd(fix0,tx);
 603             fiy0             = _mm_add_pd(fiy0,ty);
 604             fiz0             = _mm_add_pd(fiz0,tz);
 605
 606             fjx0             = _mm_add_pd(fjx0,tx);
 607             fjy0             = _mm_add_pd(fjy0,ty);
 608             fjz0             = _mm_add_pd(fjz0,tz);
 609
 610             /**************************
 611              * CALCULATE INTERACTIONS *
 612              **************************/
 613
 614             /* REACTION-FIELD ELECTROSTATICS */
 615             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 616             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 617
 618             /* Update potential sum for this i atom from the interaction with this j atom. */
 619             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 620             velecsum         = _mm_add_pd(velecsum,velec);
 621
 622             fscal            = felec;
 623
 624             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 625
 626             /* Calculate temporary vectorial force */
 627             tx               = _mm_mul_pd(fscal,dx01);
 628             ty               = _mm_mul_pd(fscal,dy01);
 629             tz               = _mm_mul_pd(fscal,dz01);
 630
 631             /* Update vectorial force */
 632             fix0             = _mm_add_pd(fix0,tx);
 633             fiy0             = _mm_add_pd(fiy0,ty);
 634             fiz0             = _mm_add_pd(fiz0,tz);
 635
 636             fjx1             = _mm_add_pd(fjx1,tx);
 637             fjy1             = _mm_add_pd(fjy1,ty);
 638             fjz1             = _mm_add_pd(fjz1,tz);
 639
 640             /**************************
 641              * CALCULATE INTERACTIONS *
 642              **************************/
 643
 644             /* REACTION-FIELD ELECTROSTATICS */
 645             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 646             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 647
 648             /* Update potential sum for this i atom from the interaction with this j atom. */
 649             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 650             velecsum         = _mm_add_pd(velecsum,velec);
 651
 652             fscal            = felec;
 653
 654             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 655
 656             /* Calculate temporary vectorial force */
 657             tx               = _mm_mul_pd(fscal,dx02);
 658             ty               = _mm_mul_pd(fscal,dy02);
 659             tz               = _mm_mul_pd(fscal,dz02);
 660
 661             /* Update vectorial force */
 662             fix0             = _mm_add_pd(fix0,tx);
 663             fiy0             = _mm_add_pd(fiy0,ty);
 664             fiz0             = _mm_add_pd(fiz0,tz);
 665
 666             fjx2             = _mm_add_pd(fjx2,tx);
 667             fjy2             = _mm_add_pd(fjy2,ty);
 668             fjz2             = _mm_add_pd(fjz2,tz);
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             /* REACTION-FIELD ELECTROSTATICS */
 675             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 676             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 677
 678             /* Update potential sum for this i atom from the interaction with this j atom. */
 679             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 680             velecsum         = _mm_add_pd(velecsum,velec);
 681
 682             fscal            = felec;
 683
 684             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 685
 686             /* Calculate temporary vectorial force */
 687             tx               = _mm_mul_pd(fscal,dx10);
 688             ty               = _mm_mul_pd(fscal,dy10);
 689             tz               = _mm_mul_pd(fscal,dz10);
 690
 691             /* Update vectorial force */
 692             fix1             = _mm_add_pd(fix1,tx);
 693             fiy1             = _mm_add_pd(fiy1,ty);
 694             fiz1             = _mm_add_pd(fiz1,tz);
 695
 696             fjx0             = _mm_add_pd(fjx0,tx);
 697             fjy0             = _mm_add_pd(fjy0,ty);
 698             fjz0             = _mm_add_pd(fjz0,tz);
 699
 700             /**************************
 701              * CALCULATE INTERACTIONS *
 702              **************************/
 703
 704             /* REACTION-FIELD ELECTROSTATICS */
 705             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 706             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 707
 708             /* Update potential sum for this i atom from the interaction with this j atom. */
 709             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 710             velecsum         = _mm_add_pd(velecsum,velec);
 711
 712             fscal            = felec;
 713
 714             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 715
 716             /* Calculate temporary vectorial force */
 717             tx               = _mm_mul_pd(fscal,dx11);
 718             ty               = _mm_mul_pd(fscal,dy11);
 719             tz               = _mm_mul_pd(fscal,dz11);
 720
 721             /* Update vectorial force */
 722             fix1             = _mm_add_pd(fix1,tx);
 723             fiy1             = _mm_add_pd(fiy1,ty);
 724             fiz1             = _mm_add_pd(fiz1,tz);
 725
 726             fjx1             = _mm_add_pd(fjx1,tx);
 727             fjy1             = _mm_add_pd(fjy1,ty);
 728             fjz1             = _mm_add_pd(fjz1,tz);
 729
 730             /**************************
 731              * CALCULATE INTERACTIONS *
 732              **************************/
 733
 734             /* REACTION-FIELD ELECTROSTATICS */
 735             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 736             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 740             velecsum         = _mm_add_pd(velecsum,velec);
 741
 742             fscal            = felec;
 743
 744             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 745
 746             /* Calculate temporary vectorial force */
 747             tx               = _mm_mul_pd(fscal,dx12);
 748             ty               = _mm_mul_pd(fscal,dy12);
 749             tz               = _mm_mul_pd(fscal,dz12);
 750
 751             /* Update vectorial force */
 752             fix1             = _mm_add_pd(fix1,tx);
 753             fiy1             = _mm_add_pd(fiy1,ty);
 754             fiz1             = _mm_add_pd(fiz1,tz);
 755
 756             fjx2             = _mm_add_pd(fjx2,tx);
 757             fjy2             = _mm_add_pd(fjy2,ty);
 758             fjz2             = _mm_add_pd(fjz2,tz);
 759
 760             /**************************
 761              * CALCULATE INTERACTIONS *
 762              **************************/
 763
 764             /* REACTION-FIELD ELECTROSTATICS */
 765             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 766             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 767
 768             /* Update potential sum for this i atom from the interaction with this j atom. */
 769             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 770             velecsum         = _mm_add_pd(velecsum,velec);
 771
 772             fscal            = felec;
 773
 774             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 775
 776             /* Calculate temporary vectorial force */
 777             tx               = _mm_mul_pd(fscal,dx20);
 778             ty               = _mm_mul_pd(fscal,dy20);
 779             tz               = _mm_mul_pd(fscal,dz20);
 780
 781             /* Update vectorial force */
 782             fix2             = _mm_add_pd(fix2,tx);
 783             fiy2             = _mm_add_pd(fiy2,ty);
 784             fiz2             = _mm_add_pd(fiz2,tz);
 785
 786             fjx0             = _mm_add_pd(fjx0,tx);
 787             fjy0             = _mm_add_pd(fjy0,ty);
 788             fjz0             = _mm_add_pd(fjz0,tz);
 789
 790             /**************************
 791              * CALCULATE INTERACTIONS *
 792              **************************/
 793
 794             /* REACTION-FIELD ELECTROSTATICS */
 795             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 796             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 797
 798             /* Update potential sum for this i atom from the interaction with this j atom. */
 799             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 800             velecsum         = _mm_add_pd(velecsum,velec);
 801
 802             fscal            = felec;
 803
 804             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 805
 806             /* Calculate temporary vectorial force */
 807             tx               = _mm_mul_pd(fscal,dx21);
 808             ty               = _mm_mul_pd(fscal,dy21);
 809             tz               = _mm_mul_pd(fscal,dz21);
 810
 811             /* Update vectorial force */
 812             fix2             = _mm_add_pd(fix2,tx);
 813             fiy2             = _mm_add_pd(fiy2,ty);
 814             fiz2             = _mm_add_pd(fiz2,tz);
 815
 816             fjx1             = _mm_add_pd(fjx1,tx);
 817             fjy1             = _mm_add_pd(fjy1,ty);
 818             fjz1             = _mm_add_pd(fjz1,tz);
 819
 820             /**************************
 821              * CALCULATE INTERACTIONS *
 822              **************************/
 823
 824             /* REACTION-FIELD ELECTROSTATICS */
 825             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 826             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 827
 828             /* Update potential sum for this i atom from the interaction with this j atom. */
 829             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 830             velecsum         = _mm_add_pd(velecsum,velec);
 831
 832             fscal            = felec;
 833
 834             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 835
 836             /* Calculate temporary vectorial force */
 837             tx               = _mm_mul_pd(fscal,dx22);
 838             ty               = _mm_mul_pd(fscal,dy22);
 839             tz               = _mm_mul_pd(fscal,dz22);
 840
 841             /* Update vectorial force */
 842             fix2             = _mm_add_pd(fix2,tx);
 843             fiy2             = _mm_add_pd(fiy2,ty);
 844             fiz2             = _mm_add_pd(fiz2,tz);
 845
 846             fjx2             = _mm_add_pd(fjx2,tx);
 847             fjy2             = _mm_add_pd(fjy2,ty);
 848             fjz2             = _mm_add_pd(fjz2,tz);
 849
 850             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 851
 852             /* Inner loop uses 288 flops */
 853         }
 854
 855         /* End of innermost loop */
 856
 857         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 858                                               f+i_coord_offset,fshift+i_shift_offset);
 859
 860         ggid                        = gid[iidx];
 861         /* Update potential energies */
 862         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 863
 864         /* Increment number of inner iterations */
 865         inneriter                  += j_index_end - j_index_start;
 866
 867         /* Outer loop uses 19 flops */
 868     }
 869
 870     /* Increment number of outer iterations */
 871     outeriter        += nri;
 872
 873     /* Update outer/inner flops */
 874
 875     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*288);
 876 }
 877 /*
 878  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse4_1_double
 879  * Electrostatics interaction: ReactionField
 880  * VdW interaction:            None
 881  * Geometry:                   Water3-Water3
 882  * Calculate force/pot:        Force
 883  */
 884 void
 885 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse4_1_double
 886                     (t_nblist * gmx_restrict                nlist,
 887                      rvec * gmx_restrict                    xx,
 888                      rvec * gmx_restrict                    ff,
 889                      t_forcerec * gmx_restrict              fr,
 890                      t_mdatoms * gmx_restrict               mdatoms,
 891                      nb_kernel_data_t * gmx_restrict        kernel_data,
 892                      t_nrnb * gmx_restrict                  nrnb)
 893 {
 894     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 895      * just 0 for non-waters.
 896      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 897      * jnr indices corresponding to data put in the four positions in the SIMD register.
 898      */
 899     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 900     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 901     int              jnrA,jnrB;
 902     int              j_coord_offsetA,j_coord_offsetB;
 903     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 904     real             rcutoff_scalar;
 905     real             *shiftvec,*fshift,*x,*f;
 906     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 907     int              vdwioffset0;
 908     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 909     int              vdwioffset1;
 910     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 911     int              vdwioffset2;
 912     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 913     int              vdwjidx0A,vdwjidx0B;
 914     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 915     int              vdwjidx1A,vdwjidx1B;
 916     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 917     int              vdwjidx2A,vdwjidx2B;
 918     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 919     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 920     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 921     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 922     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 923     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 924     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 925     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 926     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 927     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 928     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 929     real             *charge;
 930     __m128d          dummy_mask,cutoff_mask;
 931     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 932     __m128d          one     = _mm_set1_pd(1.0);
 933     __m128d          two     = _mm_set1_pd(2.0);
 934     x                = xx[0];
 935     f                = ff[0];
 936
 937     nri              = nlist->nri;
 938     iinr             = nlist->iinr;
 939     jindex           = nlist->jindex;
 940     jjnr             = nlist->jjnr;
 941     shiftidx         = nlist->shift;
 942     gid              = nlist->gid;
 943     shiftvec         = fr->shift_vec[0];
 944     fshift           = fr->fshift[0];
 945     facel            = _mm_set1_pd(fr->epsfac);
 946     charge           = mdatoms->chargeA;
 947     krf              = _mm_set1_pd(fr->ic->k_rf);
 948     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 949     crf              = _mm_set1_pd(fr->ic->c_rf);
 950
 951     /* Setup water-specific parameters */
 952     inr              = nlist->iinr[0];
 953     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 954     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 955     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 956
 957     jq0              = _mm_set1_pd(charge[inr+0]);
 958     jq1              = _mm_set1_pd(charge[inr+1]);
 959     jq2              = _mm_set1_pd(charge[inr+2]);
 960     qq00             = _mm_mul_pd(iq0,jq0);
 961     qq01             = _mm_mul_pd(iq0,jq1);
 962     qq02             = _mm_mul_pd(iq0,jq2);
 963     qq10             = _mm_mul_pd(iq1,jq0);
 964     qq11             = _mm_mul_pd(iq1,jq1);
 965     qq12             = _mm_mul_pd(iq1,jq2);
 966     qq20             = _mm_mul_pd(iq2,jq0);
 967     qq21             = _mm_mul_pd(iq2,jq1);
 968     qq22             = _mm_mul_pd(iq2,jq2);
 969
 970     /* Avoid stupid compiler warnings */
 971     jnrA = jnrB = 0;
 972     j_coord_offsetA = 0;
 973     j_coord_offsetB = 0;
 974
 975     outeriter        = 0;
 976     inneriter        = 0;
 977
 978     /* Start outer loop over neighborlists */
 979     for(iidx=0; iidx<nri; iidx++)
 980     {
 981         /* Load shift vector for this list */
 982         i_shift_offset   = DIM*shiftidx[iidx];
 983
 984         /* Load limits for loop over neighbors */
 985         j_index_start    = jindex[iidx];
 986         j_index_end      = jindex[iidx+1];
 987
 988         /* Get outer coordinate index */
 989         inr              = iinr[iidx];
 990         i_coord_offset   = DIM*inr;
 991
 992         /* Load i particle coords and add shift vector */
 993         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 994                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 995
 996         fix0             = _mm_setzero_pd();
 997         fiy0             = _mm_setzero_pd();
 998         fiz0             = _mm_setzero_pd();
 999         fix1             = _mm_setzero_pd();
1000         fiy1             = _mm_setzero_pd();
1001         fiz1             = _mm_setzero_pd();
1002         fix2             = _mm_setzero_pd();
1003         fiy2             = _mm_setzero_pd();
1004         fiz2             = _mm_setzero_pd();
1005
1006         /* Start inner kernel loop */
1007         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1008         {
1009
1010             /* Get j neighbor index, and coordinate index */
1011             jnrA             = jjnr[jidx];
1012             jnrB             = jjnr[jidx+1];
1013             j_coord_offsetA  = DIM*jnrA;
1014             j_coord_offsetB  = DIM*jnrB;
1015
1016             /* load j atom coordinates */
1017             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1018                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1019
1020             /* Calculate displacement vector */
1021             dx00             = _mm_sub_pd(ix0,jx0);
1022             dy00             = _mm_sub_pd(iy0,jy0);
1023             dz00             = _mm_sub_pd(iz0,jz0);
1024             dx01             = _mm_sub_pd(ix0,jx1);
1025             dy01             = _mm_sub_pd(iy0,jy1);
1026             dz01             = _mm_sub_pd(iz0,jz1);
1027             dx02             = _mm_sub_pd(ix0,jx2);
1028             dy02             = _mm_sub_pd(iy0,jy2);
1029             dz02             = _mm_sub_pd(iz0,jz2);
1030             dx10             = _mm_sub_pd(ix1,jx0);
1031             dy10             = _mm_sub_pd(iy1,jy0);
1032             dz10             = _mm_sub_pd(iz1,jz0);
1033             dx11             = _mm_sub_pd(ix1,jx1);
1034             dy11             = _mm_sub_pd(iy1,jy1);
1035             dz11             = _mm_sub_pd(iz1,jz1);
1036             dx12             = _mm_sub_pd(ix1,jx2);
1037             dy12             = _mm_sub_pd(iy1,jy2);
1038             dz12             = _mm_sub_pd(iz1,jz2);
1039             dx20             = _mm_sub_pd(ix2,jx0);
1040             dy20             = _mm_sub_pd(iy2,jy0);
1041             dz20             = _mm_sub_pd(iz2,jz0);
1042             dx21             = _mm_sub_pd(ix2,jx1);
1043             dy21             = _mm_sub_pd(iy2,jy1);
1044             dz21             = _mm_sub_pd(iz2,jz1);
1045             dx22             = _mm_sub_pd(ix2,jx2);
1046             dy22             = _mm_sub_pd(iy2,jy2);
1047             dz22             = _mm_sub_pd(iz2,jz2);
1048
1049             /* Calculate squared distance and things based on it */
1050             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1051             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1052             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1053             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1054             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1055             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1056             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1057             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1058             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1059
1060             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1061             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1062             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1063             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1064             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1065             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1066             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1067             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1068             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1069
1070             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1071             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1072             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1073             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1074             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1075             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1076             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1077             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1078             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1079
1080             fjx0             = _mm_setzero_pd();
1081             fjy0             = _mm_setzero_pd();
1082             fjz0             = _mm_setzero_pd();
1083             fjx1             = _mm_setzero_pd();
1084             fjy1             = _mm_setzero_pd();
1085             fjz1             = _mm_setzero_pd();
1086             fjx2             = _mm_setzero_pd();
1087             fjy2             = _mm_setzero_pd();
1088             fjz2             = _mm_setzero_pd();
1089
1090             /**************************
1091              * CALCULATE INTERACTIONS *
1092              **************************/
1093
1094             /* REACTION-FIELD ELECTROSTATICS */
1095             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1096
1097             fscal            = felec;
1098
1099             /* Calculate temporary vectorial force */
1100             tx               = _mm_mul_pd(fscal,dx00);
1101             ty               = _mm_mul_pd(fscal,dy00);
1102             tz               = _mm_mul_pd(fscal,dz00);
1103
1104             /* Update vectorial force */
1105             fix0             = _mm_add_pd(fix0,tx);
1106             fiy0             = _mm_add_pd(fiy0,ty);
1107             fiz0             = _mm_add_pd(fiz0,tz);
1108
1109             fjx0             = _mm_add_pd(fjx0,tx);
1110             fjy0             = _mm_add_pd(fjy0,ty);
1111             fjz0             = _mm_add_pd(fjz0,tz);
1112
1113             /**************************
1114              * CALCULATE INTERACTIONS *
1115              **************************/
1116
1117             /* REACTION-FIELD ELECTROSTATICS */
1118             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1119
1120             fscal            = felec;
1121
1122             /* Calculate temporary vectorial force */
1123             tx               = _mm_mul_pd(fscal,dx01);
1124             ty               = _mm_mul_pd(fscal,dy01);
1125             tz               = _mm_mul_pd(fscal,dz01);
1126
1127             /* Update vectorial force */
1128             fix0             = _mm_add_pd(fix0,tx);
1129             fiy0             = _mm_add_pd(fiy0,ty);
1130             fiz0             = _mm_add_pd(fiz0,tz);
1131
1132             fjx1             = _mm_add_pd(fjx1,tx);
1133             fjy1             = _mm_add_pd(fjy1,ty);
1134             fjz1             = _mm_add_pd(fjz1,tz);
1135
1136             /**************************
1137              * CALCULATE INTERACTIONS *
1138              **************************/
1139
1140             /* REACTION-FIELD ELECTROSTATICS */
1141             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1142
1143             fscal            = felec;
1144
1145             /* Calculate temporary vectorial force */
1146             tx               = _mm_mul_pd(fscal,dx02);
1147             ty               = _mm_mul_pd(fscal,dy02);
1148             tz               = _mm_mul_pd(fscal,dz02);
1149
1150             /* Update vectorial force */
1151             fix0             = _mm_add_pd(fix0,tx);
1152             fiy0             = _mm_add_pd(fiy0,ty);
1153             fiz0             = _mm_add_pd(fiz0,tz);
1154
1155             fjx2             = _mm_add_pd(fjx2,tx);
1156             fjy2             = _mm_add_pd(fjy2,ty);
1157             fjz2             = _mm_add_pd(fjz2,tz);
1158
1159             /**************************
1160              * CALCULATE INTERACTIONS *
1161              **************************/
1162
1163             /* REACTION-FIELD ELECTROSTATICS */
1164             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1165
1166             fscal            = felec;
1167
1168             /* Calculate temporary vectorial force */
1169             tx               = _mm_mul_pd(fscal,dx10);
1170             ty               = _mm_mul_pd(fscal,dy10);
1171             tz               = _mm_mul_pd(fscal,dz10);
1172
1173             /* Update vectorial force */
1174             fix1             = _mm_add_pd(fix1,tx);
1175             fiy1             = _mm_add_pd(fiy1,ty);
1176             fiz1             = _mm_add_pd(fiz1,tz);
1177
1178             fjx0             = _mm_add_pd(fjx0,tx);
1179             fjy0             = _mm_add_pd(fjy0,ty);
1180             fjz0             = _mm_add_pd(fjz0,tz);
1181
1182             /**************************
1183              * CALCULATE INTERACTIONS *
1184              **************************/
1185
1186             /* REACTION-FIELD ELECTROSTATICS */
1187             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1188
1189             fscal            = felec;
1190
1191             /* Calculate temporary vectorial force */
1192             tx               = _mm_mul_pd(fscal,dx11);
1193             ty               = _mm_mul_pd(fscal,dy11);
1194             tz               = _mm_mul_pd(fscal,dz11);
1195
1196             /* Update vectorial force */
1197             fix1             = _mm_add_pd(fix1,tx);
1198             fiy1             = _mm_add_pd(fiy1,ty);
1199             fiz1             = _mm_add_pd(fiz1,tz);
1200
1201             fjx1             = _mm_add_pd(fjx1,tx);
1202             fjy1             = _mm_add_pd(fjy1,ty);
1203             fjz1             = _mm_add_pd(fjz1,tz);
1204
1205             /**************************
1206              * CALCULATE INTERACTIONS *
1207              **************************/
1208
1209             /* REACTION-FIELD ELECTROSTATICS */
1210             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1211
1212             fscal            = felec;
1213
1214             /* Calculate temporary vectorial force */
1215             tx               = _mm_mul_pd(fscal,dx12);
1216             ty               = _mm_mul_pd(fscal,dy12);
1217             tz               = _mm_mul_pd(fscal,dz12);
1218
1219             /* Update vectorial force */
1220             fix1             = _mm_add_pd(fix1,tx);
1221             fiy1             = _mm_add_pd(fiy1,ty);
1222             fiz1             = _mm_add_pd(fiz1,tz);
1223
1224             fjx2             = _mm_add_pd(fjx2,tx);
1225             fjy2             = _mm_add_pd(fjy2,ty);
1226             fjz2             = _mm_add_pd(fjz2,tz);
1227
1228             /**************************
1229              * CALCULATE INTERACTIONS *
1230              **************************/
1231
1232             /* REACTION-FIELD ELECTROSTATICS */
1233             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1234
1235             fscal            = felec;
1236
1237             /* Calculate temporary vectorial force */
1238             tx               = _mm_mul_pd(fscal,dx20);
1239             ty               = _mm_mul_pd(fscal,dy20);
1240             tz               = _mm_mul_pd(fscal,dz20);
1241
1242             /* Update vectorial force */
1243             fix2             = _mm_add_pd(fix2,tx);
1244             fiy2             = _mm_add_pd(fiy2,ty);
1245             fiz2             = _mm_add_pd(fiz2,tz);
1246
1247             fjx0             = _mm_add_pd(fjx0,tx);
1248             fjy0             = _mm_add_pd(fjy0,ty);
1249             fjz0             = _mm_add_pd(fjz0,tz);
1250
1251             /**************************
1252              * CALCULATE INTERACTIONS *
1253              **************************/
1254
1255             /* REACTION-FIELD ELECTROSTATICS */
1256             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1257
1258             fscal            = felec;
1259
1260             /* Calculate temporary vectorial force */
1261             tx               = _mm_mul_pd(fscal,dx21);
1262             ty               = _mm_mul_pd(fscal,dy21);
1263             tz               = _mm_mul_pd(fscal,dz21);
1264
1265             /* Update vectorial force */
1266             fix2             = _mm_add_pd(fix2,tx);
1267             fiy2             = _mm_add_pd(fiy2,ty);
1268             fiz2             = _mm_add_pd(fiz2,tz);
1269
1270             fjx1             = _mm_add_pd(fjx1,tx);
1271             fjy1             = _mm_add_pd(fjy1,ty);
1272             fjz1             = _mm_add_pd(fjz1,tz);
1273
1274             /**************************
1275              * CALCULATE INTERACTIONS *
1276              **************************/
1277
1278             /* REACTION-FIELD ELECTROSTATICS */
1279             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1280
1281             fscal            = felec;
1282
1283             /* Calculate temporary vectorial force */
1284             tx               = _mm_mul_pd(fscal,dx22);
1285             ty               = _mm_mul_pd(fscal,dy22);
1286             tz               = _mm_mul_pd(fscal,dz22);
1287
1288             /* Update vectorial force */
1289             fix2             = _mm_add_pd(fix2,tx);
1290             fiy2             = _mm_add_pd(fiy2,ty);
1291             fiz2             = _mm_add_pd(fiz2,tz);
1292
1293             fjx2             = _mm_add_pd(fjx2,tx);
1294             fjy2             = _mm_add_pd(fjy2,ty);
1295             fjz2             = _mm_add_pd(fjz2,tz);
1296
1297             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1298
1299             /* Inner loop uses 243 flops */
1300         }
1301
1302         if(jidx<j_index_end)
1303         {
1304
1305             jnrA             = jjnr[jidx];
1306             j_coord_offsetA  = DIM*jnrA;
1307
1308             /* load j atom coordinates */
1309             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1310                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1311
1312             /* Calculate displacement vector */
1313             dx00             = _mm_sub_pd(ix0,jx0);
1314             dy00             = _mm_sub_pd(iy0,jy0);
1315             dz00             = _mm_sub_pd(iz0,jz0);
1316             dx01             = _mm_sub_pd(ix0,jx1);
1317             dy01             = _mm_sub_pd(iy0,jy1);
1318             dz01             = _mm_sub_pd(iz0,jz1);
1319             dx02             = _mm_sub_pd(ix0,jx2);
1320             dy02             = _mm_sub_pd(iy0,jy2);
1321             dz02             = _mm_sub_pd(iz0,jz2);
1322             dx10             = _mm_sub_pd(ix1,jx0);
1323             dy10             = _mm_sub_pd(iy1,jy0);
1324             dz10             = _mm_sub_pd(iz1,jz0);
1325             dx11             = _mm_sub_pd(ix1,jx1);
1326             dy11             = _mm_sub_pd(iy1,jy1);
1327             dz11             = _mm_sub_pd(iz1,jz1);
1328             dx12             = _mm_sub_pd(ix1,jx2);
1329             dy12             = _mm_sub_pd(iy1,jy2);
1330             dz12             = _mm_sub_pd(iz1,jz2);
1331             dx20             = _mm_sub_pd(ix2,jx0);
1332             dy20             = _mm_sub_pd(iy2,jy0);
1333             dz20             = _mm_sub_pd(iz2,jz0);
1334             dx21             = _mm_sub_pd(ix2,jx1);
1335             dy21             = _mm_sub_pd(iy2,jy1);
1336             dz21             = _mm_sub_pd(iz2,jz1);
1337             dx22             = _mm_sub_pd(ix2,jx2);
1338             dy22             = _mm_sub_pd(iy2,jy2);
1339             dz22             = _mm_sub_pd(iz2,jz2);
1340
1341             /* Calculate squared distance and things based on it */
1342             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1343             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1344             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1345             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1346             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1347             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1348             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1349             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1350             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1351
1352             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1353             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1354             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1355             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1356             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1357             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1358             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1359             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1360             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1361
1362             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1363             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1364             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1365             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1366             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1367             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1368             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1369             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1370             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1371
1372             fjx0             = _mm_setzero_pd();
1373             fjy0             = _mm_setzero_pd();
1374             fjz0             = _mm_setzero_pd();
1375             fjx1             = _mm_setzero_pd();
1376             fjy1             = _mm_setzero_pd();
1377             fjz1             = _mm_setzero_pd();
1378             fjx2             = _mm_setzero_pd();
1379             fjy2             = _mm_setzero_pd();
1380             fjz2             = _mm_setzero_pd();
1381
1382             /**************************
1383              * CALCULATE INTERACTIONS *
1384              **************************/
1385
1386             /* REACTION-FIELD ELECTROSTATICS */
1387             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1388
1389             fscal            = felec;
1390
1391             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1392
1393             /* Calculate temporary vectorial force */
1394             tx               = _mm_mul_pd(fscal,dx00);
1395             ty               = _mm_mul_pd(fscal,dy00);
1396             tz               = _mm_mul_pd(fscal,dz00);
1397
1398             /* Update vectorial force */
1399             fix0             = _mm_add_pd(fix0,tx);
1400             fiy0             = _mm_add_pd(fiy0,ty);
1401             fiz0             = _mm_add_pd(fiz0,tz);
1402
1403             fjx0             = _mm_add_pd(fjx0,tx);
1404             fjy0             = _mm_add_pd(fjy0,ty);
1405             fjz0             = _mm_add_pd(fjz0,tz);
1406
1407             /**************************
1408              * CALCULATE INTERACTIONS *
1409              **************************/
1410
1411             /* REACTION-FIELD ELECTROSTATICS */
1412             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1413
1414             fscal            = felec;
1415
1416             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1417
1418             /* Calculate temporary vectorial force */
1419             tx               = _mm_mul_pd(fscal,dx01);
1420             ty               = _mm_mul_pd(fscal,dy01);
1421             tz               = _mm_mul_pd(fscal,dz01);
1422
1423             /* Update vectorial force */
1424             fix0             = _mm_add_pd(fix0,tx);
1425             fiy0             = _mm_add_pd(fiy0,ty);
1426             fiz0             = _mm_add_pd(fiz0,tz);
1427
1428             fjx1             = _mm_add_pd(fjx1,tx);
1429             fjy1             = _mm_add_pd(fjy1,ty);
1430             fjz1             = _mm_add_pd(fjz1,tz);
1431
1432             /**************************
1433              * CALCULATE INTERACTIONS *
1434              **************************/
1435
1436             /* REACTION-FIELD ELECTROSTATICS */
1437             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1438
1439             fscal            = felec;
1440
1441             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1442
1443             /* Calculate temporary vectorial force */
1444             tx               = _mm_mul_pd(fscal,dx02);
1445             ty               = _mm_mul_pd(fscal,dy02);
1446             tz               = _mm_mul_pd(fscal,dz02);
1447
1448             /* Update vectorial force */
1449             fix0             = _mm_add_pd(fix0,tx);
1450             fiy0             = _mm_add_pd(fiy0,ty);
1451             fiz0             = _mm_add_pd(fiz0,tz);
1452
1453             fjx2             = _mm_add_pd(fjx2,tx);
1454             fjy2             = _mm_add_pd(fjy2,ty);
1455             fjz2             = _mm_add_pd(fjz2,tz);
1456
1457             /**************************
1458              * CALCULATE INTERACTIONS *
1459              **************************/
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1463
1464             fscal            = felec;
1465
1466             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1467
1468             /* Calculate temporary vectorial force */
1469             tx               = _mm_mul_pd(fscal,dx10);
1470             ty               = _mm_mul_pd(fscal,dy10);
1471             tz               = _mm_mul_pd(fscal,dz10);
1472
1473             /* Update vectorial force */
1474             fix1             = _mm_add_pd(fix1,tx);
1475             fiy1             = _mm_add_pd(fiy1,ty);
1476             fiz1             = _mm_add_pd(fiz1,tz);
1477
1478             fjx0             = _mm_add_pd(fjx0,tx);
1479             fjy0             = _mm_add_pd(fjy0,ty);
1480             fjz0             = _mm_add_pd(fjz0,tz);
1481
1482             /**************************
1483              * CALCULATE INTERACTIONS *
1484              **************************/
1485
1486             /* REACTION-FIELD ELECTROSTATICS */
1487             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1488
1489             fscal            = felec;
1490
1491             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1492
1493             /* Calculate temporary vectorial force */
1494             tx               = _mm_mul_pd(fscal,dx11);
1495             ty               = _mm_mul_pd(fscal,dy11);
1496             tz               = _mm_mul_pd(fscal,dz11);
1497
1498             /* Update vectorial force */
1499             fix1             = _mm_add_pd(fix1,tx);
1500             fiy1             = _mm_add_pd(fiy1,ty);
1501             fiz1             = _mm_add_pd(fiz1,tz);
1502
1503             fjx1             = _mm_add_pd(fjx1,tx);
1504             fjy1             = _mm_add_pd(fjy1,ty);
1505             fjz1             = _mm_add_pd(fjz1,tz);
1506
1507             /**************************
1508              * CALCULATE INTERACTIONS *
1509              **************************/
1510
1511             /* REACTION-FIELD ELECTROSTATICS */
1512             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1513
1514             fscal            = felec;
1515
1516             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1517
1518             /* Calculate temporary vectorial force */
1519             tx               = _mm_mul_pd(fscal,dx12);
1520             ty               = _mm_mul_pd(fscal,dy12);
1521             tz               = _mm_mul_pd(fscal,dz12);
1522
1523             /* Update vectorial force */
1524             fix1             = _mm_add_pd(fix1,tx);
1525             fiy1             = _mm_add_pd(fiy1,ty);
1526             fiz1             = _mm_add_pd(fiz1,tz);
1527
1528             fjx2             = _mm_add_pd(fjx2,tx);
1529             fjy2             = _mm_add_pd(fjy2,ty);
1530             fjz2             = _mm_add_pd(fjz2,tz);
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             /* REACTION-FIELD ELECTROSTATICS */
1537             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1538
1539             fscal            = felec;
1540
1541             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1542
1543             /* Calculate temporary vectorial force */
1544             tx               = _mm_mul_pd(fscal,dx20);
1545             ty               = _mm_mul_pd(fscal,dy20);
1546             tz               = _mm_mul_pd(fscal,dz20);
1547
1548             /* Update vectorial force */
1549             fix2             = _mm_add_pd(fix2,tx);
1550             fiy2             = _mm_add_pd(fiy2,ty);
1551             fiz2             = _mm_add_pd(fiz2,tz);
1552
1553             fjx0             = _mm_add_pd(fjx0,tx);
1554             fjy0             = _mm_add_pd(fjy0,ty);
1555             fjz0             = _mm_add_pd(fjz0,tz);
1556
1557             /**************************
1558              * CALCULATE INTERACTIONS *
1559              **************************/
1560
1561             /* REACTION-FIELD ELECTROSTATICS */
1562             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1563
1564             fscal            = felec;
1565
1566             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1567
1568             /* Calculate temporary vectorial force */
1569             tx               = _mm_mul_pd(fscal,dx21);
1570             ty               = _mm_mul_pd(fscal,dy21);
1571             tz               = _mm_mul_pd(fscal,dz21);
1572
1573             /* Update vectorial force */
1574             fix2             = _mm_add_pd(fix2,tx);
1575             fiy2             = _mm_add_pd(fiy2,ty);
1576             fiz2             = _mm_add_pd(fiz2,tz);
1577
1578             fjx1             = _mm_add_pd(fjx1,tx);
1579             fjy1             = _mm_add_pd(fjy1,ty);
1580             fjz1             = _mm_add_pd(fjz1,tz);
1581
1582             /**************************
1583              * CALCULATE INTERACTIONS *
1584              **************************/
1585
1586             /* REACTION-FIELD ELECTROSTATICS */
1587             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1588
1589             fscal            = felec;
1590
1591             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1592
1593             /* Calculate temporary vectorial force */
1594             tx               = _mm_mul_pd(fscal,dx22);
1595             ty               = _mm_mul_pd(fscal,dy22);
1596             tz               = _mm_mul_pd(fscal,dz22);
1597
1598             /* Update vectorial force */
1599             fix2             = _mm_add_pd(fix2,tx);
1600             fiy2             = _mm_add_pd(fiy2,ty);
1601             fiz2             = _mm_add_pd(fiz2,tz);
1602
1603             fjx2             = _mm_add_pd(fjx2,tx);
1604             fjy2             = _mm_add_pd(fjy2,ty);
1605             fjz2             = _mm_add_pd(fjz2,tz);
1606
1607             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1608
1609             /* Inner loop uses 243 flops */
1610         }
1611
1612         /* End of innermost loop */
1613
1614         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1615                                               f+i_coord_offset,fshift+i_shift_offset);
1616
1617         /* Increment number of inner iterations */
1618         inneriter                  += j_index_end - j_index_start;
1619
1620         /* Outer loop uses 18 flops */
1621     }
1622
1623     /* Increment number of outer iterations */
1624     outeriter        += nri;
1625
1626     /* Update outer/inner flops */
1627
1628     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1629 }