src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128d          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_macc_pd(krf,rsq00,rinv00),crf));
 296             felec            = _mm_mul_pd(qq00,_mm_msub_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_msub_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_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 310
 311             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 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_msub_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             /* Update vectorial force */
 330             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 331             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 332             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 333
 334             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 335             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 336             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 337
 338             }
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             if (gmx_mm_any_lt(rsq01,rcutoff2))
 345             {
 346
 347             /* REACTION-FIELD ELECTROSTATICS */
 348             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 349             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 350
 351             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velec            = _mm_and_pd(velec,cutoff_mask);
 355             velecsum         = _mm_add_pd(velecsum,velec);
 356
 357             fscal            = felec;
 358
 359             fscal            = _mm_and_pd(fscal,cutoff_mask);
 360
 361             /* Update vectorial force */
 362             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 363             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 364             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 365
 366             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 367             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 368             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 369
 370             }
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             if (gmx_mm_any_lt(rsq02,rcutoff2))
 377             {
 378
 379             /* REACTION-FIELD ELECTROSTATICS */
 380             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 381             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 382
 383             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velec            = _mm_and_pd(velec,cutoff_mask);
 387             velecsum         = _mm_add_pd(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391             fscal            = _mm_and_pd(fscal,cutoff_mask);
 392
 393             /* Update vectorial force */
 394             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 395             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 396             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 397
 398             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 399             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 400             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 401
 402             }
 403
 404             /**************************
 405              * CALCULATE INTERACTIONS *
 406              **************************/
 407
 408             if (gmx_mm_any_lt(rsq10,rcutoff2))
 409             {
 410
 411             /* REACTION-FIELD ELECTROSTATICS */
 412             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 413             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 414
 415             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velec            = _mm_and_pd(velec,cutoff_mask);
 419             velecsum         = _mm_add_pd(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             fscal            = _mm_and_pd(fscal,cutoff_mask);
 424
 425             /* Update vectorial force */
 426             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 427             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 428             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 429
 430             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 431             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 432             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 433
 434             }
 435
 436             /**************************
 437              * CALCULATE INTERACTIONS *
 438              **************************/
 439
 440             if (gmx_mm_any_lt(rsq11,rcutoff2))
 441             {
 442
 443             /* REACTION-FIELD ELECTROSTATICS */
 444             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 445             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 446
 447             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 448
 449             /* Update potential sum for this i atom from the interaction with this j atom. */
 450             velec            = _mm_and_pd(velec,cutoff_mask);
 451             velecsum         = _mm_add_pd(velecsum,velec);
 452
 453             fscal            = felec;
 454
 455             fscal            = _mm_and_pd(fscal,cutoff_mask);
 456
 457             /* Update vectorial force */
 458             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 459             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 460             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 461
 462             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 463             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 464             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 465
 466             }
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             if (gmx_mm_any_lt(rsq12,rcutoff2))
 473             {
 474
 475             /* REACTION-FIELD ELECTROSTATICS */
 476             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 477             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 478
 479             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 480
 481             /* Update potential sum for this i atom from the interaction with this j atom. */
 482             velec            = _mm_and_pd(velec,cutoff_mask);
 483             velecsum         = _mm_add_pd(velecsum,velec);
 484
 485             fscal            = felec;
 486
 487             fscal            = _mm_and_pd(fscal,cutoff_mask);
 488
 489             /* Update vectorial force */
 490             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 491             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 492             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 493
 494             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 495             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 496             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 497
 498             }
 499
 500             /**************************
 501              * CALCULATE INTERACTIONS *
 502              **************************/
 503
 504             if (gmx_mm_any_lt(rsq20,rcutoff2))
 505             {
 506
 507             /* REACTION-FIELD ELECTROSTATICS */
 508             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 509             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 510
 511             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 512
 513             /* Update potential sum for this i atom from the interaction with this j atom. */
 514             velec            = _mm_and_pd(velec,cutoff_mask);
 515             velecsum         = _mm_add_pd(velecsum,velec);
 516
 517             fscal            = felec;
 518
 519             fscal            = _mm_and_pd(fscal,cutoff_mask);
 520
 521             /* Update vectorial force */
 522             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 523             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 524             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 525
 526             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 527             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 528             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 529
 530             }
 531
 532             /**************************
 533              * CALCULATE INTERACTIONS *
 534              **************************/
 535
 536             if (gmx_mm_any_lt(rsq21,rcutoff2))
 537             {
 538
 539             /* REACTION-FIELD ELECTROSTATICS */
 540             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 541             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 542
 543             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 544
 545             /* Update potential sum for this i atom from the interaction with this j atom. */
 546             velec            = _mm_and_pd(velec,cutoff_mask);
 547             velecsum         = _mm_add_pd(velecsum,velec);
 548
 549             fscal            = felec;
 550
 551             fscal            = _mm_and_pd(fscal,cutoff_mask);
 552
 553             /* Update vectorial force */
 554             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 555             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 556             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 557
 558             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 559             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 560             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 561
 562             }
 563
 564             /**************************
 565              * CALCULATE INTERACTIONS *
 566              **************************/
 567
 568             if (gmx_mm_any_lt(rsq22,rcutoff2))
 569             {
 570
 571             /* REACTION-FIELD ELECTROSTATICS */
 572             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 573             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 574
 575             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 576
 577             /* Update potential sum for this i atom from the interaction with this j atom. */
 578             velec            = _mm_and_pd(velec,cutoff_mask);
 579             velecsum         = _mm_add_pd(velecsum,velec);
 580
 581             fscal            = felec;
 582
 583             fscal            = _mm_and_pd(fscal,cutoff_mask);
 584
 585             /* Update vectorial force */
 586             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 587             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 588             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 589
 590             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 591             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 592             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 593
 594             }
 595
 596             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 597
 598             /* Inner loop uses 385 flops */
 599         }
 600
 601         if(jidx<j_index_end)
 602         {
 603
 604             jnrA             = jjnr[jidx];
 605             j_coord_offsetA  = DIM*jnrA;
 606
 607             /* load j atom coordinates */
 608             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 609                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 610
 611             /* Calculate displacement vector */
 612             dx00             = _mm_sub_pd(ix0,jx0);
 613             dy00             = _mm_sub_pd(iy0,jy0);
 614             dz00             = _mm_sub_pd(iz0,jz0);
 615             dx01             = _mm_sub_pd(ix0,jx1);
 616             dy01             = _mm_sub_pd(iy0,jy1);
 617             dz01             = _mm_sub_pd(iz0,jz1);
 618             dx02             = _mm_sub_pd(ix0,jx2);
 619             dy02             = _mm_sub_pd(iy0,jy2);
 620             dz02             = _mm_sub_pd(iz0,jz2);
 621             dx10             = _mm_sub_pd(ix1,jx0);
 622             dy10             = _mm_sub_pd(iy1,jy0);
 623             dz10             = _mm_sub_pd(iz1,jz0);
 624             dx11             = _mm_sub_pd(ix1,jx1);
 625             dy11             = _mm_sub_pd(iy1,jy1);
 626             dz11             = _mm_sub_pd(iz1,jz1);
 627             dx12             = _mm_sub_pd(ix1,jx2);
 628             dy12             = _mm_sub_pd(iy1,jy2);
 629             dz12             = _mm_sub_pd(iz1,jz2);
 630             dx20             = _mm_sub_pd(ix2,jx0);
 631             dy20             = _mm_sub_pd(iy2,jy0);
 632             dz20             = _mm_sub_pd(iz2,jz0);
 633             dx21             = _mm_sub_pd(ix2,jx1);
 634             dy21             = _mm_sub_pd(iy2,jy1);
 635             dz21             = _mm_sub_pd(iz2,jz1);
 636             dx22             = _mm_sub_pd(ix2,jx2);
 637             dy22             = _mm_sub_pd(iy2,jy2);
 638             dz22             = _mm_sub_pd(iz2,jz2);
 639
 640             /* Calculate squared distance and things based on it */
 641             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 642             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 643             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 644             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 645             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 646             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 647             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 648             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 649             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 650
 651             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 652             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 653             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 654             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 655             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 656             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 657             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 658             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 659             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 660
 661             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 662             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 663             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 664             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 665             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 666             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 667             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 668             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 669             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 670
 671             fjx0             = _mm_setzero_pd();
 672             fjy0             = _mm_setzero_pd();
 673             fjz0             = _mm_setzero_pd();
 674             fjx1             = _mm_setzero_pd();
 675             fjy1             = _mm_setzero_pd();
 676             fjz1             = _mm_setzero_pd();
 677             fjx2             = _mm_setzero_pd();
 678             fjy2             = _mm_setzero_pd();
 679             fjz2             = _mm_setzero_pd();
 680
 681             /**************************
 682              * CALCULATE INTERACTIONS *
 683              **************************/
 684
 685             if (gmx_mm_any_lt(rsq00,rcutoff2))
 686             {
 687
 688             r00              = _mm_mul_pd(rsq00,rinv00);
 689
 690             /* REACTION-FIELD ELECTROSTATICS */
 691             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 692             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 693
 694             /* LENNARD-JONES DISPERSION/REPULSION */
 695
 696             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 697             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 698             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 699             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 700             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 701
 702             d                = _mm_sub_pd(r00,rswitch);
 703             d                = _mm_max_pd(d,_mm_setzero_pd());
 704             d2               = _mm_mul_pd(d,d);
 705             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 706
 707             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 708
 709             /* Evaluate switch function */
 710             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 711             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 712             vvdw             = _mm_mul_pd(vvdw,sw);
 713             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 714
 715             /* Update potential sum for this i atom from the interaction with this j atom. */
 716             velec            = _mm_and_pd(velec,cutoff_mask);
 717             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 718             velecsum         = _mm_add_pd(velecsum,velec);
 719             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 720             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 721             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 722
 723             fscal            = _mm_add_pd(felec,fvdw);
 724
 725             fscal            = _mm_and_pd(fscal,cutoff_mask);
 726
 727             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 728
 729             /* Update vectorial force */
 730             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 731             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 732             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 733
 734             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 735             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 736             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 737
 738             }
 739
 740             /**************************
 741              * CALCULATE INTERACTIONS *
 742              **************************/
 743
 744             if (gmx_mm_any_lt(rsq01,rcutoff2))
 745             {
 746
 747             /* REACTION-FIELD ELECTROSTATICS */
 748             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 749             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 750
 751             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 752
 753             /* Update potential sum for this i atom from the interaction with this j atom. */
 754             velec            = _mm_and_pd(velec,cutoff_mask);
 755             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 756             velecsum         = _mm_add_pd(velecsum,velec);
 757
 758             fscal            = felec;
 759
 760             fscal            = _mm_and_pd(fscal,cutoff_mask);
 761
 762             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 763
 764             /* Update vectorial force */
 765             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 766             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 767             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 768
 769             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 770             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 771             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 772
 773             }
 774
 775             /**************************
 776              * CALCULATE INTERACTIONS *
 777              **************************/
 778
 779             if (gmx_mm_any_lt(rsq02,rcutoff2))
 780             {
 781
 782             /* REACTION-FIELD ELECTROSTATICS */
 783             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 784             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 785
 786             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 787
 788             /* Update potential sum for this i atom from the interaction with this j atom. */
 789             velec            = _mm_and_pd(velec,cutoff_mask);
 790             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 791             velecsum         = _mm_add_pd(velecsum,velec);
 792
 793             fscal            = felec;
 794
 795             fscal            = _mm_and_pd(fscal,cutoff_mask);
 796
 797             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 798
 799             /* Update vectorial force */
 800             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 801             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 802             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 803
 804             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 805             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 806             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 807
 808             }
 809
 810             /**************************
 811              * CALCULATE INTERACTIONS *
 812              **************************/
 813
 814             if (gmx_mm_any_lt(rsq10,rcutoff2))
 815             {
 816
 817             /* REACTION-FIELD ELECTROSTATICS */
 818             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 819             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 820
 821             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 822
 823             /* Update potential sum for this i atom from the interaction with this j atom. */
 824             velec            = _mm_and_pd(velec,cutoff_mask);
 825             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 826             velecsum         = _mm_add_pd(velecsum,velec);
 827
 828             fscal            = felec;
 829
 830             fscal            = _mm_and_pd(fscal,cutoff_mask);
 831
 832             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 833
 834             /* Update vectorial force */
 835             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 836             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 837             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 838
 839             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 840             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 841             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 842
 843             }
 844
 845             /**************************
 846              * CALCULATE INTERACTIONS *
 847              **************************/
 848
 849             if (gmx_mm_any_lt(rsq11,rcutoff2))
 850             {
 851
 852             /* REACTION-FIELD ELECTROSTATICS */
 853             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 854             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 855
 856             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 857
 858             /* Update potential sum for this i atom from the interaction with this j atom. */
 859             velec            = _mm_and_pd(velec,cutoff_mask);
 860             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 861             velecsum         = _mm_add_pd(velecsum,velec);
 862
 863             fscal            = felec;
 864
 865             fscal            = _mm_and_pd(fscal,cutoff_mask);
 866
 867             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 868
 869             /* Update vectorial force */
 870             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 871             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 872             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 873
 874             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 875             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 876             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 877
 878             }
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             if (gmx_mm_any_lt(rsq12,rcutoff2))
 885             {
 886
 887             /* REACTION-FIELD ELECTROSTATICS */
 888             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 889             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 890
 891             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 892
 893             /* Update potential sum for this i atom from the interaction with this j atom. */
 894             velec            = _mm_and_pd(velec,cutoff_mask);
 895             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 896             velecsum         = _mm_add_pd(velecsum,velec);
 897
 898             fscal            = felec;
 899
 900             fscal            = _mm_and_pd(fscal,cutoff_mask);
 901
 902             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 903
 904             /* Update vectorial force */
 905             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 906             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 907             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 908
 909             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 910             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 911             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 912
 913             }
 914
 915             /**************************
 916              * CALCULATE INTERACTIONS *
 917              **************************/
 918
 919             if (gmx_mm_any_lt(rsq20,rcutoff2))
 920             {
 921
 922             /* REACTION-FIELD ELECTROSTATICS */
 923             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 924             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 925
 926             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 927
 928             /* Update potential sum for this i atom from the interaction with this j atom. */
 929             velec            = _mm_and_pd(velec,cutoff_mask);
 930             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 931             velecsum         = _mm_add_pd(velecsum,velec);
 932
 933             fscal            = felec;
 934
 935             fscal            = _mm_and_pd(fscal,cutoff_mask);
 936
 937             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 938
 939             /* Update vectorial force */
 940             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 941             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 942             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 943
 944             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 945             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 946             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 947
 948             }
 949
 950             /**************************
 951              * CALCULATE INTERACTIONS *
 952              **************************/
 953
 954             if (gmx_mm_any_lt(rsq21,rcutoff2))
 955             {
 956
 957             /* REACTION-FIELD ELECTROSTATICS */
 958             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 959             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 960
 961             cutoff_mask      = _mm_cmplt_pd(rsq21,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             /* Update vectorial force */
 975             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 976             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 977             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 978
 979             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 980             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 981             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 982
 983             }
 984
 985             /**************************
 986              * CALCULATE INTERACTIONS *
 987              **************************/
 988
 989             if (gmx_mm_any_lt(rsq22,rcutoff2))
 990             {
 991
 992             /* REACTION-FIELD ELECTROSTATICS */
 993             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 994             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 995
 996             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 997
 998             /* Update potential sum for this i atom from the interaction with this j atom. */
 999             velec            = _mm_and_pd(velec,cutoff_mask);
1000             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1001             velecsum         = _mm_add_pd(velecsum,velec);
1002
1003             fscal            = felec;
1004
1005             fscal            = _mm_and_pd(fscal,cutoff_mask);
1006
1007             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1008
1009             /* Update vectorial force */
1010             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1011             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1012             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1013
1014             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1015             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1016             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1017
1018             }
1019
1020             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1021
1022             /* Inner loop uses 385 flops */
1023         }
1024
1025         /* End of innermost loop */
1026
1027         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1028                                               f+i_coord_offset,fshift+i_shift_offset);
1029
1030         ggid                        = gid[iidx];
1031         /* Update potential energies */
1032         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1033         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1034
1035         /* Increment number of inner iterations */
1036         inneriter                  += j_index_end - j_index_start;
1037
1038         /* Outer loop uses 20 flops */
1039     }
1040
1041     /* Increment number of outer iterations */
1042     outeriter        += nri;
1043
1044     /* Update outer/inner flops */
1045
1046     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*385);
1047 }
1048 /*
1049  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_double
1050  * Electrostatics interaction: ReactionField
1051  * VdW interaction:            LennardJones
1052  * Geometry:                   Water3-Water3
1053  * Calculate force/pot:        Force
1054  */
1055 void
1056 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_double
1057                     (t_nblist * gmx_restrict                nlist,
1058                      rvec * gmx_restrict                    xx,
1059                      rvec * gmx_restrict                    ff,
1060                      t_forcerec * gmx_restrict              fr,
1061                      t_mdatoms * gmx_restrict               mdatoms,
1062                      nb_kernel_data_t * gmx_restrict        kernel_data,
1063                      t_nrnb * gmx_restrict                  nrnb)
1064 {
1065     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1066      * just 0 for non-waters.
1067      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1068      * jnr indices corresponding to data put in the four positions in the SIMD register.
1069      */
1070     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1071     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1072     int              jnrA,jnrB;
1073     int              j_coord_offsetA,j_coord_offsetB;
1074     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1075     real             rcutoff_scalar;
1076     real             *shiftvec,*fshift,*x,*f;
1077     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1078     int              vdwioffset0;
1079     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1080     int              vdwioffset1;
1081     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1082     int              vdwioffset2;
1083     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1084     int              vdwjidx0A,vdwjidx0B;
1085     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1086     int              vdwjidx1A,vdwjidx1B;
1087     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1088     int              vdwjidx2A,vdwjidx2B;
1089     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1090     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1091     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1092     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1093     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1094     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1095     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1096     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1097     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1098     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1099     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1100     real             *charge;
1101     int              nvdwtype;
1102     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1103     int              *vdwtype;
1104     real             *vdwparam;
1105     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1106     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1107     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1108     real             rswitch_scalar,d_scalar;
1109     __m128d          dummy_mask,cutoff_mask;
1110     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1111     __m128d          one     = _mm_set1_pd(1.0);
1112     __m128d          two     = _mm_set1_pd(2.0);
1113     x                = xx[0];
1114     f                = ff[0];
1115
1116     nri              = nlist->nri;
1117     iinr             = nlist->iinr;
1118     jindex           = nlist->jindex;
1119     jjnr             = nlist->jjnr;
1120     shiftidx         = nlist->shift;
1121     gid              = nlist->gid;
1122     shiftvec         = fr->shift_vec[0];
1123     fshift           = fr->fshift[0];
1124     facel            = _mm_set1_pd(fr->epsfac);
1125     charge           = mdatoms->chargeA;
1126     krf              = _mm_set1_pd(fr->ic->k_rf);
1127     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1128     crf              = _mm_set1_pd(fr->ic->c_rf);
1129     nvdwtype         = fr->ntype;
1130     vdwparam         = fr->nbfp;
1131     vdwtype          = mdatoms->typeA;
1132
1133     /* Setup water-specific parameters */
1134     inr              = nlist->iinr[0];
1135     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1136     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1137     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1138     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1139
1140     jq0              = _mm_set1_pd(charge[inr+0]);
1141     jq1              = _mm_set1_pd(charge[inr+1]);
1142     jq2              = _mm_set1_pd(charge[inr+2]);
1143     vdwjidx0A        = 2*vdwtype[inr+0];
1144     qq00             = _mm_mul_pd(iq0,jq0);
1145     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1146     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1147     qq01             = _mm_mul_pd(iq0,jq1);
1148     qq02             = _mm_mul_pd(iq0,jq2);
1149     qq10             = _mm_mul_pd(iq1,jq0);
1150     qq11             = _mm_mul_pd(iq1,jq1);
1151     qq12             = _mm_mul_pd(iq1,jq2);
1152     qq20             = _mm_mul_pd(iq2,jq0);
1153     qq21             = _mm_mul_pd(iq2,jq1);
1154     qq22             = _mm_mul_pd(iq2,jq2);
1155
1156     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1157     rcutoff_scalar   = fr->rcoulomb;
1158     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1159     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1160
1161     rswitch_scalar   = fr->rvdw_switch;
1162     rswitch          = _mm_set1_pd(rswitch_scalar);
1163     /* Setup switch parameters */
1164     d_scalar         = rcutoff_scalar-rswitch_scalar;
1165     d                = _mm_set1_pd(d_scalar);
1166     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1167     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1168     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1169     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1170     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1171     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1172
1173     /* Avoid stupid compiler warnings */
1174     jnrA = jnrB = 0;
1175     j_coord_offsetA = 0;
1176     j_coord_offsetB = 0;
1177
1178     outeriter        = 0;
1179     inneriter        = 0;
1180
1181     /* Start outer loop over neighborlists */
1182     for(iidx=0; iidx<nri; iidx++)
1183     {
1184         /* Load shift vector for this list */
1185         i_shift_offset   = DIM*shiftidx[iidx];
1186
1187         /* Load limits for loop over neighbors */
1188         j_index_start    = jindex[iidx];
1189         j_index_end      = jindex[iidx+1];
1190
1191         /* Get outer coordinate index */
1192         inr              = iinr[iidx];
1193         i_coord_offset   = DIM*inr;
1194
1195         /* Load i particle coords and add shift vector */
1196         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1197                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1198
1199         fix0             = _mm_setzero_pd();
1200         fiy0             = _mm_setzero_pd();
1201         fiz0             = _mm_setzero_pd();
1202         fix1             = _mm_setzero_pd();
1203         fiy1             = _mm_setzero_pd();
1204         fiz1             = _mm_setzero_pd();
1205         fix2             = _mm_setzero_pd();
1206         fiy2             = _mm_setzero_pd();
1207         fiz2             = _mm_setzero_pd();
1208
1209         /* Start inner kernel loop */
1210         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1211         {
1212
1213             /* Get j neighbor index, and coordinate index */
1214             jnrA             = jjnr[jidx];
1215             jnrB             = jjnr[jidx+1];
1216             j_coord_offsetA  = DIM*jnrA;
1217             j_coord_offsetB  = DIM*jnrB;
1218
1219             /* load j atom coordinates */
1220             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1221                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1222
1223             /* Calculate displacement vector */
1224             dx00             = _mm_sub_pd(ix0,jx0);
1225             dy00             = _mm_sub_pd(iy0,jy0);
1226             dz00             = _mm_sub_pd(iz0,jz0);
1227             dx01             = _mm_sub_pd(ix0,jx1);
1228             dy01             = _mm_sub_pd(iy0,jy1);
1229             dz01             = _mm_sub_pd(iz0,jz1);
1230             dx02             = _mm_sub_pd(ix0,jx2);
1231             dy02             = _mm_sub_pd(iy0,jy2);
1232             dz02             = _mm_sub_pd(iz0,jz2);
1233             dx10             = _mm_sub_pd(ix1,jx0);
1234             dy10             = _mm_sub_pd(iy1,jy0);
1235             dz10             = _mm_sub_pd(iz1,jz0);
1236             dx11             = _mm_sub_pd(ix1,jx1);
1237             dy11             = _mm_sub_pd(iy1,jy1);
1238             dz11             = _mm_sub_pd(iz1,jz1);
1239             dx12             = _mm_sub_pd(ix1,jx2);
1240             dy12             = _mm_sub_pd(iy1,jy2);
1241             dz12             = _mm_sub_pd(iz1,jz2);
1242             dx20             = _mm_sub_pd(ix2,jx0);
1243             dy20             = _mm_sub_pd(iy2,jy0);
1244             dz20             = _mm_sub_pd(iz2,jz0);
1245             dx21             = _mm_sub_pd(ix2,jx1);
1246             dy21             = _mm_sub_pd(iy2,jy1);
1247             dz21             = _mm_sub_pd(iz2,jz1);
1248             dx22             = _mm_sub_pd(ix2,jx2);
1249             dy22             = _mm_sub_pd(iy2,jy2);
1250             dz22             = _mm_sub_pd(iz2,jz2);
1251
1252             /* Calculate squared distance and things based on it */
1253             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1254             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1255             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1256             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1257             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1258             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1259             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1260             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1261             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1262
1263             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1264             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1265             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1266             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1267             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1268             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1269             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1270             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1271             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1272
1273             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1274             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1275             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1276             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1277             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1278             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1279             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1280             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1281             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1282
1283             fjx0             = _mm_setzero_pd();
1284             fjy0             = _mm_setzero_pd();
1285             fjz0             = _mm_setzero_pd();
1286             fjx1             = _mm_setzero_pd();
1287             fjy1             = _mm_setzero_pd();
1288             fjz1             = _mm_setzero_pd();
1289             fjx2             = _mm_setzero_pd();
1290             fjy2             = _mm_setzero_pd();
1291             fjz2             = _mm_setzero_pd();
1292
1293             /**************************
1294              * CALCULATE INTERACTIONS *
1295              **************************/
1296
1297             if (gmx_mm_any_lt(rsq00,rcutoff2))
1298             {
1299
1300             r00              = _mm_mul_pd(rsq00,rinv00);
1301
1302             /* REACTION-FIELD ELECTROSTATICS */
1303             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1304
1305             /* LENNARD-JONES DISPERSION/REPULSION */
1306
1307             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1308             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1309             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1310             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1311             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1312
1313             d                = _mm_sub_pd(r00,rswitch);
1314             d                = _mm_max_pd(d,_mm_setzero_pd());
1315             d2               = _mm_mul_pd(d,d);
1316             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1317
1318             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1319
1320             /* Evaluate switch function */
1321             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1322             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1323             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1324
1325             fscal            = _mm_add_pd(felec,fvdw);
1326
1327             fscal            = _mm_and_pd(fscal,cutoff_mask);
1328
1329             /* Update vectorial force */
1330             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1331             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1332             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1333
1334             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1335             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1336             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1337
1338             }
1339
1340             /**************************
1341              * CALCULATE INTERACTIONS *
1342              **************************/
1343
1344             if (gmx_mm_any_lt(rsq01,rcutoff2))
1345             {
1346
1347             /* REACTION-FIELD ELECTROSTATICS */
1348             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1349
1350             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1351
1352             fscal            = felec;
1353
1354             fscal            = _mm_and_pd(fscal,cutoff_mask);
1355
1356             /* Update vectorial force */
1357             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1358             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1359             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1360
1361             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1362             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1363             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1364
1365             }
1366
1367             /**************************
1368              * CALCULATE INTERACTIONS *
1369              **************************/
1370
1371             if (gmx_mm_any_lt(rsq02,rcutoff2))
1372             {
1373
1374             /* REACTION-FIELD ELECTROSTATICS */
1375             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1376
1377             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1378
1379             fscal            = felec;
1380
1381             fscal            = _mm_and_pd(fscal,cutoff_mask);
1382
1383             /* Update vectorial force */
1384             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1385             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1386             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1387
1388             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1389             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1390             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1391
1392             }
1393
1394             /**************************
1395              * CALCULATE INTERACTIONS *
1396              **************************/
1397
1398             if (gmx_mm_any_lt(rsq10,rcutoff2))
1399             {
1400
1401             /* REACTION-FIELD ELECTROSTATICS */
1402             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1403
1404             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1405
1406             fscal            = felec;
1407
1408             fscal            = _mm_and_pd(fscal,cutoff_mask);
1409
1410             /* Update vectorial force */
1411             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1412             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1413             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1414
1415             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1416             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1417             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1418
1419             }
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             if (gmx_mm_any_lt(rsq11,rcutoff2))
1426             {
1427
1428             /* REACTION-FIELD ELECTROSTATICS */
1429             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1430
1431             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1432
1433             fscal            = felec;
1434
1435             fscal            = _mm_and_pd(fscal,cutoff_mask);
1436
1437             /* Update vectorial force */
1438             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1439             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1440             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1441
1442             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1443             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1444             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1445
1446             }
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             if (gmx_mm_any_lt(rsq12,rcutoff2))
1453             {
1454
1455             /* REACTION-FIELD ELECTROSTATICS */
1456             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1457
1458             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1459
1460             fscal            = felec;
1461
1462             fscal            = _mm_and_pd(fscal,cutoff_mask);
1463
1464             /* Update vectorial force */
1465             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1466             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1467             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1468
1469             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1470             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1471             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1472
1473             }
1474
1475             /**************************
1476              * CALCULATE INTERACTIONS *
1477              **************************/
1478
1479             if (gmx_mm_any_lt(rsq20,rcutoff2))
1480             {
1481
1482             /* REACTION-FIELD ELECTROSTATICS */
1483             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1484
1485             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1486
1487             fscal            = felec;
1488
1489             fscal            = _mm_and_pd(fscal,cutoff_mask);
1490
1491             /* Update vectorial force */
1492             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1493             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1494             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1495
1496             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1497             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1498             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1499
1500             }
1501
1502             /**************************
1503              * CALCULATE INTERACTIONS *
1504              **************************/
1505
1506             if (gmx_mm_any_lt(rsq21,rcutoff2))
1507             {
1508
1509             /* REACTION-FIELD ELECTROSTATICS */
1510             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1511
1512             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1513
1514             fscal            = felec;
1515
1516             fscal            = _mm_and_pd(fscal,cutoff_mask);
1517
1518             /* Update vectorial force */
1519             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1520             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1521             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1522
1523             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1524             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1525             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1526
1527             }
1528
1529             /**************************
1530              * CALCULATE INTERACTIONS *
1531              **************************/
1532
1533             if (gmx_mm_any_lt(rsq22,rcutoff2))
1534             {
1535
1536             /* REACTION-FIELD ELECTROSTATICS */
1537             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1538
1539             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1540
1541             fscal            = felec;
1542
1543             fscal            = _mm_and_pd(fscal,cutoff_mask);
1544
1545             /* Update vectorial force */
1546             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1547             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1548             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1549
1550             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1551             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1552             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1553
1554             }
1555
1556             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1557
1558             /* Inner loop uses 328 flops */
1559         }
1560
1561         if(jidx<j_index_end)
1562         {
1563
1564             jnrA             = jjnr[jidx];
1565             j_coord_offsetA  = DIM*jnrA;
1566
1567             /* load j atom coordinates */
1568             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1569                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1570
1571             /* Calculate displacement vector */
1572             dx00             = _mm_sub_pd(ix0,jx0);
1573             dy00             = _mm_sub_pd(iy0,jy0);
1574             dz00             = _mm_sub_pd(iz0,jz0);
1575             dx01             = _mm_sub_pd(ix0,jx1);
1576             dy01             = _mm_sub_pd(iy0,jy1);
1577             dz01             = _mm_sub_pd(iz0,jz1);
1578             dx02             = _mm_sub_pd(ix0,jx2);
1579             dy02             = _mm_sub_pd(iy0,jy2);
1580             dz02             = _mm_sub_pd(iz0,jz2);
1581             dx10             = _mm_sub_pd(ix1,jx0);
1582             dy10             = _mm_sub_pd(iy1,jy0);
1583             dz10             = _mm_sub_pd(iz1,jz0);
1584             dx11             = _mm_sub_pd(ix1,jx1);
1585             dy11             = _mm_sub_pd(iy1,jy1);
1586             dz11             = _mm_sub_pd(iz1,jz1);
1587             dx12             = _mm_sub_pd(ix1,jx2);
1588             dy12             = _mm_sub_pd(iy1,jy2);
1589             dz12             = _mm_sub_pd(iz1,jz2);
1590             dx20             = _mm_sub_pd(ix2,jx0);
1591             dy20             = _mm_sub_pd(iy2,jy0);
1592             dz20             = _mm_sub_pd(iz2,jz0);
1593             dx21             = _mm_sub_pd(ix2,jx1);
1594             dy21             = _mm_sub_pd(iy2,jy1);
1595             dz21             = _mm_sub_pd(iz2,jz1);
1596             dx22             = _mm_sub_pd(ix2,jx2);
1597             dy22             = _mm_sub_pd(iy2,jy2);
1598             dz22             = _mm_sub_pd(iz2,jz2);
1599
1600             /* Calculate squared distance and things based on it */
1601             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1602             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1603             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1604             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1605             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1606             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1607             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1608             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1609             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1610
1611             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1612             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1613             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1614             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1615             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1616             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1617             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1618             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1619             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1620
1621             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1622             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1623             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1624             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1625             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1626             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1627             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1628             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1629             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1630
1631             fjx0             = _mm_setzero_pd();
1632             fjy0             = _mm_setzero_pd();
1633             fjz0             = _mm_setzero_pd();
1634             fjx1             = _mm_setzero_pd();
1635             fjy1             = _mm_setzero_pd();
1636             fjz1             = _mm_setzero_pd();
1637             fjx2             = _mm_setzero_pd();
1638             fjy2             = _mm_setzero_pd();
1639             fjz2             = _mm_setzero_pd();
1640
1641             /**************************
1642              * CALCULATE INTERACTIONS *
1643              **************************/
1644
1645             if (gmx_mm_any_lt(rsq00,rcutoff2))
1646             {
1647
1648             r00              = _mm_mul_pd(rsq00,rinv00);
1649
1650             /* REACTION-FIELD ELECTROSTATICS */
1651             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1652
1653             /* LENNARD-JONES DISPERSION/REPULSION */
1654
1655             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1656             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1657             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1658             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1659             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1660
1661             d                = _mm_sub_pd(r00,rswitch);
1662             d                = _mm_max_pd(d,_mm_setzero_pd());
1663             d2               = _mm_mul_pd(d,d);
1664             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1665
1666             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1667
1668             /* Evaluate switch function */
1669             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1670             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1671             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1672
1673             fscal            = _mm_add_pd(felec,fvdw);
1674
1675             fscal            = _mm_and_pd(fscal,cutoff_mask);
1676
1677             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1678
1679             /* Update vectorial force */
1680             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1681             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1682             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1683
1684             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1685             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1686             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1687
1688             }
1689
1690             /**************************
1691              * CALCULATE INTERACTIONS *
1692              **************************/
1693
1694             if (gmx_mm_any_lt(rsq01,rcutoff2))
1695             {
1696
1697             /* REACTION-FIELD ELECTROSTATICS */
1698             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1699
1700             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1701
1702             fscal            = felec;
1703
1704             fscal            = _mm_and_pd(fscal,cutoff_mask);
1705
1706             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1707
1708             /* Update vectorial force */
1709             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1710             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1711             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1712
1713             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1714             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1715             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1716
1717             }
1718
1719             /**************************
1720              * CALCULATE INTERACTIONS *
1721              **************************/
1722
1723             if (gmx_mm_any_lt(rsq02,rcutoff2))
1724             {
1725
1726             /* REACTION-FIELD ELECTROSTATICS */
1727             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1728
1729             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1730
1731             fscal            = felec;
1732
1733             fscal            = _mm_and_pd(fscal,cutoff_mask);
1734
1735             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1736
1737             /* Update vectorial force */
1738             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1739             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1740             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1741
1742             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1743             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1744             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1745
1746             }
1747
1748             /**************************
1749              * CALCULATE INTERACTIONS *
1750              **************************/
1751
1752             if (gmx_mm_any_lt(rsq10,rcutoff2))
1753             {
1754
1755             /* REACTION-FIELD ELECTROSTATICS */
1756             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1757
1758             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1759
1760             fscal            = felec;
1761
1762             fscal            = _mm_and_pd(fscal,cutoff_mask);
1763
1764             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1765
1766             /* Update vectorial force */
1767             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1768             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1769             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1770
1771             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1772             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1773             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1774
1775             }
1776
1777             /**************************
1778              * CALCULATE INTERACTIONS *
1779              **************************/
1780
1781             if (gmx_mm_any_lt(rsq11,rcutoff2))
1782             {
1783
1784             /* REACTION-FIELD ELECTROSTATICS */
1785             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1786
1787             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1788
1789             fscal            = felec;
1790
1791             fscal            = _mm_and_pd(fscal,cutoff_mask);
1792
1793             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1794
1795             /* Update vectorial force */
1796             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1797             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1798             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1799
1800             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1801             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1802             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1803
1804             }
1805
1806             /**************************
1807              * CALCULATE INTERACTIONS *
1808              **************************/
1809
1810             if (gmx_mm_any_lt(rsq12,rcutoff2))
1811             {
1812
1813             /* REACTION-FIELD ELECTROSTATICS */
1814             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1815
1816             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1817
1818             fscal            = felec;
1819
1820             fscal            = _mm_and_pd(fscal,cutoff_mask);
1821
1822             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1823
1824             /* Update vectorial force */
1825             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1826             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1827             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1828
1829             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1830             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1831             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1832
1833             }
1834
1835             /**************************
1836              * CALCULATE INTERACTIONS *
1837              **************************/
1838
1839             if (gmx_mm_any_lt(rsq20,rcutoff2))
1840             {
1841
1842             /* REACTION-FIELD ELECTROSTATICS */
1843             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1844
1845             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1846
1847             fscal            = felec;
1848
1849             fscal            = _mm_and_pd(fscal,cutoff_mask);
1850
1851             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1852
1853             /* Update vectorial force */
1854             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1855             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1856             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1857
1858             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1859             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1860             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1861
1862             }
1863
1864             /**************************
1865              * CALCULATE INTERACTIONS *
1866              **************************/
1867
1868             if (gmx_mm_any_lt(rsq21,rcutoff2))
1869             {
1870
1871             /* REACTION-FIELD ELECTROSTATICS */
1872             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1873
1874             cutoff_mask      = _mm_cmplt_pd(rsq21,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             /* Update vectorial force */
1883             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1884             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1885             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1886
1887             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1888             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1889             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1890
1891             }
1892
1893             /**************************
1894              * CALCULATE INTERACTIONS *
1895              **************************/
1896
1897             if (gmx_mm_any_lt(rsq22,rcutoff2))
1898             {
1899
1900             /* REACTION-FIELD ELECTROSTATICS */
1901             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1902
1903             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1904
1905             fscal            = felec;
1906
1907             fscal            = _mm_and_pd(fscal,cutoff_mask);
1908
1909             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1910
1911             /* Update vectorial force */
1912             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1913             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1914             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1915
1916             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1917             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1918             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1919
1920             }
1921
1922             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1923
1924             /* Inner loop uses 328 flops */
1925         }
1926
1927         /* End of innermost loop */
1928
1929         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1930                                               f+i_coord_offset,fshift+i_shift_offset);
1931
1932         /* Increment number of inner iterations */
1933         inneriter                  += j_index_end - j_index_start;
1934
1935         /* Outer loop uses 18 flops */
1936     }
1937
1938     /* Increment number of outer iterations */
1939     outeriter        += nri;
1940
1941     /* Update outer/inner flops */
1942
1943     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*328);
1944 }