src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwLJSw_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_VdwLJSw_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_VdwLJSw_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          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  96     real             rswitch_scalar,d_scalar;
  97     __m128d          dummy_mask,cutoff_mask;
  98     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  99     __m128d          one     = _mm_set1_pd(1.0);
 100     __m128d          two     = _mm_set1_pd(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm_set1_pd(fr->epsfac);
 113     charge           = mdatoms->chargeA;
 114     krf              = _mm_set1_pd(fr->ic->k_rf);
 115     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 116     crf              = _mm_set1_pd(fr->ic->c_rf);
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     /* Setup water-specific parameters */
 122     inr              = nlist->iinr[0];
 123     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 124     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 125     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 126     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 127
 128     jq0              = _mm_set1_pd(charge[inr+0]);
 129     jq1              = _mm_set1_pd(charge[inr+1]);
 130     jq2              = _mm_set1_pd(charge[inr+2]);
 131     vdwjidx0A        = 2*vdwtype[inr+0];
 132     qq00             = _mm_mul_pd(iq0,jq0);
 133     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 134     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 135     qq01             = _mm_mul_pd(iq0,jq1);
 136     qq02             = _mm_mul_pd(iq0,jq2);
 137     qq10             = _mm_mul_pd(iq1,jq0);
 138     qq11             = _mm_mul_pd(iq1,jq1);
 139     qq12             = _mm_mul_pd(iq1,jq2);
 140     qq20             = _mm_mul_pd(iq2,jq0);
 141     qq21             = _mm_mul_pd(iq2,jq1);
 142     qq22             = _mm_mul_pd(iq2,jq2);
 143
 144     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 145     rcutoff_scalar   = fr->rcoulomb;
 146     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 147     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 148
 149     rswitch_scalar   = fr->rvdw_switch;
 150     rswitch          = _mm_set1_pd(rswitch_scalar);
 151     /* Setup switch parameters */
 152     d_scalar         = rcutoff_scalar-rswitch_scalar;
 153     d                = _mm_set1_pd(d_scalar);
 154     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 155     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 156     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 157     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 158     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 159     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 160
 161     /* Avoid stupid compiler warnings */
 162     jnrA = jnrB = 0;
 163     j_coord_offsetA = 0;
 164     j_coord_offsetB = 0;
 165
 166     outeriter        = 0;
 167     inneriter        = 0;
 168
 169     /* Start outer loop over neighborlists */
 170     for(iidx=0; iidx<nri; iidx++)
 171     {
 172         /* Load shift vector for this list */
 173         i_shift_offset   = DIM*shiftidx[iidx];
 174
 175         /* Load limits for loop over neighbors */
 176         j_index_start    = jindex[iidx];
 177         j_index_end      = jindex[iidx+1];
 178
 179         /* Get outer coordinate index */
 180         inr              = iinr[iidx];
 181         i_coord_offset   = DIM*inr;
 182
 183         /* Load i particle coords and add shift vector */
 184         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 185                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 186
 187         fix0             = _mm_setzero_pd();
 188         fiy0             = _mm_setzero_pd();
 189         fiz0             = _mm_setzero_pd();
 190         fix1             = _mm_setzero_pd();
 191         fiy1             = _mm_setzero_pd();
 192         fiz1             = _mm_setzero_pd();
 193         fix2             = _mm_setzero_pd();
 194         fiy2             = _mm_setzero_pd();
 195         fiz2             = _mm_setzero_pd();
 196
 197         /* Reset potential sums */
 198         velecsum         = _mm_setzero_pd();
 199         vvdwsum          = _mm_setzero_pd();
 200
 201         /* Start inner kernel loop */
 202         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 203         {
 204
 205             /* Get j neighbor index, and coordinate index */
 206             jnrA             = jjnr[jidx];
 207             jnrB             = jjnr[jidx+1];
 208             j_coord_offsetA  = DIM*jnrA;
 209             j_coord_offsetB  = DIM*jnrB;
 210
 211             /* load j atom coordinates */
 212             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 214
 215             /* Calculate displacement vector */
 216             dx00             = _mm_sub_pd(ix0,jx0);
 217             dy00             = _mm_sub_pd(iy0,jy0);
 218             dz00             = _mm_sub_pd(iz0,jz0);
 219             dx01             = _mm_sub_pd(ix0,jx1);
 220             dy01             = _mm_sub_pd(iy0,jy1);
 221             dz01             = _mm_sub_pd(iz0,jz1);
 222             dx02             = _mm_sub_pd(ix0,jx2);
 223             dy02             = _mm_sub_pd(iy0,jy2);
 224             dz02             = _mm_sub_pd(iz0,jz2);
 225             dx10             = _mm_sub_pd(ix1,jx0);
 226             dy10             = _mm_sub_pd(iy1,jy0);
 227             dz10             = _mm_sub_pd(iz1,jz0);
 228             dx11             = _mm_sub_pd(ix1,jx1);
 229             dy11             = _mm_sub_pd(iy1,jy1);
 230             dz11             = _mm_sub_pd(iz1,jz1);
 231             dx12             = _mm_sub_pd(ix1,jx2);
 232             dy12             = _mm_sub_pd(iy1,jy2);
 233             dz12             = _mm_sub_pd(iz1,jz2);
 234             dx20             = _mm_sub_pd(ix2,jx0);
 235             dy20             = _mm_sub_pd(iy2,jy0);
 236             dz20             = _mm_sub_pd(iz2,jz0);
 237             dx21             = _mm_sub_pd(ix2,jx1);
 238             dy21             = _mm_sub_pd(iy2,jy1);
 239             dz21             = _mm_sub_pd(iz2,jz1);
 240             dx22             = _mm_sub_pd(ix2,jx2);
 241             dy22             = _mm_sub_pd(iy2,jy2);
 242             dz22             = _mm_sub_pd(iz2,jz2);
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 246             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 247             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 248             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 249             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 250             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 251             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 252             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 253             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 254
 255             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 256             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 257             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 258             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 259             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 260             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 261             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 262             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 263             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 264
 265             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 266             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 267             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 268             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 269             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 270             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 271             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 272             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 273             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 274
 275             fjx0             = _mm_setzero_pd();
 276             fjy0             = _mm_setzero_pd();
 277             fjz0             = _mm_setzero_pd();
 278             fjx1             = _mm_setzero_pd();
 279             fjy1             = _mm_setzero_pd();
 280             fjz1             = _mm_setzero_pd();
 281             fjx2             = _mm_setzero_pd();
 282             fjy2             = _mm_setzero_pd();
 283             fjz2             = _mm_setzero_pd();
 284
 285             /**************************
 286              * CALCULATE INTERACTIONS *
 287              **************************/
 288
 289             if (gmx_mm_any_lt(rsq00,rcutoff2))
 290             {
 291
 292             r00              = _mm_mul_pd(rsq00,rinv00);
 293
 294             /* REACTION-FIELD ELECTROSTATICS */
 295             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 296             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 297
 298             /* LENNARD-JONES DISPERSION/REPULSION */
 299
 300             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 301             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 302             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 303             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 304             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 305
 306             d                = _mm_sub_pd(r00,rswitch);
 307             d                = _mm_max_pd(d,_mm_setzero_pd());
 308             d2               = _mm_mul_pd(d,d);
 309             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_add_pd(swV3,_mm_mul_pd(d,_mm_add_pd(swV4,_mm_mul_pd(d,swV5)))))));
 310
 311             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
 312
 313             /* Evaluate switch function */
 314             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 315             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 316             vvdw             = _mm_mul_pd(vvdw,sw);
 317             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velec            = _mm_and_pd(velec,cutoff_mask);
 321             velecsum         = _mm_add_pd(velecsum,velec);
 322             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 323             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 324
 325             fscal            = _mm_add_pd(felec,fvdw);
 326
 327             fscal            = _mm_and_pd(fscal,cutoff_mask);
 328
 329             /* Calculate temporary vectorial force */
 330             tx               = _mm_mul_pd(fscal,dx00);
 331             ty               = _mm_mul_pd(fscal,dy00);
 332             tz               = _mm_mul_pd(fscal,dz00);
 333
 334             /* Update vectorial force */
 335             fix0             = _mm_add_pd(fix0,tx);
 336             fiy0             = _mm_add_pd(fiy0,ty);
 337             fiz0             = _mm_add_pd(fiz0,tz);
 338
 339             fjx0             = _mm_add_pd(fjx0,tx);
 340             fjy0             = _mm_add_pd(fjy0,ty);
 341             fjz0             = _mm_add_pd(fjz0,tz);
 342
 343             }
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             if (gmx_mm_any_lt(rsq01,rcutoff2))
 350             {
 351
 352             /* REACTION-FIELD ELECTROSTATICS */
 353             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 354             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 355
 356             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velec            = _mm_and_pd(velec,cutoff_mask);
 360             velecsum         = _mm_add_pd(velecsum,velec);
 361
 362             fscal            = felec;
 363
 364             fscal            = _mm_and_pd(fscal,cutoff_mask);
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = _mm_mul_pd(fscal,dx01);
 368             ty               = _mm_mul_pd(fscal,dy01);
 369             tz               = _mm_mul_pd(fscal,dz01);
 370
 371             /* Update vectorial force */
 372             fix0             = _mm_add_pd(fix0,tx);
 373             fiy0             = _mm_add_pd(fiy0,ty);
 374             fiz0             = _mm_add_pd(fiz0,tz);
 375
 376             fjx1             = _mm_add_pd(fjx1,tx);
 377             fjy1             = _mm_add_pd(fjy1,ty);
 378             fjz1             = _mm_add_pd(fjz1,tz);
 379
 380             }
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             if (gmx_mm_any_lt(rsq02,rcutoff2))
 387             {
 388
 389             /* REACTION-FIELD ELECTROSTATICS */
 390             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 391             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 392
 393             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velec            = _mm_and_pd(velec,cutoff_mask);
 397             velecsum         = _mm_add_pd(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401             fscal            = _mm_and_pd(fscal,cutoff_mask);
 402
 403             /* Calculate temporary vectorial force */
 404             tx               = _mm_mul_pd(fscal,dx02);
 405             ty               = _mm_mul_pd(fscal,dy02);
 406             tz               = _mm_mul_pd(fscal,dz02);
 407
 408             /* Update vectorial force */
 409             fix0             = _mm_add_pd(fix0,tx);
 410             fiy0             = _mm_add_pd(fiy0,ty);
 411             fiz0             = _mm_add_pd(fiz0,tz);
 412
 413             fjx2             = _mm_add_pd(fjx2,tx);
 414             fjy2             = _mm_add_pd(fjy2,ty);
 415             fjz2             = _mm_add_pd(fjz2,tz);
 416
 417             }
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             if (gmx_mm_any_lt(rsq10,rcutoff2))
 424             {
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 428             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 429
 430             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _mm_and_pd(velec,cutoff_mask);
 434             velecsum         = _mm_add_pd(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _mm_and_pd(fscal,cutoff_mask);
 439
 440             /* Calculate temporary vectorial force */
 441             tx               = _mm_mul_pd(fscal,dx10);
 442             ty               = _mm_mul_pd(fscal,dy10);
 443             tz               = _mm_mul_pd(fscal,dz10);
 444
 445             /* Update vectorial force */
 446             fix1             = _mm_add_pd(fix1,tx);
 447             fiy1             = _mm_add_pd(fiy1,ty);
 448             fiz1             = _mm_add_pd(fiz1,tz);
 449
 450             fjx0             = _mm_add_pd(fjx0,tx);
 451             fjy0             = _mm_add_pd(fjy0,ty);
 452             fjz0             = _mm_add_pd(fjz0,tz);
 453
 454             }
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             if (gmx_mm_any_lt(rsq11,rcutoff2))
 461             {
 462
 463             /* REACTION-FIELD ELECTROSTATICS */
 464             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 465             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 466
 467             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 468
 469             /* Update potential sum for this i atom from the interaction with this j atom. */
 470             velec            = _mm_and_pd(velec,cutoff_mask);
 471             velecsum         = _mm_add_pd(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475             fscal            = _mm_and_pd(fscal,cutoff_mask);
 476
 477             /* Calculate temporary vectorial force */
 478             tx               = _mm_mul_pd(fscal,dx11);
 479             ty               = _mm_mul_pd(fscal,dy11);
 480             tz               = _mm_mul_pd(fscal,dz11);
 481
 482             /* Update vectorial force */
 483             fix1             = _mm_add_pd(fix1,tx);
 484             fiy1             = _mm_add_pd(fiy1,ty);
 485             fiz1             = _mm_add_pd(fiz1,tz);
 486
 487             fjx1             = _mm_add_pd(fjx1,tx);
 488             fjy1             = _mm_add_pd(fjy1,ty);
 489             fjz1             = _mm_add_pd(fjz1,tz);
 490
 491             }
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             if (gmx_mm_any_lt(rsq12,rcutoff2))
 498             {
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 502             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 503
 504             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velec            = _mm_and_pd(velec,cutoff_mask);
 508             velecsum         = _mm_add_pd(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             fscal            = _mm_and_pd(fscal,cutoff_mask);
 513
 514             /* Calculate temporary vectorial force */
 515             tx               = _mm_mul_pd(fscal,dx12);
 516             ty               = _mm_mul_pd(fscal,dy12);
 517             tz               = _mm_mul_pd(fscal,dz12);
 518
 519             /* Update vectorial force */
 520             fix1             = _mm_add_pd(fix1,tx);
 521             fiy1             = _mm_add_pd(fiy1,ty);
 522             fiz1             = _mm_add_pd(fiz1,tz);
 523
 524             fjx2             = _mm_add_pd(fjx2,tx);
 525             fjy2             = _mm_add_pd(fjy2,ty);
 526             fjz2             = _mm_add_pd(fjz2,tz);
 527
 528             }
 529
 530             /**************************
 531              * CALCULATE INTERACTIONS *
 532              **************************/
 533
 534             if (gmx_mm_any_lt(rsq20,rcutoff2))
 535             {
 536
 537             /* REACTION-FIELD ELECTROSTATICS */
 538             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 539             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 540
 541             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 542
 543             /* Update potential sum for this i atom from the interaction with this j atom. */
 544             velec            = _mm_and_pd(velec,cutoff_mask);
 545             velecsum         = _mm_add_pd(velecsum,velec);
 546
 547             fscal            = felec;
 548
 549             fscal            = _mm_and_pd(fscal,cutoff_mask);
 550
 551             /* Calculate temporary vectorial force */
 552             tx               = _mm_mul_pd(fscal,dx20);
 553             ty               = _mm_mul_pd(fscal,dy20);
 554             tz               = _mm_mul_pd(fscal,dz20);
 555
 556             /* Update vectorial force */
 557             fix2             = _mm_add_pd(fix2,tx);
 558             fiy2             = _mm_add_pd(fiy2,ty);
 559             fiz2             = _mm_add_pd(fiz2,tz);
 560
 561             fjx0             = _mm_add_pd(fjx0,tx);
 562             fjy0             = _mm_add_pd(fjy0,ty);
 563             fjz0             = _mm_add_pd(fjz0,tz);
 564
 565             }
 566
 567             /**************************
 568              * CALCULATE INTERACTIONS *
 569              **************************/
 570
 571             if (gmx_mm_any_lt(rsq21,rcutoff2))
 572             {
 573
 574             /* REACTION-FIELD ELECTROSTATICS */
 575             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 576             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 577
 578             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 579
 580             /* Update potential sum for this i atom from the interaction with this j atom. */
 581             velec            = _mm_and_pd(velec,cutoff_mask);
 582             velecsum         = _mm_add_pd(velecsum,velec);
 583
 584             fscal            = felec;
 585
 586             fscal            = _mm_and_pd(fscal,cutoff_mask);
 587
 588             /* Calculate temporary vectorial force */
 589             tx               = _mm_mul_pd(fscal,dx21);
 590             ty               = _mm_mul_pd(fscal,dy21);
 591             tz               = _mm_mul_pd(fscal,dz21);
 592
 593             /* Update vectorial force */
 594             fix2             = _mm_add_pd(fix2,tx);
 595             fiy2             = _mm_add_pd(fiy2,ty);
 596             fiz2             = _mm_add_pd(fiz2,tz);
 597
 598             fjx1             = _mm_add_pd(fjx1,tx);
 599             fjy1             = _mm_add_pd(fjy1,ty);
 600             fjz1             = _mm_add_pd(fjz1,tz);
 601
 602             }
 603
 604             /**************************
 605              * CALCULATE INTERACTIONS *
 606              **************************/
 607
 608             if (gmx_mm_any_lt(rsq22,rcutoff2))
 609             {
 610
 611             /* REACTION-FIELD ELECTROSTATICS */
 612             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 613             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 614
 615             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 616
 617             /* Update potential sum for this i atom from the interaction with this j atom. */
 618             velec            = _mm_and_pd(velec,cutoff_mask);
 619             velecsum         = _mm_add_pd(velecsum,velec);
 620
 621             fscal            = felec;
 622
 623             fscal            = _mm_and_pd(fscal,cutoff_mask);
 624
 625             /* Calculate temporary vectorial force */
 626             tx               = _mm_mul_pd(fscal,dx22);
 627             ty               = _mm_mul_pd(fscal,dy22);
 628             tz               = _mm_mul_pd(fscal,dz22);
 629
 630             /* Update vectorial force */
 631             fix2             = _mm_add_pd(fix2,tx);
 632             fiy2             = _mm_add_pd(fiy2,ty);
 633             fiz2             = _mm_add_pd(fiz2,tz);
 634
 635             fjx2             = _mm_add_pd(fjx2,tx);
 636             fjy2             = _mm_add_pd(fjy2,ty);
 637             fjz2             = _mm_add_pd(fjz2,tz);
 638
 639             }
 640
 641             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 642
 643             /* Inner loop uses 358 flops */
 644         }
 645
 646         if(jidx<j_index_end)
 647         {
 648
 649             jnrA             = jjnr[jidx];
 650             j_coord_offsetA  = DIM*jnrA;
 651
 652             /* load j atom coordinates */
 653             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 654                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 655
 656             /* Calculate displacement vector */
 657             dx00             = _mm_sub_pd(ix0,jx0);
 658             dy00             = _mm_sub_pd(iy0,jy0);
 659             dz00             = _mm_sub_pd(iz0,jz0);
 660             dx01             = _mm_sub_pd(ix0,jx1);
 661             dy01             = _mm_sub_pd(iy0,jy1);
 662             dz01             = _mm_sub_pd(iz0,jz1);
 663             dx02             = _mm_sub_pd(ix0,jx2);
 664             dy02             = _mm_sub_pd(iy0,jy2);
 665             dz02             = _mm_sub_pd(iz0,jz2);
 666             dx10             = _mm_sub_pd(ix1,jx0);
 667             dy10             = _mm_sub_pd(iy1,jy0);
 668             dz10             = _mm_sub_pd(iz1,jz0);
 669             dx11             = _mm_sub_pd(ix1,jx1);
 670             dy11             = _mm_sub_pd(iy1,jy1);
 671             dz11             = _mm_sub_pd(iz1,jz1);
 672             dx12             = _mm_sub_pd(ix1,jx2);
 673             dy12             = _mm_sub_pd(iy1,jy2);
 674             dz12             = _mm_sub_pd(iz1,jz2);
 675             dx20             = _mm_sub_pd(ix2,jx0);
 676             dy20             = _mm_sub_pd(iy2,jy0);
 677             dz20             = _mm_sub_pd(iz2,jz0);
 678             dx21             = _mm_sub_pd(ix2,jx1);
 679             dy21             = _mm_sub_pd(iy2,jy1);
 680             dz21             = _mm_sub_pd(iz2,jz1);
 681             dx22             = _mm_sub_pd(ix2,jx2);
 682             dy22             = _mm_sub_pd(iy2,jy2);
 683             dz22             = _mm_sub_pd(iz2,jz2);
 684
 685             /* Calculate squared distance and things based on it */
 686             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 687             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 688             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 689             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 690             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 691             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 692             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 693             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 694             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 695
 696             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 697             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 698             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 699             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 700             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 701             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 702             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 703             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 704             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 705
 706             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 707             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 708             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 709             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 710             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 711             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 712             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 713             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 714             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 715
 716             fjx0             = _mm_setzero_pd();
 717             fjy0             = _mm_setzero_pd();
 718             fjz0             = _mm_setzero_pd();
 719             fjx1             = _mm_setzero_pd();
 720             fjy1             = _mm_setzero_pd();
 721             fjz1             = _mm_setzero_pd();
 722             fjx2             = _mm_setzero_pd();
 723             fjy2             = _mm_setzero_pd();
 724             fjz2             = _mm_setzero_pd();
 725
 726             /**************************
 727              * CALCULATE INTERACTIONS *
 728              **************************/
 729
 730             if (gmx_mm_any_lt(rsq00,rcutoff2))
 731             {
 732
 733             r00              = _mm_mul_pd(rsq00,rinv00);
 734
 735             /* REACTION-FIELD ELECTROSTATICS */
 736             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 737             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 738
 739             /* LENNARD-JONES DISPERSION/REPULSION */
 740
 741             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 742             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 743             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 744             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 745             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 746
 747             d                = _mm_sub_pd(r00,rswitch);
 748             d                = _mm_max_pd(d,_mm_setzero_pd());
 749             d2               = _mm_mul_pd(d,d);
 750             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_add_pd(swV3,_mm_mul_pd(d,_mm_add_pd(swV4,_mm_mul_pd(d,swV5)))))));
 751
 752             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
 753
 754             /* Evaluate switch function */
 755             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 756             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 757             vvdw             = _mm_mul_pd(vvdw,sw);
 758             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _mm_and_pd(velec,cutoff_mask);
 762             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 763             velecsum         = _mm_add_pd(velecsum,velec);
 764             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 765             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 766             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 767
 768             fscal            = _mm_add_pd(felec,fvdw);
 769
 770             fscal            = _mm_and_pd(fscal,cutoff_mask);
 771
 772             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 773
 774             /* Calculate temporary vectorial force */
 775             tx               = _mm_mul_pd(fscal,dx00);
 776             ty               = _mm_mul_pd(fscal,dy00);
 777             tz               = _mm_mul_pd(fscal,dz00);
 778
 779             /* Update vectorial force */
 780             fix0             = _mm_add_pd(fix0,tx);
 781             fiy0             = _mm_add_pd(fiy0,ty);
 782             fiz0             = _mm_add_pd(fiz0,tz);
 783
 784             fjx0             = _mm_add_pd(fjx0,tx);
 785             fjy0             = _mm_add_pd(fjy0,ty);
 786             fjz0             = _mm_add_pd(fjz0,tz);
 787
 788             }
 789
 790             /**************************
 791              * CALCULATE INTERACTIONS *
 792              **************************/
 793
 794             if (gmx_mm_any_lt(rsq01,rcutoff2))
 795             {
 796
 797             /* REACTION-FIELD ELECTROSTATICS */
 798             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 799             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 800
 801             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 802
 803             /* Update potential sum for this i atom from the interaction with this j atom. */
 804             velec            = _mm_and_pd(velec,cutoff_mask);
 805             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 806             velecsum         = _mm_add_pd(velecsum,velec);
 807
 808             fscal            = felec;
 809
 810             fscal            = _mm_and_pd(fscal,cutoff_mask);
 811
 812             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 813
 814             /* Calculate temporary vectorial force */
 815             tx               = _mm_mul_pd(fscal,dx01);
 816             ty               = _mm_mul_pd(fscal,dy01);
 817             tz               = _mm_mul_pd(fscal,dz01);
 818
 819             /* Update vectorial force */
 820             fix0             = _mm_add_pd(fix0,tx);
 821             fiy0             = _mm_add_pd(fiy0,ty);
 822             fiz0             = _mm_add_pd(fiz0,tz);
 823
 824             fjx1             = _mm_add_pd(fjx1,tx);
 825             fjy1             = _mm_add_pd(fjy1,ty);
 826             fjz1             = _mm_add_pd(fjz1,tz);
 827
 828             }
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             if (gmx_mm_any_lt(rsq02,rcutoff2))
 835             {
 836
 837             /* REACTION-FIELD ELECTROSTATICS */
 838             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 839             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 840
 841             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 842
 843             /* Update potential sum for this i atom from the interaction with this j atom. */
 844             velec            = _mm_and_pd(velec,cutoff_mask);
 845             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 846             velecsum         = _mm_add_pd(velecsum,velec);
 847
 848             fscal            = felec;
 849
 850             fscal            = _mm_and_pd(fscal,cutoff_mask);
 851
 852             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 853
 854             /* Calculate temporary vectorial force */
 855             tx               = _mm_mul_pd(fscal,dx02);
 856             ty               = _mm_mul_pd(fscal,dy02);
 857             tz               = _mm_mul_pd(fscal,dz02);
 858
 859             /* Update vectorial force */
 860             fix0             = _mm_add_pd(fix0,tx);
 861             fiy0             = _mm_add_pd(fiy0,ty);
 862             fiz0             = _mm_add_pd(fiz0,tz);
 863
 864             fjx2             = _mm_add_pd(fjx2,tx);
 865             fjy2             = _mm_add_pd(fjy2,ty);
 866             fjz2             = _mm_add_pd(fjz2,tz);
 867
 868             }
 869
 870             /**************************
 871              * CALCULATE INTERACTIONS *
 872              **************************/
 873
 874             if (gmx_mm_any_lt(rsq10,rcutoff2))
 875             {
 876
 877             /* REACTION-FIELD ELECTROSTATICS */
 878             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 879             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 880
 881             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_and_pd(velec,cutoff_mask);
 885             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 886             velecsum         = _mm_add_pd(velecsum,velec);
 887
 888             fscal            = felec;
 889
 890             fscal            = _mm_and_pd(fscal,cutoff_mask);
 891
 892             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 893
 894             /* Calculate temporary vectorial force */
 895             tx               = _mm_mul_pd(fscal,dx10);
 896             ty               = _mm_mul_pd(fscal,dy10);
 897             tz               = _mm_mul_pd(fscal,dz10);
 898
 899             /* Update vectorial force */
 900             fix1             = _mm_add_pd(fix1,tx);
 901             fiy1             = _mm_add_pd(fiy1,ty);
 902             fiz1             = _mm_add_pd(fiz1,tz);
 903
 904             fjx0             = _mm_add_pd(fjx0,tx);
 905             fjy0             = _mm_add_pd(fjy0,ty);
 906             fjz0             = _mm_add_pd(fjz0,tz);
 907
 908             }
 909
 910             /**************************
 911              * CALCULATE INTERACTIONS *
 912              **************************/
 913
 914             if (gmx_mm_any_lt(rsq11,rcutoff2))
 915             {
 916
 917             /* REACTION-FIELD ELECTROSTATICS */
 918             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 919             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 920
 921             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_and_pd(velec,cutoff_mask);
 925             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 926             velecsum         = _mm_add_pd(velecsum,velec);
 927
 928             fscal            = felec;
 929
 930             fscal            = _mm_and_pd(fscal,cutoff_mask);
 931
 932             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 933
 934             /* Calculate temporary vectorial force */
 935             tx               = _mm_mul_pd(fscal,dx11);
 936             ty               = _mm_mul_pd(fscal,dy11);
 937             tz               = _mm_mul_pd(fscal,dz11);
 938
 939             /* Update vectorial force */
 940             fix1             = _mm_add_pd(fix1,tx);
 941             fiy1             = _mm_add_pd(fiy1,ty);
 942             fiz1             = _mm_add_pd(fiz1,tz);
 943
 944             fjx1             = _mm_add_pd(fjx1,tx);
 945             fjy1             = _mm_add_pd(fjy1,ty);
 946             fjz1             = _mm_add_pd(fjz1,tz);
 947
 948             }
 949
 950             /**************************
 951              * CALCULATE INTERACTIONS *
 952              **************************/
 953
 954             if (gmx_mm_any_lt(rsq12,rcutoff2))
 955             {
 956
 957             /* REACTION-FIELD ELECTROSTATICS */
 958             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 959             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 960
 961             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 962
 963             /* Update potential sum for this i atom from the interaction with this j atom. */
 964             velec            = _mm_and_pd(velec,cutoff_mask);
 965             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 966             velecsum         = _mm_add_pd(velecsum,velec);
 967
 968             fscal            = felec;
 969
 970             fscal            = _mm_and_pd(fscal,cutoff_mask);
 971
 972             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 973
 974             /* Calculate temporary vectorial force */
 975             tx               = _mm_mul_pd(fscal,dx12);
 976             ty               = _mm_mul_pd(fscal,dy12);
 977             tz               = _mm_mul_pd(fscal,dz12);
 978
 979             /* Update vectorial force */
 980             fix1             = _mm_add_pd(fix1,tx);
 981             fiy1             = _mm_add_pd(fiy1,ty);
 982             fiz1             = _mm_add_pd(fiz1,tz);
 983
 984             fjx2             = _mm_add_pd(fjx2,tx);
 985             fjy2             = _mm_add_pd(fjy2,ty);
 986             fjz2             = _mm_add_pd(fjz2,tz);
 987
 988             }
 989
 990             /**************************
 991              * CALCULATE INTERACTIONS *
 992              **************************/
 993
 994             if (gmx_mm_any_lt(rsq20,rcutoff2))
 995             {
 996
 997             /* REACTION-FIELD ELECTROSTATICS */
 998             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 999             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1000
1001             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1002
1003             /* Update potential sum for this i atom from the interaction with this j atom. */
1004             velec            = _mm_and_pd(velec,cutoff_mask);
1005             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1006             velecsum         = _mm_add_pd(velecsum,velec);
1007
1008             fscal            = felec;
1009
1010             fscal            = _mm_and_pd(fscal,cutoff_mask);
1011
1012             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1013
1014             /* Calculate temporary vectorial force */
1015             tx               = _mm_mul_pd(fscal,dx20);
1016             ty               = _mm_mul_pd(fscal,dy20);
1017             tz               = _mm_mul_pd(fscal,dz20);
1018
1019             /* Update vectorial force */
1020             fix2             = _mm_add_pd(fix2,tx);
1021             fiy2             = _mm_add_pd(fiy2,ty);
1022             fiz2             = _mm_add_pd(fiz2,tz);
1023
1024             fjx0             = _mm_add_pd(fjx0,tx);
1025             fjy0             = _mm_add_pd(fjy0,ty);
1026             fjz0             = _mm_add_pd(fjz0,tz);
1027
1028             }
1029
1030             /**************************
1031              * CALCULATE INTERACTIONS *
1032              **************************/
1033
1034             if (gmx_mm_any_lt(rsq21,rcutoff2))
1035             {
1036
1037             /* REACTION-FIELD ELECTROSTATICS */
1038             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
1039             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1040
1041             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1042
1043             /* Update potential sum for this i atom from the interaction with this j atom. */
1044             velec            = _mm_and_pd(velec,cutoff_mask);
1045             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1046             velecsum         = _mm_add_pd(velecsum,velec);
1047
1048             fscal            = felec;
1049
1050             fscal            = _mm_and_pd(fscal,cutoff_mask);
1051
1052             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1053
1054             /* Calculate temporary vectorial force */
1055             tx               = _mm_mul_pd(fscal,dx21);
1056             ty               = _mm_mul_pd(fscal,dy21);
1057             tz               = _mm_mul_pd(fscal,dz21);
1058
1059             /* Update vectorial force */
1060             fix2             = _mm_add_pd(fix2,tx);
1061             fiy2             = _mm_add_pd(fiy2,ty);
1062             fiz2             = _mm_add_pd(fiz2,tz);
1063
1064             fjx1             = _mm_add_pd(fjx1,tx);
1065             fjy1             = _mm_add_pd(fjy1,ty);
1066             fjz1             = _mm_add_pd(fjz1,tz);
1067
1068             }
1069
1070             /**************************
1071              * CALCULATE INTERACTIONS *
1072              **************************/
1073
1074             if (gmx_mm_any_lt(rsq22,rcutoff2))
1075             {
1076
1077             /* REACTION-FIELD ELECTROSTATICS */
1078             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1079             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1080
1081             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1082
1083             /* Update potential sum for this i atom from the interaction with this j atom. */
1084             velec            = _mm_and_pd(velec,cutoff_mask);
1085             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1086             velecsum         = _mm_add_pd(velecsum,velec);
1087
1088             fscal            = felec;
1089
1090             fscal            = _mm_and_pd(fscal,cutoff_mask);
1091
1092             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1093
1094             /* Calculate temporary vectorial force */
1095             tx               = _mm_mul_pd(fscal,dx22);
1096             ty               = _mm_mul_pd(fscal,dy22);
1097             tz               = _mm_mul_pd(fscal,dz22);
1098
1099             /* Update vectorial force */
1100             fix2             = _mm_add_pd(fix2,tx);
1101             fiy2             = _mm_add_pd(fiy2,ty);
1102             fiz2             = _mm_add_pd(fiz2,tz);
1103
1104             fjx2             = _mm_add_pd(fjx2,tx);
1105             fjy2             = _mm_add_pd(fjy2,ty);
1106             fjz2             = _mm_add_pd(fjz2,tz);
1107
1108             }
1109
1110             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1111
1112             /* Inner loop uses 358 flops */
1113         }
1114
1115         /* End of innermost loop */
1116
1117         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1118                                               f+i_coord_offset,fshift+i_shift_offset);
1119
1120         ggid                        = gid[iidx];
1121         /* Update potential energies */
1122         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1123         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1124
1125         /* Increment number of inner iterations */
1126         inneriter                  += j_index_end - j_index_start;
1127
1128         /* Outer loop uses 20 flops */
1129     }
1130
1131     /* Increment number of outer iterations */
1132     outeriter        += nri;
1133
1134     /* Update outer/inner flops */
1135
1136     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*358);
1137 }
1138 /*
1139  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_sse4_1_double
1140  * Electrostatics interaction: ReactionField
1141  * VdW interaction:            LennardJones
1142  * Geometry:                   Water3-Water3
1143  * Calculate force/pot:        Force
1144  */
1145 void
1146 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_sse4_1_double
1147                     (t_nblist * gmx_restrict                nlist,
1148                      rvec * gmx_restrict                    xx,
1149                      rvec * gmx_restrict                    ff,
1150                      t_forcerec * gmx_restrict              fr,
1151                      t_mdatoms * gmx_restrict               mdatoms,
1152                      nb_kernel_data_t * gmx_restrict        kernel_data,
1153                      t_nrnb * gmx_restrict                  nrnb)
1154 {
1155     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1156      * just 0 for non-waters.
1157      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1158      * jnr indices corresponding to data put in the four positions in the SIMD register.
1159      */
1160     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1161     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1162     int              jnrA,jnrB;
1163     int              j_coord_offsetA,j_coord_offsetB;
1164     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1165     real             rcutoff_scalar;
1166     real             *shiftvec,*fshift,*x,*f;
1167     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1168     int              vdwioffset0;
1169     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1170     int              vdwioffset1;
1171     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1172     int              vdwioffset2;
1173     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1174     int              vdwjidx0A,vdwjidx0B;
1175     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1176     int              vdwjidx1A,vdwjidx1B;
1177     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1178     int              vdwjidx2A,vdwjidx2B;
1179     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1180     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1181     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1182     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1183     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1184     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1185     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1186     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1187     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1188     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1189     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1190     real             *charge;
1191     int              nvdwtype;
1192     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1193     int              *vdwtype;
1194     real             *vdwparam;
1195     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1196     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1197     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1198     real             rswitch_scalar,d_scalar;
1199     __m128d          dummy_mask,cutoff_mask;
1200     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1201     __m128d          one     = _mm_set1_pd(1.0);
1202     __m128d          two     = _mm_set1_pd(2.0);
1203     x                = xx[0];
1204     f                = ff[0];
1205
1206     nri              = nlist->nri;
1207     iinr             = nlist->iinr;
1208     jindex           = nlist->jindex;
1209     jjnr             = nlist->jjnr;
1210     shiftidx         = nlist->shift;
1211     gid              = nlist->gid;
1212     shiftvec         = fr->shift_vec[0];
1213     fshift           = fr->fshift[0];
1214     facel            = _mm_set1_pd(fr->epsfac);
1215     charge           = mdatoms->chargeA;
1216     krf              = _mm_set1_pd(fr->ic->k_rf);
1217     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1218     crf              = _mm_set1_pd(fr->ic->c_rf);
1219     nvdwtype         = fr->ntype;
1220     vdwparam         = fr->nbfp;
1221     vdwtype          = mdatoms->typeA;
1222
1223     /* Setup water-specific parameters */
1224     inr              = nlist->iinr[0];
1225     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1226     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1227     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1228     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1229
1230     jq0              = _mm_set1_pd(charge[inr+0]);
1231     jq1              = _mm_set1_pd(charge[inr+1]);
1232     jq2              = _mm_set1_pd(charge[inr+2]);
1233     vdwjidx0A        = 2*vdwtype[inr+0];
1234     qq00             = _mm_mul_pd(iq0,jq0);
1235     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1236     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1237     qq01             = _mm_mul_pd(iq0,jq1);
1238     qq02             = _mm_mul_pd(iq0,jq2);
1239     qq10             = _mm_mul_pd(iq1,jq0);
1240     qq11             = _mm_mul_pd(iq1,jq1);
1241     qq12             = _mm_mul_pd(iq1,jq2);
1242     qq20             = _mm_mul_pd(iq2,jq0);
1243     qq21             = _mm_mul_pd(iq2,jq1);
1244     qq22             = _mm_mul_pd(iq2,jq2);
1245
1246     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1247     rcutoff_scalar   = fr->rcoulomb;
1248     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1249     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1250
1251     rswitch_scalar   = fr->rvdw_switch;
1252     rswitch          = _mm_set1_pd(rswitch_scalar);
1253     /* Setup switch parameters */
1254     d_scalar         = rcutoff_scalar-rswitch_scalar;
1255     d                = _mm_set1_pd(d_scalar);
1256     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1257     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1258     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1259     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1260     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1261     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1262
1263     /* Avoid stupid compiler warnings */
1264     jnrA = jnrB = 0;
1265     j_coord_offsetA = 0;
1266     j_coord_offsetB = 0;
1267
1268     outeriter        = 0;
1269     inneriter        = 0;
1270
1271     /* Start outer loop over neighborlists */
1272     for(iidx=0; iidx<nri; iidx++)
1273     {
1274         /* Load shift vector for this list */
1275         i_shift_offset   = DIM*shiftidx[iidx];
1276
1277         /* Load limits for loop over neighbors */
1278         j_index_start    = jindex[iidx];
1279         j_index_end      = jindex[iidx+1];
1280
1281         /* Get outer coordinate index */
1282         inr              = iinr[iidx];
1283         i_coord_offset   = DIM*inr;
1284
1285         /* Load i particle coords and add shift vector */
1286         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1287                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1288
1289         fix0             = _mm_setzero_pd();
1290         fiy0             = _mm_setzero_pd();
1291         fiz0             = _mm_setzero_pd();
1292         fix1             = _mm_setzero_pd();
1293         fiy1             = _mm_setzero_pd();
1294         fiz1             = _mm_setzero_pd();
1295         fix2             = _mm_setzero_pd();
1296         fiy2             = _mm_setzero_pd();
1297         fiz2             = _mm_setzero_pd();
1298
1299         /* Start inner kernel loop */
1300         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1301         {
1302
1303             /* Get j neighbor index, and coordinate index */
1304             jnrA             = jjnr[jidx];
1305             jnrB             = jjnr[jidx+1];
1306             j_coord_offsetA  = DIM*jnrA;
1307             j_coord_offsetB  = DIM*jnrB;
1308
1309             /* load j atom coordinates */
1310             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1311                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1312
1313             /* Calculate displacement vector */
1314             dx00             = _mm_sub_pd(ix0,jx0);
1315             dy00             = _mm_sub_pd(iy0,jy0);
1316             dz00             = _mm_sub_pd(iz0,jz0);
1317             dx01             = _mm_sub_pd(ix0,jx1);
1318             dy01             = _mm_sub_pd(iy0,jy1);
1319             dz01             = _mm_sub_pd(iz0,jz1);
1320             dx02             = _mm_sub_pd(ix0,jx2);
1321             dy02             = _mm_sub_pd(iy0,jy2);
1322             dz02             = _mm_sub_pd(iz0,jz2);
1323             dx10             = _mm_sub_pd(ix1,jx0);
1324             dy10             = _mm_sub_pd(iy1,jy0);
1325             dz10             = _mm_sub_pd(iz1,jz0);
1326             dx11             = _mm_sub_pd(ix1,jx1);
1327             dy11             = _mm_sub_pd(iy1,jy1);
1328             dz11             = _mm_sub_pd(iz1,jz1);
1329             dx12             = _mm_sub_pd(ix1,jx2);
1330             dy12             = _mm_sub_pd(iy1,jy2);
1331             dz12             = _mm_sub_pd(iz1,jz2);
1332             dx20             = _mm_sub_pd(ix2,jx0);
1333             dy20             = _mm_sub_pd(iy2,jy0);
1334             dz20             = _mm_sub_pd(iz2,jz0);
1335             dx21             = _mm_sub_pd(ix2,jx1);
1336             dy21             = _mm_sub_pd(iy2,jy1);
1337             dz21             = _mm_sub_pd(iz2,jz1);
1338             dx22             = _mm_sub_pd(ix2,jx2);
1339             dy22             = _mm_sub_pd(iy2,jy2);
1340             dz22             = _mm_sub_pd(iz2,jz2);
1341
1342             /* Calculate squared distance and things based on it */
1343             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1344             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1345             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1346             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1347             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1348             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1349             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1350             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1351             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1352
1353             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1354             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1355             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1356             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1357             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1358             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1359             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1360             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1361             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1362
1363             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1364             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1365             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1366             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1367             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1368             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1369             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1370             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1371             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1372
1373             fjx0             = _mm_setzero_pd();
1374             fjy0             = _mm_setzero_pd();
1375             fjz0             = _mm_setzero_pd();
1376             fjx1             = _mm_setzero_pd();
1377             fjy1             = _mm_setzero_pd();
1378             fjz1             = _mm_setzero_pd();
1379             fjx2             = _mm_setzero_pd();
1380             fjy2             = _mm_setzero_pd();
1381             fjz2             = _mm_setzero_pd();
1382
1383             /**************************
1384              * CALCULATE INTERACTIONS *
1385              **************************/
1386
1387             if (gmx_mm_any_lt(rsq00,rcutoff2))
1388             {
1389
1390             r00              = _mm_mul_pd(rsq00,rinv00);
1391
1392             /* REACTION-FIELD ELECTROSTATICS */
1393             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1394
1395             /* LENNARD-JONES DISPERSION/REPULSION */
1396
1397             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1398             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1399             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1400             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
1401             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1402
1403             d                = _mm_sub_pd(r00,rswitch);
1404             d                = _mm_max_pd(d,_mm_setzero_pd());
1405             d2               = _mm_mul_pd(d,d);
1406             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_add_pd(swV3,_mm_mul_pd(d,_mm_add_pd(swV4,_mm_mul_pd(d,swV5)))))));
1407
1408             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
1409
1410             /* Evaluate switch function */
1411             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1412             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1413             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1414
1415             fscal            = _mm_add_pd(felec,fvdw);
1416
1417             fscal            = _mm_and_pd(fscal,cutoff_mask);
1418
1419             /* Calculate temporary vectorial force */
1420             tx               = _mm_mul_pd(fscal,dx00);
1421             ty               = _mm_mul_pd(fscal,dy00);
1422             tz               = _mm_mul_pd(fscal,dz00);
1423
1424             /* Update vectorial force */
1425             fix0             = _mm_add_pd(fix0,tx);
1426             fiy0             = _mm_add_pd(fiy0,ty);
1427             fiz0             = _mm_add_pd(fiz0,tz);
1428
1429             fjx0             = _mm_add_pd(fjx0,tx);
1430             fjy0             = _mm_add_pd(fjy0,ty);
1431             fjz0             = _mm_add_pd(fjz0,tz);
1432
1433             }
1434
1435             /**************************
1436              * CALCULATE INTERACTIONS *
1437              **************************/
1438
1439             if (gmx_mm_any_lt(rsq01,rcutoff2))
1440             {
1441
1442             /* REACTION-FIELD ELECTROSTATICS */
1443             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1444
1445             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1446
1447             fscal            = felec;
1448
1449             fscal            = _mm_and_pd(fscal,cutoff_mask);
1450
1451             /* Calculate temporary vectorial force */
1452             tx               = _mm_mul_pd(fscal,dx01);
1453             ty               = _mm_mul_pd(fscal,dy01);
1454             tz               = _mm_mul_pd(fscal,dz01);
1455
1456             /* Update vectorial force */
1457             fix0             = _mm_add_pd(fix0,tx);
1458             fiy0             = _mm_add_pd(fiy0,ty);
1459             fiz0             = _mm_add_pd(fiz0,tz);
1460
1461             fjx1             = _mm_add_pd(fjx1,tx);
1462             fjy1             = _mm_add_pd(fjy1,ty);
1463             fjz1             = _mm_add_pd(fjz1,tz);
1464
1465             }
1466
1467             /**************************
1468              * CALCULATE INTERACTIONS *
1469              **************************/
1470
1471             if (gmx_mm_any_lt(rsq02,rcutoff2))
1472             {
1473
1474             /* REACTION-FIELD ELECTROSTATICS */
1475             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1476
1477             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1478
1479             fscal            = felec;
1480
1481             fscal            = _mm_and_pd(fscal,cutoff_mask);
1482
1483             /* Calculate temporary vectorial force */
1484             tx               = _mm_mul_pd(fscal,dx02);
1485             ty               = _mm_mul_pd(fscal,dy02);
1486             tz               = _mm_mul_pd(fscal,dz02);
1487
1488             /* Update vectorial force */
1489             fix0             = _mm_add_pd(fix0,tx);
1490             fiy0             = _mm_add_pd(fiy0,ty);
1491             fiz0             = _mm_add_pd(fiz0,tz);
1492
1493             fjx2             = _mm_add_pd(fjx2,tx);
1494             fjy2             = _mm_add_pd(fjy2,ty);
1495             fjz2             = _mm_add_pd(fjz2,tz);
1496
1497             }
1498
1499             /**************************
1500              * CALCULATE INTERACTIONS *
1501              **************************/
1502
1503             if (gmx_mm_any_lt(rsq10,rcutoff2))
1504             {
1505
1506             /* REACTION-FIELD ELECTROSTATICS */
1507             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1508
1509             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1510
1511             fscal            = felec;
1512
1513             fscal            = _mm_and_pd(fscal,cutoff_mask);
1514
1515             /* Calculate temporary vectorial force */
1516             tx               = _mm_mul_pd(fscal,dx10);
1517             ty               = _mm_mul_pd(fscal,dy10);
1518             tz               = _mm_mul_pd(fscal,dz10);
1519
1520             /* Update vectorial force */
1521             fix1             = _mm_add_pd(fix1,tx);
1522             fiy1             = _mm_add_pd(fiy1,ty);
1523             fiz1             = _mm_add_pd(fiz1,tz);
1524
1525             fjx0             = _mm_add_pd(fjx0,tx);
1526             fjy0             = _mm_add_pd(fjy0,ty);
1527             fjz0             = _mm_add_pd(fjz0,tz);
1528
1529             }
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             if (gmx_mm_any_lt(rsq11,rcutoff2))
1536             {
1537
1538             /* REACTION-FIELD ELECTROSTATICS */
1539             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1540
1541             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_and_pd(fscal,cutoff_mask);
1546
1547             /* Calculate temporary vectorial force */
1548             tx               = _mm_mul_pd(fscal,dx11);
1549             ty               = _mm_mul_pd(fscal,dy11);
1550             tz               = _mm_mul_pd(fscal,dz11);
1551
1552             /* Update vectorial force */
1553             fix1             = _mm_add_pd(fix1,tx);
1554             fiy1             = _mm_add_pd(fiy1,ty);
1555             fiz1             = _mm_add_pd(fiz1,tz);
1556
1557             fjx1             = _mm_add_pd(fjx1,tx);
1558             fjy1             = _mm_add_pd(fjy1,ty);
1559             fjz1             = _mm_add_pd(fjz1,tz);
1560
1561             }
1562
1563             /**************************
1564              * CALCULATE INTERACTIONS *
1565              **************************/
1566
1567             if (gmx_mm_any_lt(rsq12,rcutoff2))
1568             {
1569
1570             /* REACTION-FIELD ELECTROSTATICS */
1571             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1572
1573             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1574
1575             fscal            = felec;
1576
1577             fscal            = _mm_and_pd(fscal,cutoff_mask);
1578
1579             /* Calculate temporary vectorial force */
1580             tx               = _mm_mul_pd(fscal,dx12);
1581             ty               = _mm_mul_pd(fscal,dy12);
1582             tz               = _mm_mul_pd(fscal,dz12);
1583
1584             /* Update vectorial force */
1585             fix1             = _mm_add_pd(fix1,tx);
1586             fiy1             = _mm_add_pd(fiy1,ty);
1587             fiz1             = _mm_add_pd(fiz1,tz);
1588
1589             fjx2             = _mm_add_pd(fjx2,tx);
1590             fjy2             = _mm_add_pd(fjy2,ty);
1591             fjz2             = _mm_add_pd(fjz2,tz);
1592
1593             }
1594
1595             /**************************
1596              * CALCULATE INTERACTIONS *
1597              **************************/
1598
1599             if (gmx_mm_any_lt(rsq20,rcutoff2))
1600             {
1601
1602             /* REACTION-FIELD ELECTROSTATICS */
1603             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1604
1605             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1606
1607             fscal            = felec;
1608
1609             fscal            = _mm_and_pd(fscal,cutoff_mask);
1610
1611             /* Calculate temporary vectorial force */
1612             tx               = _mm_mul_pd(fscal,dx20);
1613             ty               = _mm_mul_pd(fscal,dy20);
1614             tz               = _mm_mul_pd(fscal,dz20);
1615
1616             /* Update vectorial force */
1617             fix2             = _mm_add_pd(fix2,tx);
1618             fiy2             = _mm_add_pd(fiy2,ty);
1619             fiz2             = _mm_add_pd(fiz2,tz);
1620
1621             fjx0             = _mm_add_pd(fjx0,tx);
1622             fjy0             = _mm_add_pd(fjy0,ty);
1623             fjz0             = _mm_add_pd(fjz0,tz);
1624
1625             }
1626
1627             /**************************
1628              * CALCULATE INTERACTIONS *
1629              **************************/
1630
1631             if (gmx_mm_any_lt(rsq21,rcutoff2))
1632             {
1633
1634             /* REACTION-FIELD ELECTROSTATICS */
1635             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1636
1637             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1638
1639             fscal            = felec;
1640
1641             fscal            = _mm_and_pd(fscal,cutoff_mask);
1642
1643             /* Calculate temporary vectorial force */
1644             tx               = _mm_mul_pd(fscal,dx21);
1645             ty               = _mm_mul_pd(fscal,dy21);
1646             tz               = _mm_mul_pd(fscal,dz21);
1647
1648             /* Update vectorial force */
1649             fix2             = _mm_add_pd(fix2,tx);
1650             fiy2             = _mm_add_pd(fiy2,ty);
1651             fiz2             = _mm_add_pd(fiz2,tz);
1652
1653             fjx1             = _mm_add_pd(fjx1,tx);
1654             fjy1             = _mm_add_pd(fjy1,ty);
1655             fjz1             = _mm_add_pd(fjz1,tz);
1656
1657             }
1658
1659             /**************************
1660              * CALCULATE INTERACTIONS *
1661              **************************/
1662
1663             if (gmx_mm_any_lt(rsq22,rcutoff2))
1664             {
1665
1666             /* REACTION-FIELD ELECTROSTATICS */
1667             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1668
1669             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1670
1671             fscal            = felec;
1672
1673             fscal            = _mm_and_pd(fscal,cutoff_mask);
1674
1675             /* Calculate temporary vectorial force */
1676             tx               = _mm_mul_pd(fscal,dx22);
1677             ty               = _mm_mul_pd(fscal,dy22);
1678             tz               = _mm_mul_pd(fscal,dz22);
1679
1680             /* Update vectorial force */
1681             fix2             = _mm_add_pd(fix2,tx);
1682             fiy2             = _mm_add_pd(fiy2,ty);
1683             fiz2             = _mm_add_pd(fiz2,tz);
1684
1685             fjx2             = _mm_add_pd(fjx2,tx);
1686             fjy2             = _mm_add_pd(fjy2,ty);
1687             fjz2             = _mm_add_pd(fjz2,tz);
1688
1689             }
1690
1691             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1692
1693             /* Inner loop uses 301 flops */
1694         }
1695
1696         if(jidx<j_index_end)
1697         {
1698
1699             jnrA             = jjnr[jidx];
1700             j_coord_offsetA  = DIM*jnrA;
1701
1702             /* load j atom coordinates */
1703             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1704                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1705
1706             /* Calculate displacement vector */
1707             dx00             = _mm_sub_pd(ix0,jx0);
1708             dy00             = _mm_sub_pd(iy0,jy0);
1709             dz00             = _mm_sub_pd(iz0,jz0);
1710             dx01             = _mm_sub_pd(ix0,jx1);
1711             dy01             = _mm_sub_pd(iy0,jy1);
1712             dz01             = _mm_sub_pd(iz0,jz1);
1713             dx02             = _mm_sub_pd(ix0,jx2);
1714             dy02             = _mm_sub_pd(iy0,jy2);
1715             dz02             = _mm_sub_pd(iz0,jz2);
1716             dx10             = _mm_sub_pd(ix1,jx0);
1717             dy10             = _mm_sub_pd(iy1,jy0);
1718             dz10             = _mm_sub_pd(iz1,jz0);
1719             dx11             = _mm_sub_pd(ix1,jx1);
1720             dy11             = _mm_sub_pd(iy1,jy1);
1721             dz11             = _mm_sub_pd(iz1,jz1);
1722             dx12             = _mm_sub_pd(ix1,jx2);
1723             dy12             = _mm_sub_pd(iy1,jy2);
1724             dz12             = _mm_sub_pd(iz1,jz2);
1725             dx20             = _mm_sub_pd(ix2,jx0);
1726             dy20             = _mm_sub_pd(iy2,jy0);
1727             dz20             = _mm_sub_pd(iz2,jz0);
1728             dx21             = _mm_sub_pd(ix2,jx1);
1729             dy21             = _mm_sub_pd(iy2,jy1);
1730             dz21             = _mm_sub_pd(iz2,jz1);
1731             dx22             = _mm_sub_pd(ix2,jx2);
1732             dy22             = _mm_sub_pd(iy2,jy2);
1733             dz22             = _mm_sub_pd(iz2,jz2);
1734
1735             /* Calculate squared distance and things based on it */
1736             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1737             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1738             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1739             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1740             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1741             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1742             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1743             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1744             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1745
1746             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1747             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1748             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1749             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1750             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1751             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1752             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1753             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1754             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1755
1756             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1757             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1758             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1759             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1760             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1761             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1762             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1763             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1764             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1765
1766             fjx0             = _mm_setzero_pd();
1767             fjy0             = _mm_setzero_pd();
1768             fjz0             = _mm_setzero_pd();
1769             fjx1             = _mm_setzero_pd();
1770             fjy1             = _mm_setzero_pd();
1771             fjz1             = _mm_setzero_pd();
1772             fjx2             = _mm_setzero_pd();
1773             fjy2             = _mm_setzero_pd();
1774             fjz2             = _mm_setzero_pd();
1775
1776             /**************************
1777              * CALCULATE INTERACTIONS *
1778              **************************/
1779
1780             if (gmx_mm_any_lt(rsq00,rcutoff2))
1781             {
1782
1783             r00              = _mm_mul_pd(rsq00,rinv00);
1784
1785             /* REACTION-FIELD ELECTROSTATICS */
1786             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1787
1788             /* LENNARD-JONES DISPERSION/REPULSION */
1789
1790             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1791             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1792             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1793             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
1794             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1795
1796             d                = _mm_sub_pd(r00,rswitch);
1797             d                = _mm_max_pd(d,_mm_setzero_pd());
1798             d2               = _mm_mul_pd(d,d);
1799             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_add_pd(swV3,_mm_mul_pd(d,_mm_add_pd(swV4,_mm_mul_pd(d,swV5)))))));
1800
1801             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
1802
1803             /* Evaluate switch function */
1804             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1805             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1806             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1807
1808             fscal            = _mm_add_pd(felec,fvdw);
1809
1810             fscal            = _mm_and_pd(fscal,cutoff_mask);
1811
1812             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1813
1814             /* Calculate temporary vectorial force */
1815             tx               = _mm_mul_pd(fscal,dx00);
1816             ty               = _mm_mul_pd(fscal,dy00);
1817             tz               = _mm_mul_pd(fscal,dz00);
1818
1819             /* Update vectorial force */
1820             fix0             = _mm_add_pd(fix0,tx);
1821             fiy0             = _mm_add_pd(fiy0,ty);
1822             fiz0             = _mm_add_pd(fiz0,tz);
1823
1824             fjx0             = _mm_add_pd(fjx0,tx);
1825             fjy0             = _mm_add_pd(fjy0,ty);
1826             fjz0             = _mm_add_pd(fjz0,tz);
1827
1828             }
1829
1830             /**************************
1831              * CALCULATE INTERACTIONS *
1832              **************************/
1833
1834             if (gmx_mm_any_lt(rsq01,rcutoff2))
1835             {
1836
1837             /* REACTION-FIELD ELECTROSTATICS */
1838             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1839
1840             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1841
1842             fscal            = felec;
1843
1844             fscal            = _mm_and_pd(fscal,cutoff_mask);
1845
1846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1847
1848             /* Calculate temporary vectorial force */
1849             tx               = _mm_mul_pd(fscal,dx01);
1850             ty               = _mm_mul_pd(fscal,dy01);
1851             tz               = _mm_mul_pd(fscal,dz01);
1852
1853             /* Update vectorial force */
1854             fix0             = _mm_add_pd(fix0,tx);
1855             fiy0             = _mm_add_pd(fiy0,ty);
1856             fiz0             = _mm_add_pd(fiz0,tz);
1857
1858             fjx1             = _mm_add_pd(fjx1,tx);
1859             fjy1             = _mm_add_pd(fjy1,ty);
1860             fjz1             = _mm_add_pd(fjz1,tz);
1861
1862             }
1863
1864             /**************************
1865              * CALCULATE INTERACTIONS *
1866              **************************/
1867
1868             if (gmx_mm_any_lt(rsq02,rcutoff2))
1869             {
1870
1871             /* REACTION-FIELD ELECTROSTATICS */
1872             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1873
1874             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1875
1876             fscal            = felec;
1877
1878             fscal            = _mm_and_pd(fscal,cutoff_mask);
1879
1880             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1881
1882             /* Calculate temporary vectorial force */
1883             tx               = _mm_mul_pd(fscal,dx02);
1884             ty               = _mm_mul_pd(fscal,dy02);
1885             tz               = _mm_mul_pd(fscal,dz02);
1886
1887             /* Update vectorial force */
1888             fix0             = _mm_add_pd(fix0,tx);
1889             fiy0             = _mm_add_pd(fiy0,ty);
1890             fiz0             = _mm_add_pd(fiz0,tz);
1891
1892             fjx2             = _mm_add_pd(fjx2,tx);
1893             fjy2             = _mm_add_pd(fjy2,ty);
1894             fjz2             = _mm_add_pd(fjz2,tz);
1895
1896             }
1897
1898             /**************************
1899              * CALCULATE INTERACTIONS *
1900              **************************/
1901
1902             if (gmx_mm_any_lt(rsq10,rcutoff2))
1903             {
1904
1905             /* REACTION-FIELD ELECTROSTATICS */
1906             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1907
1908             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1909
1910             fscal            = felec;
1911
1912             fscal            = _mm_and_pd(fscal,cutoff_mask);
1913
1914             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1915
1916             /* Calculate temporary vectorial force */
1917             tx               = _mm_mul_pd(fscal,dx10);
1918             ty               = _mm_mul_pd(fscal,dy10);
1919             tz               = _mm_mul_pd(fscal,dz10);
1920
1921             /* Update vectorial force */
1922             fix1             = _mm_add_pd(fix1,tx);
1923             fiy1             = _mm_add_pd(fiy1,ty);
1924             fiz1             = _mm_add_pd(fiz1,tz);
1925
1926             fjx0             = _mm_add_pd(fjx0,tx);
1927             fjy0             = _mm_add_pd(fjy0,ty);
1928             fjz0             = _mm_add_pd(fjz0,tz);
1929
1930             }
1931
1932             /**************************
1933              * CALCULATE INTERACTIONS *
1934              **************************/
1935
1936             if (gmx_mm_any_lt(rsq11,rcutoff2))
1937             {
1938
1939             /* REACTION-FIELD ELECTROSTATICS */
1940             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1941
1942             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1943
1944             fscal            = felec;
1945
1946             fscal            = _mm_and_pd(fscal,cutoff_mask);
1947
1948             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1949
1950             /* Calculate temporary vectorial force */
1951             tx               = _mm_mul_pd(fscal,dx11);
1952             ty               = _mm_mul_pd(fscal,dy11);
1953             tz               = _mm_mul_pd(fscal,dz11);
1954
1955             /* Update vectorial force */
1956             fix1             = _mm_add_pd(fix1,tx);
1957             fiy1             = _mm_add_pd(fiy1,ty);
1958             fiz1             = _mm_add_pd(fiz1,tz);
1959
1960             fjx1             = _mm_add_pd(fjx1,tx);
1961             fjy1             = _mm_add_pd(fjy1,ty);
1962             fjz1             = _mm_add_pd(fjz1,tz);
1963
1964             }
1965
1966             /**************************
1967              * CALCULATE INTERACTIONS *
1968              **************************/
1969
1970             if (gmx_mm_any_lt(rsq12,rcutoff2))
1971             {
1972
1973             /* REACTION-FIELD ELECTROSTATICS */
1974             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1975
1976             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1977
1978             fscal            = felec;
1979
1980             fscal            = _mm_and_pd(fscal,cutoff_mask);
1981
1982             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1983
1984             /* Calculate temporary vectorial force */
1985             tx               = _mm_mul_pd(fscal,dx12);
1986             ty               = _mm_mul_pd(fscal,dy12);
1987             tz               = _mm_mul_pd(fscal,dz12);
1988
1989             /* Update vectorial force */
1990             fix1             = _mm_add_pd(fix1,tx);
1991             fiy1             = _mm_add_pd(fiy1,ty);
1992             fiz1             = _mm_add_pd(fiz1,tz);
1993
1994             fjx2             = _mm_add_pd(fjx2,tx);
1995             fjy2             = _mm_add_pd(fjy2,ty);
1996             fjz2             = _mm_add_pd(fjz2,tz);
1997
1998             }
1999
2000             /**************************
2001              * CALCULATE INTERACTIONS *
2002              **************************/
2003
2004             if (gmx_mm_any_lt(rsq20,rcutoff2))
2005             {
2006
2007             /* REACTION-FIELD ELECTROSTATICS */
2008             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
2009
2010             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
2011
2012             fscal            = felec;
2013
2014             fscal            = _mm_and_pd(fscal,cutoff_mask);
2015
2016             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2017
2018             /* Calculate temporary vectorial force */
2019             tx               = _mm_mul_pd(fscal,dx20);
2020             ty               = _mm_mul_pd(fscal,dy20);
2021             tz               = _mm_mul_pd(fscal,dz20);
2022
2023             /* Update vectorial force */
2024             fix2             = _mm_add_pd(fix2,tx);
2025             fiy2             = _mm_add_pd(fiy2,ty);
2026             fiz2             = _mm_add_pd(fiz2,tz);
2027
2028             fjx0             = _mm_add_pd(fjx0,tx);
2029             fjy0             = _mm_add_pd(fjy0,ty);
2030             fjz0             = _mm_add_pd(fjz0,tz);
2031
2032             }
2033
2034             /**************************
2035              * CALCULATE INTERACTIONS *
2036              **************************/
2037
2038             if (gmx_mm_any_lt(rsq21,rcutoff2))
2039             {
2040
2041             /* REACTION-FIELD ELECTROSTATICS */
2042             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2043
2044             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2045
2046             fscal            = felec;
2047
2048             fscal            = _mm_and_pd(fscal,cutoff_mask);
2049
2050             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2051
2052             /* Calculate temporary vectorial force */
2053             tx               = _mm_mul_pd(fscal,dx21);
2054             ty               = _mm_mul_pd(fscal,dy21);
2055             tz               = _mm_mul_pd(fscal,dz21);
2056
2057             /* Update vectorial force */
2058             fix2             = _mm_add_pd(fix2,tx);
2059             fiy2             = _mm_add_pd(fiy2,ty);
2060             fiz2             = _mm_add_pd(fiz2,tz);
2061
2062             fjx1             = _mm_add_pd(fjx1,tx);
2063             fjy1             = _mm_add_pd(fjy1,ty);
2064             fjz1             = _mm_add_pd(fjz1,tz);
2065
2066             }
2067
2068             /**************************
2069              * CALCULATE INTERACTIONS *
2070              **************************/
2071
2072             if (gmx_mm_any_lt(rsq22,rcutoff2))
2073             {
2074
2075             /* REACTION-FIELD ELECTROSTATICS */
2076             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2077
2078             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2079
2080             fscal            = felec;
2081
2082             fscal            = _mm_and_pd(fscal,cutoff_mask);
2083
2084             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2085
2086             /* Calculate temporary vectorial force */
2087             tx               = _mm_mul_pd(fscal,dx22);
2088             ty               = _mm_mul_pd(fscal,dy22);
2089             tz               = _mm_mul_pd(fscal,dz22);
2090
2091             /* Update vectorial force */
2092             fix2             = _mm_add_pd(fix2,tx);
2093             fiy2             = _mm_add_pd(fiy2,ty);
2094             fiz2             = _mm_add_pd(fiz2,tz);
2095
2096             fjx2             = _mm_add_pd(fjx2,tx);
2097             fjy2             = _mm_add_pd(fjy2,ty);
2098             fjz2             = _mm_add_pd(fjz2,tz);
2099
2100             }
2101
2102             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2103
2104             /* Inner loop uses 301 flops */
2105         }
2106
2107         /* End of innermost loop */
2108
2109         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2110                                               f+i_coord_offset,fshift+i_shift_offset);
2111
2112         /* Increment number of inner iterations */
2113         inneriter                  += j_index_end - j_index_start;
2114
2115         /* Outer loop uses 18 flops */
2116     }
2117
2118     /* Increment number of outer iterations */
2119     outeriter        += nri;
2120
2121     /* Update outer/inner flops */
2122
2123     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*301);
2124 }