src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwLJSh_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_ElecRFCut_VdwLJSh_GeomW3W3_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSh_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     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     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 143     rcutoff_scalar   = fr->rcoulomb;
 144     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 145     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 146
 147     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 148     rvdw             = _mm_set1_pd(fr->rvdw);
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154
 155     outeriter        = 0;
 156     inneriter        = 0;
 157
 158     /* Start outer loop over neighborlists */
 159     for(iidx=0; iidx<nri; iidx++)
 160     {
 161         /* Load shift vector for this list */
 162         i_shift_offset   = DIM*shiftidx[iidx];
 163
 164         /* Load limits for loop over neighbors */
 165         j_index_start    = jindex[iidx];
 166         j_index_end      = jindex[iidx+1];
 167
 168         /* Get outer coordinate index */
 169         inr              = iinr[iidx];
 170         i_coord_offset   = DIM*inr;
 171
 172         /* Load i particle coords and add shift vector */
 173         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 174                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 175
 176         fix0             = _mm_setzero_pd();
 177         fiy0             = _mm_setzero_pd();
 178         fiz0             = _mm_setzero_pd();
 179         fix1             = _mm_setzero_pd();
 180         fiy1             = _mm_setzero_pd();
 181         fiz1             = _mm_setzero_pd();
 182         fix2             = _mm_setzero_pd();
 183         fiy2             = _mm_setzero_pd();
 184         fiz2             = _mm_setzero_pd();
 185
 186         /* Reset potential sums */
 187         velecsum         = _mm_setzero_pd();
 188         vvdwsum          = _mm_setzero_pd();
 189
 190         /* Start inner kernel loop */
 191         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 192         {
 193
 194             /* Get j neighbor index, and coordinate index */
 195             jnrA             = jjnr[jidx];
 196             jnrB             = jjnr[jidx+1];
 197             j_coord_offsetA  = DIM*jnrA;
 198             j_coord_offsetB  = DIM*jnrB;
 199
 200             /* load j atom coordinates */
 201             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 202                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 203
 204             /* Calculate displacement vector */
 205             dx00             = _mm_sub_pd(ix0,jx0);
 206             dy00             = _mm_sub_pd(iy0,jy0);
 207             dz00             = _mm_sub_pd(iz0,jz0);
 208             dx01             = _mm_sub_pd(ix0,jx1);
 209             dy01             = _mm_sub_pd(iy0,jy1);
 210             dz01             = _mm_sub_pd(iz0,jz1);
 211             dx02             = _mm_sub_pd(ix0,jx2);
 212             dy02             = _mm_sub_pd(iy0,jy2);
 213             dz02             = _mm_sub_pd(iz0,jz2);
 214             dx10             = _mm_sub_pd(ix1,jx0);
 215             dy10             = _mm_sub_pd(iy1,jy0);
 216             dz10             = _mm_sub_pd(iz1,jz0);
 217             dx11             = _mm_sub_pd(ix1,jx1);
 218             dy11             = _mm_sub_pd(iy1,jy1);
 219             dz11             = _mm_sub_pd(iz1,jz1);
 220             dx12             = _mm_sub_pd(ix1,jx2);
 221             dy12             = _mm_sub_pd(iy1,jy2);
 222             dz12             = _mm_sub_pd(iz1,jz2);
 223             dx20             = _mm_sub_pd(ix2,jx0);
 224             dy20             = _mm_sub_pd(iy2,jy0);
 225             dz20             = _mm_sub_pd(iz2,jz0);
 226             dx21             = _mm_sub_pd(ix2,jx1);
 227             dy21             = _mm_sub_pd(iy2,jy1);
 228             dz21             = _mm_sub_pd(iz2,jz1);
 229             dx22             = _mm_sub_pd(ix2,jx2);
 230             dy22             = _mm_sub_pd(iy2,jy2);
 231             dz22             = _mm_sub_pd(iz2,jz2);
 232
 233             /* Calculate squared distance and things based on it */
 234             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 235             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 236             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 237             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 238             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 239             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 240             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 241             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 242             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 243
 244             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 245             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 246             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 247             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 248             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 249             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 250             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 251             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 252             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 253
 254             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 255             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 256             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 257             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 258             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 259             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 260             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 261             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 262             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 263
 264             fjx0             = _mm_setzero_pd();
 265             fjy0             = _mm_setzero_pd();
 266             fjz0             = _mm_setzero_pd();
 267             fjx1             = _mm_setzero_pd();
 268             fjy1             = _mm_setzero_pd();
 269             fjz1             = _mm_setzero_pd();
 270             fjx2             = _mm_setzero_pd();
 271             fjy2             = _mm_setzero_pd();
 272             fjz2             = _mm_setzero_pd();
 273
 274             /**************************
 275              * CALCULATE INTERACTIONS *
 276              **************************/
 277
 278             if (gmx_mm_any_lt(rsq00,rcutoff2))
 279             {
 280
 281             /* REACTION-FIELD ELECTROSTATICS */
 282             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 283             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 284
 285             /* LENNARD-JONES DISPERSION/REPULSION */
 286
 287             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 288             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 289             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 290             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 291                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 292             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 293
 294             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 295
 296             /* Update potential sum for this i atom from the interaction with this j atom. */
 297             velec            = _mm_and_pd(velec,cutoff_mask);
 298             velecsum         = _mm_add_pd(velecsum,velec);
 299             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 300             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 301
 302             fscal            = _mm_add_pd(felec,fvdw);
 303
 304             fscal            = _mm_and_pd(fscal,cutoff_mask);
 305
 306             /* Calculate temporary vectorial force */
 307             tx               = _mm_mul_pd(fscal,dx00);
 308             ty               = _mm_mul_pd(fscal,dy00);
 309             tz               = _mm_mul_pd(fscal,dz00);
 310
 311             /* Update vectorial force */
 312             fix0             = _mm_add_pd(fix0,tx);
 313             fiy0             = _mm_add_pd(fiy0,ty);
 314             fiz0             = _mm_add_pd(fiz0,tz);
 315
 316             fjx0             = _mm_add_pd(fjx0,tx);
 317             fjy0             = _mm_add_pd(fjy0,ty);
 318             fjz0             = _mm_add_pd(fjz0,tz);
 319
 320             }
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             if (gmx_mm_any_lt(rsq01,rcutoff2))
 327             {
 328
 329             /* REACTION-FIELD ELECTROSTATICS */
 330             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 331             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 332
 333             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             velec            = _mm_and_pd(velec,cutoff_mask);
 337             velecsum         = _mm_add_pd(velecsum,velec);
 338
 339             fscal            = felec;
 340
 341             fscal            = _mm_and_pd(fscal,cutoff_mask);
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = _mm_mul_pd(fscal,dx01);
 345             ty               = _mm_mul_pd(fscal,dy01);
 346             tz               = _mm_mul_pd(fscal,dz01);
 347
 348             /* Update vectorial force */
 349             fix0             = _mm_add_pd(fix0,tx);
 350             fiy0             = _mm_add_pd(fiy0,ty);
 351             fiz0             = _mm_add_pd(fiz0,tz);
 352
 353             fjx1             = _mm_add_pd(fjx1,tx);
 354             fjy1             = _mm_add_pd(fjy1,ty);
 355             fjz1             = _mm_add_pd(fjz1,tz);
 356
 357             }
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             if (gmx_mm_any_lt(rsq02,rcutoff2))
 364             {
 365
 366             /* REACTION-FIELD ELECTROSTATICS */
 367             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 368             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 369
 370             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velec            = _mm_and_pd(velec,cutoff_mask);
 374             velecsum         = _mm_add_pd(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm_and_pd(fscal,cutoff_mask);
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = _mm_mul_pd(fscal,dx02);
 382             ty               = _mm_mul_pd(fscal,dy02);
 383             tz               = _mm_mul_pd(fscal,dz02);
 384
 385             /* Update vectorial force */
 386             fix0             = _mm_add_pd(fix0,tx);
 387             fiy0             = _mm_add_pd(fiy0,ty);
 388             fiz0             = _mm_add_pd(fiz0,tz);
 389
 390             fjx2             = _mm_add_pd(fjx2,tx);
 391             fjy2             = _mm_add_pd(fjy2,ty);
 392             fjz2             = _mm_add_pd(fjz2,tz);
 393
 394             }
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             if (gmx_mm_any_lt(rsq10,rcutoff2))
 401             {
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 405             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 406
 407             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 408
 409             /* Update potential sum for this i atom from the interaction with this j atom. */
 410             velec            = _mm_and_pd(velec,cutoff_mask);
 411             velecsum         = _mm_add_pd(velecsum,velec);
 412
 413             fscal            = felec;
 414
 415             fscal            = _mm_and_pd(fscal,cutoff_mask);
 416
 417             /* Calculate temporary vectorial force */
 418             tx               = _mm_mul_pd(fscal,dx10);
 419             ty               = _mm_mul_pd(fscal,dy10);
 420             tz               = _mm_mul_pd(fscal,dz10);
 421
 422             /* Update vectorial force */
 423             fix1             = _mm_add_pd(fix1,tx);
 424             fiy1             = _mm_add_pd(fiy1,ty);
 425             fiz1             = _mm_add_pd(fiz1,tz);
 426
 427             fjx0             = _mm_add_pd(fjx0,tx);
 428             fjy0             = _mm_add_pd(fjy0,ty);
 429             fjz0             = _mm_add_pd(fjz0,tz);
 430
 431             }
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             if (gmx_mm_any_lt(rsq11,rcutoff2))
 438             {
 439
 440             /* REACTION-FIELD ELECTROSTATICS */
 441             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 442             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 443
 444             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velec            = _mm_and_pd(velec,cutoff_mask);
 448             velecsum         = _mm_add_pd(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             fscal            = _mm_and_pd(fscal,cutoff_mask);
 453
 454             /* Calculate temporary vectorial force */
 455             tx               = _mm_mul_pd(fscal,dx11);
 456             ty               = _mm_mul_pd(fscal,dy11);
 457             tz               = _mm_mul_pd(fscal,dz11);
 458
 459             /* Update vectorial force */
 460             fix1             = _mm_add_pd(fix1,tx);
 461             fiy1             = _mm_add_pd(fiy1,ty);
 462             fiz1             = _mm_add_pd(fiz1,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
 470             /**************************
 471              * CALCULATE INTERACTIONS *
 472              **************************/
 473
 474             if (gmx_mm_any_lt(rsq12,rcutoff2))
 475             {
 476
 477             /* REACTION-FIELD ELECTROSTATICS */
 478             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 479             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 480
 481             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 482
 483             /* Update potential sum for this i atom from the interaction with this j atom. */
 484             velec            = _mm_and_pd(velec,cutoff_mask);
 485             velecsum         = _mm_add_pd(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             fscal            = _mm_and_pd(fscal,cutoff_mask);
 490
 491             /* Calculate temporary vectorial force */
 492             tx               = _mm_mul_pd(fscal,dx12);
 493             ty               = _mm_mul_pd(fscal,dy12);
 494             tz               = _mm_mul_pd(fscal,dz12);
 495
 496             /* Update vectorial force */
 497             fix1             = _mm_add_pd(fix1,tx);
 498             fiy1             = _mm_add_pd(fiy1,ty);
 499             fiz1             = _mm_add_pd(fiz1,tz);
 500
 501             fjx2             = _mm_add_pd(fjx2,tx);
 502             fjy2             = _mm_add_pd(fjy2,ty);
 503             fjz2             = _mm_add_pd(fjz2,tz);
 504
 505             }
 506
 507             /**************************
 508              * CALCULATE INTERACTIONS *
 509              **************************/
 510
 511             if (gmx_mm_any_lt(rsq20,rcutoff2))
 512             {
 513
 514             /* REACTION-FIELD ELECTROSTATICS */
 515             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 516             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 517
 518             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 519
 520             /* Update potential sum for this i atom from the interaction with this j atom. */
 521             velec            = _mm_and_pd(velec,cutoff_mask);
 522             velecsum         = _mm_add_pd(velecsum,velec);
 523
 524             fscal            = felec;
 525
 526             fscal            = _mm_and_pd(fscal,cutoff_mask);
 527
 528             /* Calculate temporary vectorial force */
 529             tx               = _mm_mul_pd(fscal,dx20);
 530             ty               = _mm_mul_pd(fscal,dy20);
 531             tz               = _mm_mul_pd(fscal,dz20);
 532
 533             /* Update vectorial force */
 534             fix2             = _mm_add_pd(fix2,tx);
 535             fiy2             = _mm_add_pd(fiy2,ty);
 536             fiz2             = _mm_add_pd(fiz2,tz);
 537
 538             fjx0             = _mm_add_pd(fjx0,tx);
 539             fjy0             = _mm_add_pd(fjy0,ty);
 540             fjz0             = _mm_add_pd(fjz0,tz);
 541
 542             }
 543
 544             /**************************
 545              * CALCULATE INTERACTIONS *
 546              **************************/
 547
 548             if (gmx_mm_any_lt(rsq21,rcutoff2))
 549             {
 550
 551             /* REACTION-FIELD ELECTROSTATICS */
 552             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 553             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 554
 555             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 556
 557             /* Update potential sum for this i atom from the interaction with this j atom. */
 558             velec            = _mm_and_pd(velec,cutoff_mask);
 559             velecsum         = _mm_add_pd(velecsum,velec);
 560
 561             fscal            = felec;
 562
 563             fscal            = _mm_and_pd(fscal,cutoff_mask);
 564
 565             /* Calculate temporary vectorial force */
 566             tx               = _mm_mul_pd(fscal,dx21);
 567             ty               = _mm_mul_pd(fscal,dy21);
 568             tz               = _mm_mul_pd(fscal,dz21);
 569
 570             /* Update vectorial force */
 571             fix2             = _mm_add_pd(fix2,tx);
 572             fiy2             = _mm_add_pd(fiy2,ty);
 573             fiz2             = _mm_add_pd(fiz2,tz);
 574
 575             fjx1             = _mm_add_pd(fjx1,tx);
 576             fjy1             = _mm_add_pd(fjy1,ty);
 577             fjz1             = _mm_add_pd(fjz1,tz);
 578
 579             }
 580
 581             /**************************
 582              * CALCULATE INTERACTIONS *
 583              **************************/
 584
 585             if (gmx_mm_any_lt(rsq22,rcutoff2))
 586             {
 587
 588             /* REACTION-FIELD ELECTROSTATICS */
 589             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 590             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 591
 592             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 593
 594             /* Update potential sum for this i atom from the interaction with this j atom. */
 595             velec            = _mm_and_pd(velec,cutoff_mask);
 596             velecsum         = _mm_add_pd(velecsum,velec);
 597
 598             fscal            = felec;
 599
 600             fscal            = _mm_and_pd(fscal,cutoff_mask);
 601
 602             /* Calculate temporary vectorial force */
 603             tx               = _mm_mul_pd(fscal,dx22);
 604             ty               = _mm_mul_pd(fscal,dy22);
 605             tz               = _mm_mul_pd(fscal,dz22);
 606
 607             /* Update vectorial force */
 608             fix2             = _mm_add_pd(fix2,tx);
 609             fiy2             = _mm_add_pd(fiy2,ty);
 610             fiz2             = _mm_add_pd(fiz2,tz);
 611
 612             fjx2             = _mm_add_pd(fjx2,tx);
 613             fjy2             = _mm_add_pd(fjy2,ty);
 614             fjz2             = _mm_add_pd(fjz2,tz);
 615
 616             }
 617
 618             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 619
 620             /* Inner loop uses 342 flops */
 621         }
 622
 623         if(jidx<j_index_end)
 624         {
 625
 626             jnrA             = jjnr[jidx];
 627             j_coord_offsetA  = DIM*jnrA;
 628
 629             /* load j atom coordinates */
 630             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 631                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 632
 633             /* Calculate displacement vector */
 634             dx00             = _mm_sub_pd(ix0,jx0);
 635             dy00             = _mm_sub_pd(iy0,jy0);
 636             dz00             = _mm_sub_pd(iz0,jz0);
 637             dx01             = _mm_sub_pd(ix0,jx1);
 638             dy01             = _mm_sub_pd(iy0,jy1);
 639             dz01             = _mm_sub_pd(iz0,jz1);
 640             dx02             = _mm_sub_pd(ix0,jx2);
 641             dy02             = _mm_sub_pd(iy0,jy2);
 642             dz02             = _mm_sub_pd(iz0,jz2);
 643             dx10             = _mm_sub_pd(ix1,jx0);
 644             dy10             = _mm_sub_pd(iy1,jy0);
 645             dz10             = _mm_sub_pd(iz1,jz0);
 646             dx11             = _mm_sub_pd(ix1,jx1);
 647             dy11             = _mm_sub_pd(iy1,jy1);
 648             dz11             = _mm_sub_pd(iz1,jz1);
 649             dx12             = _mm_sub_pd(ix1,jx2);
 650             dy12             = _mm_sub_pd(iy1,jy2);
 651             dz12             = _mm_sub_pd(iz1,jz2);
 652             dx20             = _mm_sub_pd(ix2,jx0);
 653             dy20             = _mm_sub_pd(iy2,jy0);
 654             dz20             = _mm_sub_pd(iz2,jz0);
 655             dx21             = _mm_sub_pd(ix2,jx1);
 656             dy21             = _mm_sub_pd(iy2,jy1);
 657             dz21             = _mm_sub_pd(iz2,jz1);
 658             dx22             = _mm_sub_pd(ix2,jx2);
 659             dy22             = _mm_sub_pd(iy2,jy2);
 660             dz22             = _mm_sub_pd(iz2,jz2);
 661
 662             /* Calculate squared distance and things based on it */
 663             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 664             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 665             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 666             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 667             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 668             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 669             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 670             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 671             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 672
 673             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 674             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 675             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 676             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 677             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 678             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 679             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 680             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 681             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 682
 683             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 684             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 685             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 686             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 687             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 688             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 689             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 690             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 691             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 692
 693             fjx0             = _mm_setzero_pd();
 694             fjy0             = _mm_setzero_pd();
 695             fjz0             = _mm_setzero_pd();
 696             fjx1             = _mm_setzero_pd();
 697             fjy1             = _mm_setzero_pd();
 698             fjz1             = _mm_setzero_pd();
 699             fjx2             = _mm_setzero_pd();
 700             fjy2             = _mm_setzero_pd();
 701             fjz2             = _mm_setzero_pd();
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             if (gmx_mm_any_lt(rsq00,rcutoff2))
 708             {
 709
 710             /* REACTION-FIELD ELECTROSTATICS */
 711             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 712             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 713
 714             /* LENNARD-JONES DISPERSION/REPULSION */
 715
 716             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 717             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 718             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 719             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 720                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 721             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 722
 723             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 724
 725             /* Update potential sum for this i atom from the interaction with this j atom. */
 726             velec            = _mm_and_pd(velec,cutoff_mask);
 727             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 728             velecsum         = _mm_add_pd(velecsum,velec);
 729             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 730             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 731             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 732
 733             fscal            = _mm_add_pd(felec,fvdw);
 734
 735             fscal            = _mm_and_pd(fscal,cutoff_mask);
 736
 737             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 738
 739             /* Calculate temporary vectorial force */
 740             tx               = _mm_mul_pd(fscal,dx00);
 741             ty               = _mm_mul_pd(fscal,dy00);
 742             tz               = _mm_mul_pd(fscal,dz00);
 743
 744             /* Update vectorial force */
 745             fix0             = _mm_add_pd(fix0,tx);
 746             fiy0             = _mm_add_pd(fiy0,ty);
 747             fiz0             = _mm_add_pd(fiz0,tz);
 748
 749             fjx0             = _mm_add_pd(fjx0,tx);
 750             fjy0             = _mm_add_pd(fjy0,ty);
 751             fjz0             = _mm_add_pd(fjz0,tz);
 752
 753             }
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             if (gmx_mm_any_lt(rsq01,rcutoff2))
 760             {
 761
 762             /* REACTION-FIELD ELECTROSTATICS */
 763             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 764             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 765
 766             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 767
 768             /* Update potential sum for this i atom from the interaction with this j atom. */
 769             velec            = _mm_and_pd(velec,cutoff_mask);
 770             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 771             velecsum         = _mm_add_pd(velecsum,velec);
 772
 773             fscal            = felec;
 774
 775             fscal            = _mm_and_pd(fscal,cutoff_mask);
 776
 777             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 778
 779             /* Calculate temporary vectorial force */
 780             tx               = _mm_mul_pd(fscal,dx01);
 781             ty               = _mm_mul_pd(fscal,dy01);
 782             tz               = _mm_mul_pd(fscal,dz01);
 783
 784             /* Update vectorial force */
 785             fix0             = _mm_add_pd(fix0,tx);
 786             fiy0             = _mm_add_pd(fiy0,ty);
 787             fiz0             = _mm_add_pd(fiz0,tz);
 788
 789             fjx1             = _mm_add_pd(fjx1,tx);
 790             fjy1             = _mm_add_pd(fjy1,ty);
 791             fjz1             = _mm_add_pd(fjz1,tz);
 792
 793             }
 794
 795             /**************************
 796              * CALCULATE INTERACTIONS *
 797              **************************/
 798
 799             if (gmx_mm_any_lt(rsq02,rcutoff2))
 800             {
 801
 802             /* REACTION-FIELD ELECTROSTATICS */
 803             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 804             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 805
 806             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 807
 808             /* Update potential sum for this i atom from the interaction with this j atom. */
 809             velec            = _mm_and_pd(velec,cutoff_mask);
 810             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 811             velecsum         = _mm_add_pd(velecsum,velec);
 812
 813             fscal            = felec;
 814
 815             fscal            = _mm_and_pd(fscal,cutoff_mask);
 816
 817             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 818
 819             /* Calculate temporary vectorial force */
 820             tx               = _mm_mul_pd(fscal,dx02);
 821             ty               = _mm_mul_pd(fscal,dy02);
 822             tz               = _mm_mul_pd(fscal,dz02);
 823
 824             /* Update vectorial force */
 825             fix0             = _mm_add_pd(fix0,tx);
 826             fiy0             = _mm_add_pd(fiy0,ty);
 827             fiz0             = _mm_add_pd(fiz0,tz);
 828
 829             fjx2             = _mm_add_pd(fjx2,tx);
 830             fjy2             = _mm_add_pd(fjy2,ty);
 831             fjz2             = _mm_add_pd(fjz2,tz);
 832
 833             }
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             if (gmx_mm_any_lt(rsq10,rcutoff2))
 840             {
 841
 842             /* REACTION-FIELD ELECTROSTATICS */
 843             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 844             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 845
 846             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             velec            = _mm_and_pd(velec,cutoff_mask);
 850             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 851             velecsum         = _mm_add_pd(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_and_pd(fscal,cutoff_mask);
 856
 857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 858
 859             /* Calculate temporary vectorial force */
 860             tx               = _mm_mul_pd(fscal,dx10);
 861             ty               = _mm_mul_pd(fscal,dy10);
 862             tz               = _mm_mul_pd(fscal,dz10);
 863
 864             /* Update vectorial force */
 865             fix1             = _mm_add_pd(fix1,tx);
 866             fiy1             = _mm_add_pd(fiy1,ty);
 867             fiz1             = _mm_add_pd(fiz1,tz);
 868
 869             fjx0             = _mm_add_pd(fjx0,tx);
 870             fjy0             = _mm_add_pd(fjy0,ty);
 871             fjz0             = _mm_add_pd(fjz0,tz);
 872
 873             }
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             if (gmx_mm_any_lt(rsq11,rcutoff2))
 880             {
 881
 882             /* REACTION-FIELD ELECTROSTATICS */
 883             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 884             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 885
 886             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _mm_and_pd(velec,cutoff_mask);
 890             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 891             velecsum         = _mm_add_pd(velecsum,velec);
 892
 893             fscal            = felec;
 894
 895             fscal            = _mm_and_pd(fscal,cutoff_mask);
 896
 897             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 898
 899             /* Calculate temporary vectorial force */
 900             tx               = _mm_mul_pd(fscal,dx11);
 901             ty               = _mm_mul_pd(fscal,dy11);
 902             tz               = _mm_mul_pd(fscal,dz11);
 903
 904             /* Update vectorial force */
 905             fix1             = _mm_add_pd(fix1,tx);
 906             fiy1             = _mm_add_pd(fiy1,ty);
 907             fiz1             = _mm_add_pd(fiz1,tz);
 908
 909             fjx1             = _mm_add_pd(fjx1,tx);
 910             fjy1             = _mm_add_pd(fjy1,ty);
 911             fjz1             = _mm_add_pd(fjz1,tz);
 912
 913             }
 914
 915             /**************************
 916              * CALCULATE INTERACTIONS *
 917              **************************/
 918
 919             if (gmx_mm_any_lt(rsq12,rcutoff2))
 920             {
 921
 922             /* REACTION-FIELD ELECTROSTATICS */
 923             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 924             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 925
 926             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 927
 928             /* Update potential sum for this i atom from the interaction with this j atom. */
 929             velec            = _mm_and_pd(velec,cutoff_mask);
 930             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 931             velecsum         = _mm_add_pd(velecsum,velec);
 932
 933             fscal            = felec;
 934
 935             fscal            = _mm_and_pd(fscal,cutoff_mask);
 936
 937             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 938
 939             /* Calculate temporary vectorial force */
 940             tx               = _mm_mul_pd(fscal,dx12);
 941             ty               = _mm_mul_pd(fscal,dy12);
 942             tz               = _mm_mul_pd(fscal,dz12);
 943
 944             /* Update vectorial force */
 945             fix1             = _mm_add_pd(fix1,tx);
 946             fiy1             = _mm_add_pd(fiy1,ty);
 947             fiz1             = _mm_add_pd(fiz1,tz);
 948
 949             fjx2             = _mm_add_pd(fjx2,tx);
 950             fjy2             = _mm_add_pd(fjy2,ty);
 951             fjz2             = _mm_add_pd(fjz2,tz);
 952
 953             }
 954
 955             /**************************
 956              * CALCULATE INTERACTIONS *
 957              **************************/
 958
 959             if (gmx_mm_any_lt(rsq20,rcutoff2))
 960             {
 961
 962             /* REACTION-FIELD ELECTROSTATICS */
 963             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 964             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 965
 966             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 967
 968             /* Update potential sum for this i atom from the interaction with this j atom. */
 969             velec            = _mm_and_pd(velec,cutoff_mask);
 970             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 971             velecsum         = _mm_add_pd(velecsum,velec);
 972
 973             fscal            = felec;
 974
 975             fscal            = _mm_and_pd(fscal,cutoff_mask);
 976
 977             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 978
 979             /* Calculate temporary vectorial force */
 980             tx               = _mm_mul_pd(fscal,dx20);
 981             ty               = _mm_mul_pd(fscal,dy20);
 982             tz               = _mm_mul_pd(fscal,dz20);
 983
 984             /* Update vectorial force */
 985             fix2             = _mm_add_pd(fix2,tx);
 986             fiy2             = _mm_add_pd(fiy2,ty);
 987             fiz2             = _mm_add_pd(fiz2,tz);
 988
 989             fjx0             = _mm_add_pd(fjx0,tx);
 990             fjy0             = _mm_add_pd(fjy0,ty);
 991             fjz0             = _mm_add_pd(fjz0,tz);
 992
 993             }
 994
 995             /**************************
 996              * CALCULATE INTERACTIONS *
 997              **************************/
 998
 999             if (gmx_mm_any_lt(rsq21,rcutoff2))
1000             {
1001
1002             /* REACTION-FIELD ELECTROSTATICS */
1003             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
1004             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1005
1006             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1007
1008             /* Update potential sum for this i atom from the interaction with this j atom. */
1009             velec            = _mm_and_pd(velec,cutoff_mask);
1010             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1011             velecsum         = _mm_add_pd(velecsum,velec);
1012
1013             fscal            = felec;
1014
1015             fscal            = _mm_and_pd(fscal,cutoff_mask);
1016
1017             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1018
1019             /* Calculate temporary vectorial force */
1020             tx               = _mm_mul_pd(fscal,dx21);
1021             ty               = _mm_mul_pd(fscal,dy21);
1022             tz               = _mm_mul_pd(fscal,dz21);
1023
1024             /* Update vectorial force */
1025             fix2             = _mm_add_pd(fix2,tx);
1026             fiy2             = _mm_add_pd(fiy2,ty);
1027             fiz2             = _mm_add_pd(fiz2,tz);
1028
1029             fjx1             = _mm_add_pd(fjx1,tx);
1030             fjy1             = _mm_add_pd(fjy1,ty);
1031             fjz1             = _mm_add_pd(fjz1,tz);
1032
1033             }
1034
1035             /**************************
1036              * CALCULATE INTERACTIONS *
1037              **************************/
1038
1039             if (gmx_mm_any_lt(rsq22,rcutoff2))
1040             {
1041
1042             /* REACTION-FIELD ELECTROSTATICS */
1043             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1044             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1045
1046             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1047
1048             /* Update potential sum for this i atom from the interaction with this j atom. */
1049             velec            = _mm_and_pd(velec,cutoff_mask);
1050             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1051             velecsum         = _mm_add_pd(velecsum,velec);
1052
1053             fscal            = felec;
1054
1055             fscal            = _mm_and_pd(fscal,cutoff_mask);
1056
1057             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1058
1059             /* Calculate temporary vectorial force */
1060             tx               = _mm_mul_pd(fscal,dx22);
1061             ty               = _mm_mul_pd(fscal,dy22);
1062             tz               = _mm_mul_pd(fscal,dz22);
1063
1064             /* Update vectorial force */
1065             fix2             = _mm_add_pd(fix2,tx);
1066             fiy2             = _mm_add_pd(fiy2,ty);
1067             fiz2             = _mm_add_pd(fiz2,tz);
1068
1069             fjx2             = _mm_add_pd(fjx2,tx);
1070             fjy2             = _mm_add_pd(fjy2,ty);
1071             fjz2             = _mm_add_pd(fjz2,tz);
1072
1073             }
1074
1075             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1076
1077             /* Inner loop uses 342 flops */
1078         }
1079
1080         /* End of innermost loop */
1081
1082         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1083                                               f+i_coord_offset,fshift+i_shift_offset);
1084
1085         ggid                        = gid[iidx];
1086         /* Update potential energies */
1087         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1088         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1089
1090         /* Increment number of inner iterations */
1091         inneriter                  += j_index_end - j_index_start;
1092
1093         /* Outer loop uses 20 flops */
1094     }
1095
1096     /* Increment number of outer iterations */
1097     outeriter        += nri;
1098
1099     /* Update outer/inner flops */
1100
1101     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*342);
1102 }
1103 /*
1104  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sse4_1_double
1105  * Electrostatics interaction: ReactionField
1106  * VdW interaction:            LennardJones
1107  * Geometry:                   Water3-Water3
1108  * Calculate force/pot:        Force
1109  */
1110 void
1111 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sse4_1_double
1112                     (t_nblist * gmx_restrict                nlist,
1113                      rvec * gmx_restrict                    xx,
1114                      rvec * gmx_restrict                    ff,
1115                      t_forcerec * gmx_restrict              fr,
1116                      t_mdatoms * gmx_restrict               mdatoms,
1117                      nb_kernel_data_t * gmx_restrict        kernel_data,
1118                      t_nrnb * gmx_restrict                  nrnb)
1119 {
1120     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1121      * just 0 for non-waters.
1122      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1123      * jnr indices corresponding to data put in the four positions in the SIMD register.
1124      */
1125     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1126     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1127     int              jnrA,jnrB;
1128     int              j_coord_offsetA,j_coord_offsetB;
1129     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1130     real             rcutoff_scalar;
1131     real             *shiftvec,*fshift,*x,*f;
1132     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1133     int              vdwioffset0;
1134     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1135     int              vdwioffset1;
1136     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1137     int              vdwioffset2;
1138     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1139     int              vdwjidx0A,vdwjidx0B;
1140     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1141     int              vdwjidx1A,vdwjidx1B;
1142     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1143     int              vdwjidx2A,vdwjidx2B;
1144     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1145     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1146     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1147     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1148     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1149     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1150     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1151     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1152     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1153     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1154     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1155     real             *charge;
1156     int              nvdwtype;
1157     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1158     int              *vdwtype;
1159     real             *vdwparam;
1160     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1161     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1162     __m128d          dummy_mask,cutoff_mask;
1163     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1164     __m128d          one     = _mm_set1_pd(1.0);
1165     __m128d          two     = _mm_set1_pd(2.0);
1166     x                = xx[0];
1167     f                = ff[0];
1168
1169     nri              = nlist->nri;
1170     iinr             = nlist->iinr;
1171     jindex           = nlist->jindex;
1172     jjnr             = nlist->jjnr;
1173     shiftidx         = nlist->shift;
1174     gid              = nlist->gid;
1175     shiftvec         = fr->shift_vec[0];
1176     fshift           = fr->fshift[0];
1177     facel            = _mm_set1_pd(fr->epsfac);
1178     charge           = mdatoms->chargeA;
1179     krf              = _mm_set1_pd(fr->ic->k_rf);
1180     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1181     crf              = _mm_set1_pd(fr->ic->c_rf);
1182     nvdwtype         = fr->ntype;
1183     vdwparam         = fr->nbfp;
1184     vdwtype          = mdatoms->typeA;
1185
1186     /* Setup water-specific parameters */
1187     inr              = nlist->iinr[0];
1188     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1189     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1190     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1191     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1192
1193     jq0              = _mm_set1_pd(charge[inr+0]);
1194     jq1              = _mm_set1_pd(charge[inr+1]);
1195     jq2              = _mm_set1_pd(charge[inr+2]);
1196     vdwjidx0A        = 2*vdwtype[inr+0];
1197     qq00             = _mm_mul_pd(iq0,jq0);
1198     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1199     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1200     qq01             = _mm_mul_pd(iq0,jq1);
1201     qq02             = _mm_mul_pd(iq0,jq2);
1202     qq10             = _mm_mul_pd(iq1,jq0);
1203     qq11             = _mm_mul_pd(iq1,jq1);
1204     qq12             = _mm_mul_pd(iq1,jq2);
1205     qq20             = _mm_mul_pd(iq2,jq0);
1206     qq21             = _mm_mul_pd(iq2,jq1);
1207     qq22             = _mm_mul_pd(iq2,jq2);
1208
1209     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1210     rcutoff_scalar   = fr->rcoulomb;
1211     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1212     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1213
1214     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1215     rvdw             = _mm_set1_pd(fr->rvdw);
1216
1217     /* Avoid stupid compiler warnings */
1218     jnrA = jnrB = 0;
1219     j_coord_offsetA = 0;
1220     j_coord_offsetB = 0;
1221
1222     outeriter        = 0;
1223     inneriter        = 0;
1224
1225     /* Start outer loop over neighborlists */
1226     for(iidx=0; iidx<nri; iidx++)
1227     {
1228         /* Load shift vector for this list */
1229         i_shift_offset   = DIM*shiftidx[iidx];
1230
1231         /* Load limits for loop over neighbors */
1232         j_index_start    = jindex[iidx];
1233         j_index_end      = jindex[iidx+1];
1234
1235         /* Get outer coordinate index */
1236         inr              = iinr[iidx];
1237         i_coord_offset   = DIM*inr;
1238
1239         /* Load i particle coords and add shift vector */
1240         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1241                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1242
1243         fix0             = _mm_setzero_pd();
1244         fiy0             = _mm_setzero_pd();
1245         fiz0             = _mm_setzero_pd();
1246         fix1             = _mm_setzero_pd();
1247         fiy1             = _mm_setzero_pd();
1248         fiz1             = _mm_setzero_pd();
1249         fix2             = _mm_setzero_pd();
1250         fiy2             = _mm_setzero_pd();
1251         fiz2             = _mm_setzero_pd();
1252
1253         /* Start inner kernel loop */
1254         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1255         {
1256
1257             /* Get j neighbor index, and coordinate index */
1258             jnrA             = jjnr[jidx];
1259             jnrB             = jjnr[jidx+1];
1260             j_coord_offsetA  = DIM*jnrA;
1261             j_coord_offsetB  = DIM*jnrB;
1262
1263             /* load j atom coordinates */
1264             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1265                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1266
1267             /* Calculate displacement vector */
1268             dx00             = _mm_sub_pd(ix0,jx0);
1269             dy00             = _mm_sub_pd(iy0,jy0);
1270             dz00             = _mm_sub_pd(iz0,jz0);
1271             dx01             = _mm_sub_pd(ix0,jx1);
1272             dy01             = _mm_sub_pd(iy0,jy1);
1273             dz01             = _mm_sub_pd(iz0,jz1);
1274             dx02             = _mm_sub_pd(ix0,jx2);
1275             dy02             = _mm_sub_pd(iy0,jy2);
1276             dz02             = _mm_sub_pd(iz0,jz2);
1277             dx10             = _mm_sub_pd(ix1,jx0);
1278             dy10             = _mm_sub_pd(iy1,jy0);
1279             dz10             = _mm_sub_pd(iz1,jz0);
1280             dx11             = _mm_sub_pd(ix1,jx1);
1281             dy11             = _mm_sub_pd(iy1,jy1);
1282             dz11             = _mm_sub_pd(iz1,jz1);
1283             dx12             = _mm_sub_pd(ix1,jx2);
1284             dy12             = _mm_sub_pd(iy1,jy2);
1285             dz12             = _mm_sub_pd(iz1,jz2);
1286             dx20             = _mm_sub_pd(ix2,jx0);
1287             dy20             = _mm_sub_pd(iy2,jy0);
1288             dz20             = _mm_sub_pd(iz2,jz0);
1289             dx21             = _mm_sub_pd(ix2,jx1);
1290             dy21             = _mm_sub_pd(iy2,jy1);
1291             dz21             = _mm_sub_pd(iz2,jz1);
1292             dx22             = _mm_sub_pd(ix2,jx2);
1293             dy22             = _mm_sub_pd(iy2,jy2);
1294             dz22             = _mm_sub_pd(iz2,jz2);
1295
1296             /* Calculate squared distance and things based on it */
1297             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1298             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1299             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1300             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1301             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1302             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1303             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1304             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1305             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1306
1307             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1308             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1309             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1310             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1311             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1312             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1313             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1314             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1315             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1316
1317             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1318             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1319             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1320             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1321             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1322             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1323             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1324             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1325             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1326
1327             fjx0             = _mm_setzero_pd();
1328             fjy0             = _mm_setzero_pd();
1329             fjz0             = _mm_setzero_pd();
1330             fjx1             = _mm_setzero_pd();
1331             fjy1             = _mm_setzero_pd();
1332             fjz1             = _mm_setzero_pd();
1333             fjx2             = _mm_setzero_pd();
1334             fjy2             = _mm_setzero_pd();
1335             fjz2             = _mm_setzero_pd();
1336
1337             /**************************
1338              * CALCULATE INTERACTIONS *
1339              **************************/
1340
1341             if (gmx_mm_any_lt(rsq00,rcutoff2))
1342             {
1343
1344             /* REACTION-FIELD ELECTROSTATICS */
1345             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1346
1347             /* LENNARD-JONES DISPERSION/REPULSION */
1348
1349             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1350             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1351
1352             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1353
1354             fscal            = _mm_add_pd(felec,fvdw);
1355
1356             fscal            = _mm_and_pd(fscal,cutoff_mask);
1357
1358             /* Calculate temporary vectorial force */
1359             tx               = _mm_mul_pd(fscal,dx00);
1360             ty               = _mm_mul_pd(fscal,dy00);
1361             tz               = _mm_mul_pd(fscal,dz00);
1362
1363             /* Update vectorial force */
1364             fix0             = _mm_add_pd(fix0,tx);
1365             fiy0             = _mm_add_pd(fiy0,ty);
1366             fiz0             = _mm_add_pd(fiz0,tz);
1367
1368             fjx0             = _mm_add_pd(fjx0,tx);
1369             fjy0             = _mm_add_pd(fjy0,ty);
1370             fjz0             = _mm_add_pd(fjz0,tz);
1371
1372             }
1373
1374             /**************************
1375              * CALCULATE INTERACTIONS *
1376              **************************/
1377
1378             if (gmx_mm_any_lt(rsq01,rcutoff2))
1379             {
1380
1381             /* REACTION-FIELD ELECTROSTATICS */
1382             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1383
1384             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1385
1386             fscal            = felec;
1387
1388             fscal            = _mm_and_pd(fscal,cutoff_mask);
1389
1390             /* Calculate temporary vectorial force */
1391             tx               = _mm_mul_pd(fscal,dx01);
1392             ty               = _mm_mul_pd(fscal,dy01);
1393             tz               = _mm_mul_pd(fscal,dz01);
1394
1395             /* Update vectorial force */
1396             fix0             = _mm_add_pd(fix0,tx);
1397             fiy0             = _mm_add_pd(fiy0,ty);
1398             fiz0             = _mm_add_pd(fiz0,tz);
1399
1400             fjx1             = _mm_add_pd(fjx1,tx);
1401             fjy1             = _mm_add_pd(fjy1,ty);
1402             fjz1             = _mm_add_pd(fjz1,tz);
1403
1404             }
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             if (gmx_mm_any_lt(rsq02,rcutoff2))
1411             {
1412
1413             /* REACTION-FIELD ELECTROSTATICS */
1414             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1415
1416             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1417
1418             fscal            = felec;
1419
1420             fscal            = _mm_and_pd(fscal,cutoff_mask);
1421
1422             /* Calculate temporary vectorial force */
1423             tx               = _mm_mul_pd(fscal,dx02);
1424             ty               = _mm_mul_pd(fscal,dy02);
1425             tz               = _mm_mul_pd(fscal,dz02);
1426
1427             /* Update vectorial force */
1428             fix0             = _mm_add_pd(fix0,tx);
1429             fiy0             = _mm_add_pd(fiy0,ty);
1430             fiz0             = _mm_add_pd(fiz0,tz);
1431
1432             fjx2             = _mm_add_pd(fjx2,tx);
1433             fjy2             = _mm_add_pd(fjy2,ty);
1434             fjz2             = _mm_add_pd(fjz2,tz);
1435
1436             }
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             if (gmx_mm_any_lt(rsq10,rcutoff2))
1443             {
1444
1445             /* REACTION-FIELD ELECTROSTATICS */
1446             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1447
1448             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1449
1450             fscal            = felec;
1451
1452             fscal            = _mm_and_pd(fscal,cutoff_mask);
1453
1454             /* Calculate temporary vectorial force */
1455             tx               = _mm_mul_pd(fscal,dx10);
1456             ty               = _mm_mul_pd(fscal,dy10);
1457             tz               = _mm_mul_pd(fscal,dz10);
1458
1459             /* Update vectorial force */
1460             fix1             = _mm_add_pd(fix1,tx);
1461             fiy1             = _mm_add_pd(fiy1,ty);
1462             fiz1             = _mm_add_pd(fiz1,tz);
1463
1464             fjx0             = _mm_add_pd(fjx0,tx);
1465             fjy0             = _mm_add_pd(fjy0,ty);
1466             fjz0             = _mm_add_pd(fjz0,tz);
1467
1468             }
1469
1470             /**************************
1471              * CALCULATE INTERACTIONS *
1472              **************************/
1473
1474             if (gmx_mm_any_lt(rsq11,rcutoff2))
1475             {
1476
1477             /* REACTION-FIELD ELECTROSTATICS */
1478             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1479
1480             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1481
1482             fscal            = felec;
1483
1484             fscal            = _mm_and_pd(fscal,cutoff_mask);
1485
1486             /* Calculate temporary vectorial force */
1487             tx               = _mm_mul_pd(fscal,dx11);
1488             ty               = _mm_mul_pd(fscal,dy11);
1489             tz               = _mm_mul_pd(fscal,dz11);
1490
1491             /* Update vectorial force */
1492             fix1             = _mm_add_pd(fix1,tx);
1493             fiy1             = _mm_add_pd(fiy1,ty);
1494             fiz1             = _mm_add_pd(fiz1,tz);
1495
1496             fjx1             = _mm_add_pd(fjx1,tx);
1497             fjy1             = _mm_add_pd(fjy1,ty);
1498             fjz1             = _mm_add_pd(fjz1,tz);
1499
1500             }
1501
1502             /**************************
1503              * CALCULATE INTERACTIONS *
1504              **************************/
1505
1506             if (gmx_mm_any_lt(rsq12,rcutoff2))
1507             {
1508
1509             /* REACTION-FIELD ELECTROSTATICS */
1510             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1511
1512             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1513
1514             fscal            = felec;
1515
1516             fscal            = _mm_and_pd(fscal,cutoff_mask);
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
1534             /**************************
1535              * CALCULATE INTERACTIONS *
1536              **************************/
1537
1538             if (gmx_mm_any_lt(rsq20,rcutoff2))
1539             {
1540
1541             /* REACTION-FIELD ELECTROSTATICS */
1542             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1543
1544             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_and_pd(fscal,cutoff_mask);
1549
1550             /* Calculate temporary vectorial force */
1551             tx               = _mm_mul_pd(fscal,dx20);
1552             ty               = _mm_mul_pd(fscal,dy20);
1553             tz               = _mm_mul_pd(fscal,dz20);
1554
1555             /* Update vectorial force */
1556             fix2             = _mm_add_pd(fix2,tx);
1557             fiy2             = _mm_add_pd(fiy2,ty);
1558             fiz2             = _mm_add_pd(fiz2,tz);
1559
1560             fjx0             = _mm_add_pd(fjx0,tx);
1561             fjy0             = _mm_add_pd(fjy0,ty);
1562             fjz0             = _mm_add_pd(fjz0,tz);
1563
1564             }
1565
1566             /**************************
1567              * CALCULATE INTERACTIONS *
1568              **************************/
1569
1570             if (gmx_mm_any_lt(rsq21,rcutoff2))
1571             {
1572
1573             /* REACTION-FIELD ELECTROSTATICS */
1574             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1575
1576             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1577
1578             fscal            = felec;
1579
1580             fscal            = _mm_and_pd(fscal,cutoff_mask);
1581
1582             /* Calculate temporary vectorial force */
1583             tx               = _mm_mul_pd(fscal,dx21);
1584             ty               = _mm_mul_pd(fscal,dy21);
1585             tz               = _mm_mul_pd(fscal,dz21);
1586
1587             /* Update vectorial force */
1588             fix2             = _mm_add_pd(fix2,tx);
1589             fiy2             = _mm_add_pd(fiy2,ty);
1590             fiz2             = _mm_add_pd(fiz2,tz);
1591
1592             fjx1             = _mm_add_pd(fjx1,tx);
1593             fjy1             = _mm_add_pd(fjy1,ty);
1594             fjz1             = _mm_add_pd(fjz1,tz);
1595
1596             }
1597
1598             /**************************
1599              * CALCULATE INTERACTIONS *
1600              **************************/
1601
1602             if (gmx_mm_any_lt(rsq22,rcutoff2))
1603             {
1604
1605             /* REACTION-FIELD ELECTROSTATICS */
1606             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1607
1608             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1609
1610             fscal            = felec;
1611
1612             fscal            = _mm_and_pd(fscal,cutoff_mask);
1613
1614             /* Calculate temporary vectorial force */
1615             tx               = _mm_mul_pd(fscal,dx22);
1616             ty               = _mm_mul_pd(fscal,dy22);
1617             tz               = _mm_mul_pd(fscal,dz22);
1618
1619             /* Update vectorial force */
1620             fix2             = _mm_add_pd(fix2,tx);
1621             fiy2             = _mm_add_pd(fiy2,ty);
1622             fiz2             = _mm_add_pd(fiz2,tz);
1623
1624             fjx2             = _mm_add_pd(fjx2,tx);
1625             fjy2             = _mm_add_pd(fjy2,ty);
1626             fjz2             = _mm_add_pd(fjz2,tz);
1627
1628             }
1629
1630             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1631
1632             /* Inner loop uses 277 flops */
1633         }
1634
1635         if(jidx<j_index_end)
1636         {
1637
1638             jnrA             = jjnr[jidx];
1639             j_coord_offsetA  = DIM*jnrA;
1640
1641             /* load j atom coordinates */
1642             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1643                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1644
1645             /* Calculate displacement vector */
1646             dx00             = _mm_sub_pd(ix0,jx0);
1647             dy00             = _mm_sub_pd(iy0,jy0);
1648             dz00             = _mm_sub_pd(iz0,jz0);
1649             dx01             = _mm_sub_pd(ix0,jx1);
1650             dy01             = _mm_sub_pd(iy0,jy1);
1651             dz01             = _mm_sub_pd(iz0,jz1);
1652             dx02             = _mm_sub_pd(ix0,jx2);
1653             dy02             = _mm_sub_pd(iy0,jy2);
1654             dz02             = _mm_sub_pd(iz0,jz2);
1655             dx10             = _mm_sub_pd(ix1,jx0);
1656             dy10             = _mm_sub_pd(iy1,jy0);
1657             dz10             = _mm_sub_pd(iz1,jz0);
1658             dx11             = _mm_sub_pd(ix1,jx1);
1659             dy11             = _mm_sub_pd(iy1,jy1);
1660             dz11             = _mm_sub_pd(iz1,jz1);
1661             dx12             = _mm_sub_pd(ix1,jx2);
1662             dy12             = _mm_sub_pd(iy1,jy2);
1663             dz12             = _mm_sub_pd(iz1,jz2);
1664             dx20             = _mm_sub_pd(ix2,jx0);
1665             dy20             = _mm_sub_pd(iy2,jy0);
1666             dz20             = _mm_sub_pd(iz2,jz0);
1667             dx21             = _mm_sub_pd(ix2,jx1);
1668             dy21             = _mm_sub_pd(iy2,jy1);
1669             dz21             = _mm_sub_pd(iz2,jz1);
1670             dx22             = _mm_sub_pd(ix2,jx2);
1671             dy22             = _mm_sub_pd(iy2,jy2);
1672             dz22             = _mm_sub_pd(iz2,jz2);
1673
1674             /* Calculate squared distance and things based on it */
1675             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1676             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1677             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1678             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1679             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1680             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1681             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1682             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1683             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1684
1685             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1686             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1687             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1688             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1689             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1690             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1691             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1692             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1693             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1694
1695             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1696             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1697             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1698             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1699             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1700             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1701             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1702             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1703             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1704
1705             fjx0             = _mm_setzero_pd();
1706             fjy0             = _mm_setzero_pd();
1707             fjz0             = _mm_setzero_pd();
1708             fjx1             = _mm_setzero_pd();
1709             fjy1             = _mm_setzero_pd();
1710             fjz1             = _mm_setzero_pd();
1711             fjx2             = _mm_setzero_pd();
1712             fjy2             = _mm_setzero_pd();
1713             fjz2             = _mm_setzero_pd();
1714
1715             /**************************
1716              * CALCULATE INTERACTIONS *
1717              **************************/
1718
1719             if (gmx_mm_any_lt(rsq00,rcutoff2))
1720             {
1721
1722             /* REACTION-FIELD ELECTROSTATICS */
1723             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1724
1725             /* LENNARD-JONES DISPERSION/REPULSION */
1726
1727             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1728             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1729
1730             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1731
1732             fscal            = _mm_add_pd(felec,fvdw);
1733
1734             fscal            = _mm_and_pd(fscal,cutoff_mask);
1735
1736             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1737
1738             /* Calculate temporary vectorial force */
1739             tx               = _mm_mul_pd(fscal,dx00);
1740             ty               = _mm_mul_pd(fscal,dy00);
1741             tz               = _mm_mul_pd(fscal,dz00);
1742
1743             /* Update vectorial force */
1744             fix0             = _mm_add_pd(fix0,tx);
1745             fiy0             = _mm_add_pd(fiy0,ty);
1746             fiz0             = _mm_add_pd(fiz0,tz);
1747
1748             fjx0             = _mm_add_pd(fjx0,tx);
1749             fjy0             = _mm_add_pd(fjy0,ty);
1750             fjz0             = _mm_add_pd(fjz0,tz);
1751
1752             }
1753
1754             /**************************
1755              * CALCULATE INTERACTIONS *
1756              **************************/
1757
1758             if (gmx_mm_any_lt(rsq01,rcutoff2))
1759             {
1760
1761             /* REACTION-FIELD ELECTROSTATICS */
1762             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1763
1764             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1765
1766             fscal            = felec;
1767
1768             fscal            = _mm_and_pd(fscal,cutoff_mask);
1769
1770             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1771
1772             /* Calculate temporary vectorial force */
1773             tx               = _mm_mul_pd(fscal,dx01);
1774             ty               = _mm_mul_pd(fscal,dy01);
1775             tz               = _mm_mul_pd(fscal,dz01);
1776
1777             /* Update vectorial force */
1778             fix0             = _mm_add_pd(fix0,tx);
1779             fiy0             = _mm_add_pd(fiy0,ty);
1780             fiz0             = _mm_add_pd(fiz0,tz);
1781
1782             fjx1             = _mm_add_pd(fjx1,tx);
1783             fjy1             = _mm_add_pd(fjy1,ty);
1784             fjz1             = _mm_add_pd(fjz1,tz);
1785
1786             }
1787
1788             /**************************
1789              * CALCULATE INTERACTIONS *
1790              **************************/
1791
1792             if (gmx_mm_any_lt(rsq02,rcutoff2))
1793             {
1794
1795             /* REACTION-FIELD ELECTROSTATICS */
1796             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1797
1798             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1799
1800             fscal            = felec;
1801
1802             fscal            = _mm_and_pd(fscal,cutoff_mask);
1803
1804             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1805
1806             /* Calculate temporary vectorial force */
1807             tx               = _mm_mul_pd(fscal,dx02);
1808             ty               = _mm_mul_pd(fscal,dy02);
1809             tz               = _mm_mul_pd(fscal,dz02);
1810
1811             /* Update vectorial force */
1812             fix0             = _mm_add_pd(fix0,tx);
1813             fiy0             = _mm_add_pd(fiy0,ty);
1814             fiz0             = _mm_add_pd(fiz0,tz);
1815
1816             fjx2             = _mm_add_pd(fjx2,tx);
1817             fjy2             = _mm_add_pd(fjy2,ty);
1818             fjz2             = _mm_add_pd(fjz2,tz);
1819
1820             }
1821
1822             /**************************
1823              * CALCULATE INTERACTIONS *
1824              **************************/
1825
1826             if (gmx_mm_any_lt(rsq10,rcutoff2))
1827             {
1828
1829             /* REACTION-FIELD ELECTROSTATICS */
1830             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1831
1832             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1833
1834             fscal            = felec;
1835
1836             fscal            = _mm_and_pd(fscal,cutoff_mask);
1837
1838             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1839
1840             /* Calculate temporary vectorial force */
1841             tx               = _mm_mul_pd(fscal,dx10);
1842             ty               = _mm_mul_pd(fscal,dy10);
1843             tz               = _mm_mul_pd(fscal,dz10);
1844
1845             /* Update vectorial force */
1846             fix1             = _mm_add_pd(fix1,tx);
1847             fiy1             = _mm_add_pd(fiy1,ty);
1848             fiz1             = _mm_add_pd(fiz1,tz);
1849
1850             fjx0             = _mm_add_pd(fjx0,tx);
1851             fjy0             = _mm_add_pd(fjy0,ty);
1852             fjz0             = _mm_add_pd(fjz0,tz);
1853
1854             }
1855
1856             /**************************
1857              * CALCULATE INTERACTIONS *
1858              **************************/
1859
1860             if (gmx_mm_any_lt(rsq11,rcutoff2))
1861             {
1862
1863             /* REACTION-FIELD ELECTROSTATICS */
1864             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1865
1866             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1867
1868             fscal            = felec;
1869
1870             fscal            = _mm_and_pd(fscal,cutoff_mask);
1871
1872             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1873
1874             /* Calculate temporary vectorial force */
1875             tx               = _mm_mul_pd(fscal,dx11);
1876             ty               = _mm_mul_pd(fscal,dy11);
1877             tz               = _mm_mul_pd(fscal,dz11);
1878
1879             /* Update vectorial force */
1880             fix1             = _mm_add_pd(fix1,tx);
1881             fiy1             = _mm_add_pd(fiy1,ty);
1882             fiz1             = _mm_add_pd(fiz1,tz);
1883
1884             fjx1             = _mm_add_pd(fjx1,tx);
1885             fjy1             = _mm_add_pd(fjy1,ty);
1886             fjz1             = _mm_add_pd(fjz1,tz);
1887
1888             }
1889
1890             /**************************
1891              * CALCULATE INTERACTIONS *
1892              **************************/
1893
1894             if (gmx_mm_any_lt(rsq12,rcutoff2))
1895             {
1896
1897             /* REACTION-FIELD ELECTROSTATICS */
1898             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1899
1900             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1901
1902             fscal            = felec;
1903
1904             fscal            = _mm_and_pd(fscal,cutoff_mask);
1905
1906             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1907
1908             /* Calculate temporary vectorial force */
1909             tx               = _mm_mul_pd(fscal,dx12);
1910             ty               = _mm_mul_pd(fscal,dy12);
1911             tz               = _mm_mul_pd(fscal,dz12);
1912
1913             /* Update vectorial force */
1914             fix1             = _mm_add_pd(fix1,tx);
1915             fiy1             = _mm_add_pd(fiy1,ty);
1916             fiz1             = _mm_add_pd(fiz1,tz);
1917
1918             fjx2             = _mm_add_pd(fjx2,tx);
1919             fjy2             = _mm_add_pd(fjy2,ty);
1920             fjz2             = _mm_add_pd(fjz2,tz);
1921
1922             }
1923
1924             /**************************
1925              * CALCULATE INTERACTIONS *
1926              **************************/
1927
1928             if (gmx_mm_any_lt(rsq20,rcutoff2))
1929             {
1930
1931             /* REACTION-FIELD ELECTROSTATICS */
1932             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1933
1934             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1935
1936             fscal            = felec;
1937
1938             fscal            = _mm_and_pd(fscal,cutoff_mask);
1939
1940             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1941
1942             /* Calculate temporary vectorial force */
1943             tx               = _mm_mul_pd(fscal,dx20);
1944             ty               = _mm_mul_pd(fscal,dy20);
1945             tz               = _mm_mul_pd(fscal,dz20);
1946
1947             /* Update vectorial force */
1948             fix2             = _mm_add_pd(fix2,tx);
1949             fiy2             = _mm_add_pd(fiy2,ty);
1950             fiz2             = _mm_add_pd(fiz2,tz);
1951
1952             fjx0             = _mm_add_pd(fjx0,tx);
1953             fjy0             = _mm_add_pd(fjy0,ty);
1954             fjz0             = _mm_add_pd(fjz0,tz);
1955
1956             }
1957
1958             /**************************
1959              * CALCULATE INTERACTIONS *
1960              **************************/
1961
1962             if (gmx_mm_any_lt(rsq21,rcutoff2))
1963             {
1964
1965             /* REACTION-FIELD ELECTROSTATICS */
1966             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1967
1968             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1969
1970             fscal            = felec;
1971
1972             fscal            = _mm_and_pd(fscal,cutoff_mask);
1973
1974             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1975
1976             /* Calculate temporary vectorial force */
1977             tx               = _mm_mul_pd(fscal,dx21);
1978             ty               = _mm_mul_pd(fscal,dy21);
1979             tz               = _mm_mul_pd(fscal,dz21);
1980
1981             /* Update vectorial force */
1982             fix2             = _mm_add_pd(fix2,tx);
1983             fiy2             = _mm_add_pd(fiy2,ty);
1984             fiz2             = _mm_add_pd(fiz2,tz);
1985
1986             fjx1             = _mm_add_pd(fjx1,tx);
1987             fjy1             = _mm_add_pd(fjy1,ty);
1988             fjz1             = _mm_add_pd(fjz1,tz);
1989
1990             }
1991
1992             /**************************
1993              * CALCULATE INTERACTIONS *
1994              **************************/
1995
1996             if (gmx_mm_any_lt(rsq22,rcutoff2))
1997             {
1998
1999             /* REACTION-FIELD ELECTROSTATICS */
2000             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2001
2002             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2003
2004             fscal            = felec;
2005
2006             fscal            = _mm_and_pd(fscal,cutoff_mask);
2007
2008             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2009
2010             /* Calculate temporary vectorial force */
2011             tx               = _mm_mul_pd(fscal,dx22);
2012             ty               = _mm_mul_pd(fscal,dy22);
2013             tz               = _mm_mul_pd(fscal,dz22);
2014
2015             /* Update vectorial force */
2016             fix2             = _mm_add_pd(fix2,tx);
2017             fiy2             = _mm_add_pd(fiy2,ty);
2018             fiz2             = _mm_add_pd(fiz2,tz);
2019
2020             fjx2             = _mm_add_pd(fjx2,tx);
2021             fjy2             = _mm_add_pd(fjy2,ty);
2022             fjz2             = _mm_add_pd(fjz2,tz);
2023
2024             }
2025
2026             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2027
2028             /* Inner loop uses 277 flops */
2029         }
2030
2031         /* End of innermost loop */
2032
2033         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2034                                               f+i_coord_offset,fshift+i_shift_offset);
2035
2036         /* Increment number of inner iterations */
2037         inneriter                  += j_index_end - j_index_start;
2038
2039         /* Outer loop uses 18 flops */
2040     }
2041
2042     /* Increment number of outer iterations */
2043     outeriter        += nri;
2044
2045     /* Update outer/inner flops */
2046
2047     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
2048 }