src/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_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_GeomW4W4_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_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              vdwioffset3;
  73     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B;
  75     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B;
  77     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B;
  79     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B;
  81     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  99     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 100     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 101     real             rswitch_scalar,d_scalar;
 102     __m128d          dummy_mask,cutoff_mask;
 103     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 104     __m128d          one     = _mm_set1_pd(1.0);
 105     __m128d          two     = _mm_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_pd(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = _mm_set1_pd(fr->ic->k_rf);
 120     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 121     crf              = _mm_set1_pd(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 129     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 130     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     jq1              = _mm_set1_pd(charge[inr+1]);
 134     jq2              = _mm_set1_pd(charge[inr+2]);
 135     jq3              = _mm_set1_pd(charge[inr+3]);
 136     vdwjidx0A        = 2*vdwtype[inr+0];
 137     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 138     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 139     qq11             = _mm_mul_pd(iq1,jq1);
 140     qq12             = _mm_mul_pd(iq1,jq2);
 141     qq13             = _mm_mul_pd(iq1,jq3);
 142     qq21             = _mm_mul_pd(iq2,jq1);
 143     qq22             = _mm_mul_pd(iq2,jq2);
 144     qq23             = _mm_mul_pd(iq2,jq3);
 145     qq31             = _mm_mul_pd(iq3,jq1);
 146     qq32             = _mm_mul_pd(iq3,jq2);
 147     qq33             = _mm_mul_pd(iq3,jq3);
 148
 149     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 150     rcutoff_scalar   = fr->rcoulomb;
 151     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 152     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 153
 154     rswitch_scalar   = fr->rvdw_switch;
 155     rswitch          = _mm_set1_pd(rswitch_scalar);
 156     /* Setup switch parameters */
 157     d_scalar         = rcutoff_scalar-rswitch_scalar;
 158     d                = _mm_set1_pd(d_scalar);
 159     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 160     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 161     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 162     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 163     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 164     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 165
 166     /* Avoid stupid compiler warnings */
 167     jnrA = jnrB = 0;
 168     j_coord_offsetA = 0;
 169     j_coord_offsetB = 0;
 170
 171     outeriter        = 0;
 172     inneriter        = 0;
 173
 174     /* Start outer loop over neighborlists */
 175     for(iidx=0; iidx<nri; iidx++)
 176     {
 177         /* Load shift vector for this list */
 178         i_shift_offset   = DIM*shiftidx[iidx];
 179
 180         /* Load limits for loop over neighbors */
 181         j_index_start    = jindex[iidx];
 182         j_index_end      = jindex[iidx+1];
 183
 184         /* Get outer coordinate index */
 185         inr              = iinr[iidx];
 186         i_coord_offset   = DIM*inr;
 187
 188         /* Load i particle coords and add shift vector */
 189         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 190                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 191
 192         fix0             = _mm_setzero_pd();
 193         fiy0             = _mm_setzero_pd();
 194         fiz0             = _mm_setzero_pd();
 195         fix1             = _mm_setzero_pd();
 196         fiy1             = _mm_setzero_pd();
 197         fiz1             = _mm_setzero_pd();
 198         fix2             = _mm_setzero_pd();
 199         fiy2             = _mm_setzero_pd();
 200         fiz2             = _mm_setzero_pd();
 201         fix3             = _mm_setzero_pd();
 202         fiy3             = _mm_setzero_pd();
 203         fiz3             = _mm_setzero_pd();
 204
 205         /* Reset potential sums */
 206         velecsum         = _mm_setzero_pd();
 207         vvdwsum          = _mm_setzero_pd();
 208
 209         /* Start inner kernel loop */
 210         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 211         {
 212
 213             /* Get j neighbor index, and coordinate index */
 214             jnrA             = jjnr[jidx];
 215             jnrB             = jjnr[jidx+1];
 216             j_coord_offsetA  = DIM*jnrA;
 217             j_coord_offsetB  = DIM*jnrB;
 218
 219             /* load j atom coordinates */
 220             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 221                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 222                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 223
 224             /* Calculate displacement vector */
 225             dx00             = _mm_sub_pd(ix0,jx0);
 226             dy00             = _mm_sub_pd(iy0,jy0);
 227             dz00             = _mm_sub_pd(iz0,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             dx13             = _mm_sub_pd(ix1,jx3);
 235             dy13             = _mm_sub_pd(iy1,jy3);
 236             dz13             = _mm_sub_pd(iz1,jz3);
 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             dx23             = _mm_sub_pd(ix2,jx3);
 244             dy23             = _mm_sub_pd(iy2,jy3);
 245             dz23             = _mm_sub_pd(iz2,jz3);
 246             dx31             = _mm_sub_pd(ix3,jx1);
 247             dy31             = _mm_sub_pd(iy3,jy1);
 248             dz31             = _mm_sub_pd(iz3,jz1);
 249             dx32             = _mm_sub_pd(ix3,jx2);
 250             dy32             = _mm_sub_pd(iy3,jy2);
 251             dz32             = _mm_sub_pd(iz3,jz2);
 252             dx33             = _mm_sub_pd(ix3,jx3);
 253             dy33             = _mm_sub_pd(iy3,jy3);
 254             dz33             = _mm_sub_pd(iz3,jz3);
 255
 256             /* Calculate squared distance and things based on it */
 257             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 258             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 259             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 260             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 261             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 262             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 263             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 264             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 265             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 266             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 267
 268             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 269             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 270             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 271             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 272             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 273             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 274             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 275             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 276             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 277             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 278
 279             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 280             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 281             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 282             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 283             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 284             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 285             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 286             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 287             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 288             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 289
 290             fjx0             = _mm_setzero_pd();
 291             fjy0             = _mm_setzero_pd();
 292             fjz0             = _mm_setzero_pd();
 293             fjx1             = _mm_setzero_pd();
 294             fjy1             = _mm_setzero_pd();
 295             fjz1             = _mm_setzero_pd();
 296             fjx2             = _mm_setzero_pd();
 297             fjy2             = _mm_setzero_pd();
 298             fjz2             = _mm_setzero_pd();
 299             fjx3             = _mm_setzero_pd();
 300             fjy3             = _mm_setzero_pd();
 301             fjz3             = _mm_setzero_pd();
 302
 303             /**************************
 304              * CALCULATE INTERACTIONS *
 305              **************************/
 306
 307             if (gmx_mm_any_lt(rsq00,rcutoff2))
 308             {
 309
 310             r00              = _mm_mul_pd(rsq00,rinv00);
 311
 312             /* LENNARD-JONES DISPERSION/REPULSION */
 313
 314             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 315             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 316             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 317             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 318             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 319
 320             d                = _mm_sub_pd(r00,rswitch);
 321             d                = _mm_max_pd(d,_mm_setzero_pd());
 322             d2               = _mm_mul_pd(d,d);
 323             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)))))));
 324
 325             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
 326
 327             /* Evaluate switch function */
 328             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 329             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 330             vvdw             = _mm_mul_pd(vvdw,sw);
 331             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 335             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 336
 337             fscal            = fvdw;
 338
 339             fscal            = _mm_and_pd(fscal,cutoff_mask);
 340
 341             /* Calculate temporary vectorial force */
 342             tx               = _mm_mul_pd(fscal,dx00);
 343             ty               = _mm_mul_pd(fscal,dy00);
 344             tz               = _mm_mul_pd(fscal,dz00);
 345
 346             /* Update vectorial force */
 347             fix0             = _mm_add_pd(fix0,tx);
 348             fiy0             = _mm_add_pd(fiy0,ty);
 349             fiz0             = _mm_add_pd(fiz0,tz);
 350
 351             fjx0             = _mm_add_pd(fjx0,tx);
 352             fjy0             = _mm_add_pd(fjy0,ty);
 353             fjz0             = _mm_add_pd(fjz0,tz);
 354
 355             }
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             if (gmx_mm_any_lt(rsq11,rcutoff2))
 362             {
 363
 364             /* REACTION-FIELD ELECTROSTATICS */
 365             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 366             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 367
 368             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velec            = _mm_and_pd(velec,cutoff_mask);
 372             velecsum         = _mm_add_pd(velecsum,velec);
 373
 374             fscal            = felec;
 375
 376             fscal            = _mm_and_pd(fscal,cutoff_mask);
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm_mul_pd(fscal,dx11);
 380             ty               = _mm_mul_pd(fscal,dy11);
 381             tz               = _mm_mul_pd(fscal,dz11);
 382
 383             /* Update vectorial force */
 384             fix1             = _mm_add_pd(fix1,tx);
 385             fiy1             = _mm_add_pd(fiy1,ty);
 386             fiz1             = _mm_add_pd(fiz1,tz);
 387
 388             fjx1             = _mm_add_pd(fjx1,tx);
 389             fjy1             = _mm_add_pd(fjy1,ty);
 390             fjz1             = _mm_add_pd(fjz1,tz);
 391
 392             }
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             if (gmx_mm_any_lt(rsq12,rcutoff2))
 399             {
 400
 401             /* REACTION-FIELD ELECTROSTATICS */
 402             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 403             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 404
 405             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velec            = _mm_and_pd(velec,cutoff_mask);
 409             velecsum         = _mm_add_pd(velecsum,velec);
 410
 411             fscal            = felec;
 412
 413             fscal            = _mm_and_pd(fscal,cutoff_mask);
 414
 415             /* Calculate temporary vectorial force */
 416             tx               = _mm_mul_pd(fscal,dx12);
 417             ty               = _mm_mul_pd(fscal,dy12);
 418             tz               = _mm_mul_pd(fscal,dz12);
 419
 420             /* Update vectorial force */
 421             fix1             = _mm_add_pd(fix1,tx);
 422             fiy1             = _mm_add_pd(fiy1,ty);
 423             fiz1             = _mm_add_pd(fiz1,tz);
 424
 425             fjx2             = _mm_add_pd(fjx2,tx);
 426             fjy2             = _mm_add_pd(fjy2,ty);
 427             fjz2             = _mm_add_pd(fjz2,tz);
 428
 429             }
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             if (gmx_mm_any_lt(rsq13,rcutoff2))
 436             {
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 440             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 441
 442             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 443
 444             /* Update potential sum for this i atom from the interaction with this j atom. */
 445             velec            = _mm_and_pd(velec,cutoff_mask);
 446             velecsum         = _mm_add_pd(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             fscal            = _mm_and_pd(fscal,cutoff_mask);
 451
 452             /* Calculate temporary vectorial force */
 453             tx               = _mm_mul_pd(fscal,dx13);
 454             ty               = _mm_mul_pd(fscal,dy13);
 455             tz               = _mm_mul_pd(fscal,dz13);
 456
 457             /* Update vectorial force */
 458             fix1             = _mm_add_pd(fix1,tx);
 459             fiy1             = _mm_add_pd(fiy1,ty);
 460             fiz1             = _mm_add_pd(fiz1,tz);
 461
 462             fjx3             = _mm_add_pd(fjx3,tx);
 463             fjy3             = _mm_add_pd(fjy3,ty);
 464             fjz3             = _mm_add_pd(fjz3,tz);
 465
 466             }
 467
 468             /**************************
 469              * CALCULATE INTERACTIONS *
 470              **************************/
 471
 472             if (gmx_mm_any_lt(rsq21,rcutoff2))
 473             {
 474
 475             /* REACTION-FIELD ELECTROSTATICS */
 476             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 477             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 478
 479             cutoff_mask      = _mm_cmplt_pd(rsq21,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             /* Calculate temporary vectorial force */
 490             tx               = _mm_mul_pd(fscal,dx21);
 491             ty               = _mm_mul_pd(fscal,dy21);
 492             tz               = _mm_mul_pd(fscal,dz21);
 493
 494             /* Update vectorial force */
 495             fix2             = _mm_add_pd(fix2,tx);
 496             fiy2             = _mm_add_pd(fiy2,ty);
 497             fiz2             = _mm_add_pd(fiz2,tz);
 498
 499             fjx1             = _mm_add_pd(fjx1,tx);
 500             fjy1             = _mm_add_pd(fjy1,ty);
 501             fjz1             = _mm_add_pd(fjz1,tz);
 502
 503             }
 504
 505             /**************************
 506              * CALCULATE INTERACTIONS *
 507              **************************/
 508
 509             if (gmx_mm_any_lt(rsq22,rcutoff2))
 510             {
 511
 512             /* REACTION-FIELD ELECTROSTATICS */
 513             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 514             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 515
 516             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 517
 518             /* Update potential sum for this i atom from the interaction with this j atom. */
 519             velec            = _mm_and_pd(velec,cutoff_mask);
 520             velecsum         = _mm_add_pd(velecsum,velec);
 521
 522             fscal            = felec;
 523
 524             fscal            = _mm_and_pd(fscal,cutoff_mask);
 525
 526             /* Calculate temporary vectorial force */
 527             tx               = _mm_mul_pd(fscal,dx22);
 528             ty               = _mm_mul_pd(fscal,dy22);
 529             tz               = _mm_mul_pd(fscal,dz22);
 530
 531             /* Update vectorial force */
 532             fix2             = _mm_add_pd(fix2,tx);
 533             fiy2             = _mm_add_pd(fiy2,ty);
 534             fiz2             = _mm_add_pd(fiz2,tz);
 535
 536             fjx2             = _mm_add_pd(fjx2,tx);
 537             fjy2             = _mm_add_pd(fjy2,ty);
 538             fjz2             = _mm_add_pd(fjz2,tz);
 539
 540             }
 541
 542             /**************************
 543              * CALCULATE INTERACTIONS *
 544              **************************/
 545
 546             if (gmx_mm_any_lt(rsq23,rcutoff2))
 547             {
 548
 549             /* REACTION-FIELD ELECTROSTATICS */
 550             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 551             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 552
 553             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 554
 555             /* Update potential sum for this i atom from the interaction with this j atom. */
 556             velec            = _mm_and_pd(velec,cutoff_mask);
 557             velecsum         = _mm_add_pd(velecsum,velec);
 558
 559             fscal            = felec;
 560
 561             fscal            = _mm_and_pd(fscal,cutoff_mask);
 562
 563             /* Calculate temporary vectorial force */
 564             tx               = _mm_mul_pd(fscal,dx23);
 565             ty               = _mm_mul_pd(fscal,dy23);
 566             tz               = _mm_mul_pd(fscal,dz23);
 567
 568             /* Update vectorial force */
 569             fix2             = _mm_add_pd(fix2,tx);
 570             fiy2             = _mm_add_pd(fiy2,ty);
 571             fiz2             = _mm_add_pd(fiz2,tz);
 572
 573             fjx3             = _mm_add_pd(fjx3,tx);
 574             fjy3             = _mm_add_pd(fjy3,ty);
 575             fjz3             = _mm_add_pd(fjz3,tz);
 576
 577             }
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             if (gmx_mm_any_lt(rsq31,rcutoff2))
 584             {
 585
 586             /* REACTION-FIELD ELECTROSTATICS */
 587             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 588             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 589
 590             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velec            = _mm_and_pd(velec,cutoff_mask);
 594             velecsum         = _mm_add_pd(velecsum,velec);
 595
 596             fscal            = felec;
 597
 598             fscal            = _mm_and_pd(fscal,cutoff_mask);
 599
 600             /* Calculate temporary vectorial force */
 601             tx               = _mm_mul_pd(fscal,dx31);
 602             ty               = _mm_mul_pd(fscal,dy31);
 603             tz               = _mm_mul_pd(fscal,dz31);
 604
 605             /* Update vectorial force */
 606             fix3             = _mm_add_pd(fix3,tx);
 607             fiy3             = _mm_add_pd(fiy3,ty);
 608             fiz3             = _mm_add_pd(fiz3,tz);
 609
 610             fjx1             = _mm_add_pd(fjx1,tx);
 611             fjy1             = _mm_add_pd(fjy1,ty);
 612             fjz1             = _mm_add_pd(fjz1,tz);
 613
 614             }
 615
 616             /**************************
 617              * CALCULATE INTERACTIONS *
 618              **************************/
 619
 620             if (gmx_mm_any_lt(rsq32,rcutoff2))
 621             {
 622
 623             /* REACTION-FIELD ELECTROSTATICS */
 624             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 625             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 626
 627             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 628
 629             /* Update potential sum for this i atom from the interaction with this j atom. */
 630             velec            = _mm_and_pd(velec,cutoff_mask);
 631             velecsum         = _mm_add_pd(velecsum,velec);
 632
 633             fscal            = felec;
 634
 635             fscal            = _mm_and_pd(fscal,cutoff_mask);
 636
 637             /* Calculate temporary vectorial force */
 638             tx               = _mm_mul_pd(fscal,dx32);
 639             ty               = _mm_mul_pd(fscal,dy32);
 640             tz               = _mm_mul_pd(fscal,dz32);
 641
 642             /* Update vectorial force */
 643             fix3             = _mm_add_pd(fix3,tx);
 644             fiy3             = _mm_add_pd(fiy3,ty);
 645             fiz3             = _mm_add_pd(fiz3,tz);
 646
 647             fjx2             = _mm_add_pd(fjx2,tx);
 648             fjy2             = _mm_add_pd(fjy2,ty);
 649             fjz2             = _mm_add_pd(fjz2,tz);
 650
 651             }
 652
 653             /**************************
 654              * CALCULATE INTERACTIONS *
 655              **************************/
 656
 657             if (gmx_mm_any_lt(rsq33,rcutoff2))
 658             {
 659
 660             /* REACTION-FIELD ELECTROSTATICS */
 661             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 662             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 663
 664             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _mm_and_pd(velec,cutoff_mask);
 668             velecsum         = _mm_add_pd(velecsum,velec);
 669
 670             fscal            = felec;
 671
 672             fscal            = _mm_and_pd(fscal,cutoff_mask);
 673
 674             /* Calculate temporary vectorial force */
 675             tx               = _mm_mul_pd(fscal,dx33);
 676             ty               = _mm_mul_pd(fscal,dy33);
 677             tz               = _mm_mul_pd(fscal,dz33);
 678
 679             /* Update vectorial force */
 680             fix3             = _mm_add_pd(fix3,tx);
 681             fiy3             = _mm_add_pd(fiy3,ty);
 682             fiz3             = _mm_add_pd(fiz3,tz);
 683
 684             fjx3             = _mm_add_pd(fjx3,tx);
 685             fjy3             = _mm_add_pd(fjy3,ty);
 686             fjz3             = _mm_add_pd(fjz3,tz);
 687
 688             }
 689
 690             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 691
 692             /* Inner loop uses 386 flops */
 693         }
 694
 695         if(jidx<j_index_end)
 696         {
 697
 698             jnrA             = jjnr[jidx];
 699             j_coord_offsetA  = DIM*jnrA;
 700
 701             /* load j atom coordinates */
 702             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 703                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 704                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 705
 706             /* Calculate displacement vector */
 707             dx00             = _mm_sub_pd(ix0,jx0);
 708             dy00             = _mm_sub_pd(iy0,jy0);
 709             dz00             = _mm_sub_pd(iz0,jz0);
 710             dx11             = _mm_sub_pd(ix1,jx1);
 711             dy11             = _mm_sub_pd(iy1,jy1);
 712             dz11             = _mm_sub_pd(iz1,jz1);
 713             dx12             = _mm_sub_pd(ix1,jx2);
 714             dy12             = _mm_sub_pd(iy1,jy2);
 715             dz12             = _mm_sub_pd(iz1,jz2);
 716             dx13             = _mm_sub_pd(ix1,jx3);
 717             dy13             = _mm_sub_pd(iy1,jy3);
 718             dz13             = _mm_sub_pd(iz1,jz3);
 719             dx21             = _mm_sub_pd(ix2,jx1);
 720             dy21             = _mm_sub_pd(iy2,jy1);
 721             dz21             = _mm_sub_pd(iz2,jz1);
 722             dx22             = _mm_sub_pd(ix2,jx2);
 723             dy22             = _mm_sub_pd(iy2,jy2);
 724             dz22             = _mm_sub_pd(iz2,jz2);
 725             dx23             = _mm_sub_pd(ix2,jx3);
 726             dy23             = _mm_sub_pd(iy2,jy3);
 727             dz23             = _mm_sub_pd(iz2,jz3);
 728             dx31             = _mm_sub_pd(ix3,jx1);
 729             dy31             = _mm_sub_pd(iy3,jy1);
 730             dz31             = _mm_sub_pd(iz3,jz1);
 731             dx32             = _mm_sub_pd(ix3,jx2);
 732             dy32             = _mm_sub_pd(iy3,jy2);
 733             dz32             = _mm_sub_pd(iz3,jz2);
 734             dx33             = _mm_sub_pd(ix3,jx3);
 735             dy33             = _mm_sub_pd(iy3,jy3);
 736             dz33             = _mm_sub_pd(iz3,jz3);
 737
 738             /* Calculate squared distance and things based on it */
 739             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 740             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 741             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 742             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 743             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 744             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 745             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 746             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 747             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 748             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 749
 750             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 751             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 752             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 753             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 754             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 755             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 756             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 757             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 758             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 759             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 760
 761             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 762             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 763             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 764             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 765             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 766             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 767             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 768             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 769             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 770             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 771
 772             fjx0             = _mm_setzero_pd();
 773             fjy0             = _mm_setzero_pd();
 774             fjz0             = _mm_setzero_pd();
 775             fjx1             = _mm_setzero_pd();
 776             fjy1             = _mm_setzero_pd();
 777             fjz1             = _mm_setzero_pd();
 778             fjx2             = _mm_setzero_pd();
 779             fjy2             = _mm_setzero_pd();
 780             fjz2             = _mm_setzero_pd();
 781             fjx3             = _mm_setzero_pd();
 782             fjy3             = _mm_setzero_pd();
 783             fjz3             = _mm_setzero_pd();
 784
 785             /**************************
 786              * CALCULATE INTERACTIONS *
 787              **************************/
 788
 789             if (gmx_mm_any_lt(rsq00,rcutoff2))
 790             {
 791
 792             r00              = _mm_mul_pd(rsq00,rinv00);
 793
 794             /* LENNARD-JONES DISPERSION/REPULSION */
 795
 796             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 797             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 798             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 799             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 800             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 801
 802             d                = _mm_sub_pd(r00,rswitch);
 803             d                = _mm_max_pd(d,_mm_setzero_pd());
 804             d2               = _mm_mul_pd(d,d);
 805             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)))))));
 806
 807             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
 808
 809             /* Evaluate switch function */
 810             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 811             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 812             vvdw             = _mm_mul_pd(vvdw,sw);
 813             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 814
 815             /* Update potential sum for this i atom from the interaction with this j atom. */
 816             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 817             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 818             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 819
 820             fscal            = fvdw;
 821
 822             fscal            = _mm_and_pd(fscal,cutoff_mask);
 823
 824             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 825
 826             /* Calculate temporary vectorial force */
 827             tx               = _mm_mul_pd(fscal,dx00);
 828             ty               = _mm_mul_pd(fscal,dy00);
 829             tz               = _mm_mul_pd(fscal,dz00);
 830
 831             /* Update vectorial force */
 832             fix0             = _mm_add_pd(fix0,tx);
 833             fiy0             = _mm_add_pd(fiy0,ty);
 834             fiz0             = _mm_add_pd(fiz0,tz);
 835
 836             fjx0             = _mm_add_pd(fjx0,tx);
 837             fjy0             = _mm_add_pd(fjy0,ty);
 838             fjz0             = _mm_add_pd(fjz0,tz);
 839
 840             }
 841
 842             /**************************
 843              * CALCULATE INTERACTIONS *
 844              **************************/
 845
 846             if (gmx_mm_any_lt(rsq11,rcutoff2))
 847             {
 848
 849             /* REACTION-FIELD ELECTROSTATICS */
 850             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 851             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 852
 853             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 854
 855             /* Update potential sum for this i atom from the interaction with this j atom. */
 856             velec            = _mm_and_pd(velec,cutoff_mask);
 857             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 858             velecsum         = _mm_add_pd(velecsum,velec);
 859
 860             fscal            = felec;
 861
 862             fscal            = _mm_and_pd(fscal,cutoff_mask);
 863
 864             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 865
 866             /* Calculate temporary vectorial force */
 867             tx               = _mm_mul_pd(fscal,dx11);
 868             ty               = _mm_mul_pd(fscal,dy11);
 869             tz               = _mm_mul_pd(fscal,dz11);
 870
 871             /* Update vectorial force */
 872             fix1             = _mm_add_pd(fix1,tx);
 873             fiy1             = _mm_add_pd(fiy1,ty);
 874             fiz1             = _mm_add_pd(fiz1,tz);
 875
 876             fjx1             = _mm_add_pd(fjx1,tx);
 877             fjy1             = _mm_add_pd(fjy1,ty);
 878             fjz1             = _mm_add_pd(fjz1,tz);
 879
 880             }
 881
 882             /**************************
 883              * CALCULATE INTERACTIONS *
 884              **************************/
 885
 886             if (gmx_mm_any_lt(rsq12,rcutoff2))
 887             {
 888
 889             /* REACTION-FIELD ELECTROSTATICS */
 890             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 891             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 892
 893             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 894
 895             /* Update potential sum for this i atom from the interaction with this j atom. */
 896             velec            = _mm_and_pd(velec,cutoff_mask);
 897             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 898             velecsum         = _mm_add_pd(velecsum,velec);
 899
 900             fscal            = felec;
 901
 902             fscal            = _mm_and_pd(fscal,cutoff_mask);
 903
 904             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 905
 906             /* Calculate temporary vectorial force */
 907             tx               = _mm_mul_pd(fscal,dx12);
 908             ty               = _mm_mul_pd(fscal,dy12);
 909             tz               = _mm_mul_pd(fscal,dz12);
 910
 911             /* Update vectorial force */
 912             fix1             = _mm_add_pd(fix1,tx);
 913             fiy1             = _mm_add_pd(fiy1,ty);
 914             fiz1             = _mm_add_pd(fiz1,tz);
 915
 916             fjx2             = _mm_add_pd(fjx2,tx);
 917             fjy2             = _mm_add_pd(fjy2,ty);
 918             fjz2             = _mm_add_pd(fjz2,tz);
 919
 920             }
 921
 922             /**************************
 923              * CALCULATE INTERACTIONS *
 924              **************************/
 925
 926             if (gmx_mm_any_lt(rsq13,rcutoff2))
 927             {
 928
 929             /* REACTION-FIELD ELECTROSTATICS */
 930             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 931             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 932
 933             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 934
 935             /* Update potential sum for this i atom from the interaction with this j atom. */
 936             velec            = _mm_and_pd(velec,cutoff_mask);
 937             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 938             velecsum         = _mm_add_pd(velecsum,velec);
 939
 940             fscal            = felec;
 941
 942             fscal            = _mm_and_pd(fscal,cutoff_mask);
 943
 944             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 945
 946             /* Calculate temporary vectorial force */
 947             tx               = _mm_mul_pd(fscal,dx13);
 948             ty               = _mm_mul_pd(fscal,dy13);
 949             tz               = _mm_mul_pd(fscal,dz13);
 950
 951             /* Update vectorial force */
 952             fix1             = _mm_add_pd(fix1,tx);
 953             fiy1             = _mm_add_pd(fiy1,ty);
 954             fiz1             = _mm_add_pd(fiz1,tz);
 955
 956             fjx3             = _mm_add_pd(fjx3,tx);
 957             fjy3             = _mm_add_pd(fjy3,ty);
 958             fjz3             = _mm_add_pd(fjz3,tz);
 959
 960             }
 961
 962             /**************************
 963              * CALCULATE INTERACTIONS *
 964              **************************/
 965
 966             if (gmx_mm_any_lt(rsq21,rcutoff2))
 967             {
 968
 969             /* REACTION-FIELD ELECTROSTATICS */
 970             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 971             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 972
 973             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 974
 975             /* Update potential sum for this i atom from the interaction with this j atom. */
 976             velec            = _mm_and_pd(velec,cutoff_mask);
 977             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 978             velecsum         = _mm_add_pd(velecsum,velec);
 979
 980             fscal            = felec;
 981
 982             fscal            = _mm_and_pd(fscal,cutoff_mask);
 983
 984             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 985
 986             /* Calculate temporary vectorial force */
 987             tx               = _mm_mul_pd(fscal,dx21);
 988             ty               = _mm_mul_pd(fscal,dy21);
 989             tz               = _mm_mul_pd(fscal,dz21);
 990
 991             /* Update vectorial force */
 992             fix2             = _mm_add_pd(fix2,tx);
 993             fiy2             = _mm_add_pd(fiy2,ty);
 994             fiz2             = _mm_add_pd(fiz2,tz);
 995
 996             fjx1             = _mm_add_pd(fjx1,tx);
 997             fjy1             = _mm_add_pd(fjy1,ty);
 998             fjz1             = _mm_add_pd(fjz1,tz);
 999
1000             }
1001
1002             /**************************
1003              * CALCULATE INTERACTIONS *
1004              **************************/
1005
1006             if (gmx_mm_any_lt(rsq22,rcutoff2))
1007             {
1008
1009             /* REACTION-FIELD ELECTROSTATICS */
1010             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1011             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1012
1013             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1014
1015             /* Update potential sum for this i atom from the interaction with this j atom. */
1016             velec            = _mm_and_pd(velec,cutoff_mask);
1017             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1018             velecsum         = _mm_add_pd(velecsum,velec);
1019
1020             fscal            = felec;
1021
1022             fscal            = _mm_and_pd(fscal,cutoff_mask);
1023
1024             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1025
1026             /* Calculate temporary vectorial force */
1027             tx               = _mm_mul_pd(fscal,dx22);
1028             ty               = _mm_mul_pd(fscal,dy22);
1029             tz               = _mm_mul_pd(fscal,dz22);
1030
1031             /* Update vectorial force */
1032             fix2             = _mm_add_pd(fix2,tx);
1033             fiy2             = _mm_add_pd(fiy2,ty);
1034             fiz2             = _mm_add_pd(fiz2,tz);
1035
1036             fjx2             = _mm_add_pd(fjx2,tx);
1037             fjy2             = _mm_add_pd(fjy2,ty);
1038             fjz2             = _mm_add_pd(fjz2,tz);
1039
1040             }
1041
1042             /**************************
1043              * CALCULATE INTERACTIONS *
1044              **************************/
1045
1046             if (gmx_mm_any_lt(rsq23,rcutoff2))
1047             {
1048
1049             /* REACTION-FIELD ELECTROSTATICS */
1050             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
1051             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1052
1053             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1054
1055             /* Update potential sum for this i atom from the interaction with this j atom. */
1056             velec            = _mm_and_pd(velec,cutoff_mask);
1057             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1058             velecsum         = _mm_add_pd(velecsum,velec);
1059
1060             fscal            = felec;
1061
1062             fscal            = _mm_and_pd(fscal,cutoff_mask);
1063
1064             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1065
1066             /* Calculate temporary vectorial force */
1067             tx               = _mm_mul_pd(fscal,dx23);
1068             ty               = _mm_mul_pd(fscal,dy23);
1069             tz               = _mm_mul_pd(fscal,dz23);
1070
1071             /* Update vectorial force */
1072             fix2             = _mm_add_pd(fix2,tx);
1073             fiy2             = _mm_add_pd(fiy2,ty);
1074             fiz2             = _mm_add_pd(fiz2,tz);
1075
1076             fjx3             = _mm_add_pd(fjx3,tx);
1077             fjy3             = _mm_add_pd(fjy3,ty);
1078             fjz3             = _mm_add_pd(fjz3,tz);
1079
1080             }
1081
1082             /**************************
1083              * CALCULATE INTERACTIONS *
1084              **************************/
1085
1086             if (gmx_mm_any_lt(rsq31,rcutoff2))
1087             {
1088
1089             /* REACTION-FIELD ELECTROSTATICS */
1090             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
1091             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1092
1093             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1094
1095             /* Update potential sum for this i atom from the interaction with this j atom. */
1096             velec            = _mm_and_pd(velec,cutoff_mask);
1097             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1098             velecsum         = _mm_add_pd(velecsum,velec);
1099
1100             fscal            = felec;
1101
1102             fscal            = _mm_and_pd(fscal,cutoff_mask);
1103
1104             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1105
1106             /* Calculate temporary vectorial force */
1107             tx               = _mm_mul_pd(fscal,dx31);
1108             ty               = _mm_mul_pd(fscal,dy31);
1109             tz               = _mm_mul_pd(fscal,dz31);
1110
1111             /* Update vectorial force */
1112             fix3             = _mm_add_pd(fix3,tx);
1113             fiy3             = _mm_add_pd(fiy3,ty);
1114             fiz3             = _mm_add_pd(fiz3,tz);
1115
1116             fjx1             = _mm_add_pd(fjx1,tx);
1117             fjy1             = _mm_add_pd(fjy1,ty);
1118             fjz1             = _mm_add_pd(fjz1,tz);
1119
1120             }
1121
1122             /**************************
1123              * CALCULATE INTERACTIONS *
1124              **************************/
1125
1126             if (gmx_mm_any_lt(rsq32,rcutoff2))
1127             {
1128
1129             /* REACTION-FIELD ELECTROSTATICS */
1130             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
1131             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1132
1133             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1134
1135             /* Update potential sum for this i atom from the interaction with this j atom. */
1136             velec            = _mm_and_pd(velec,cutoff_mask);
1137             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1138             velecsum         = _mm_add_pd(velecsum,velec);
1139
1140             fscal            = felec;
1141
1142             fscal            = _mm_and_pd(fscal,cutoff_mask);
1143
1144             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1145
1146             /* Calculate temporary vectorial force */
1147             tx               = _mm_mul_pd(fscal,dx32);
1148             ty               = _mm_mul_pd(fscal,dy32);
1149             tz               = _mm_mul_pd(fscal,dz32);
1150
1151             /* Update vectorial force */
1152             fix3             = _mm_add_pd(fix3,tx);
1153             fiy3             = _mm_add_pd(fiy3,ty);
1154             fiz3             = _mm_add_pd(fiz3,tz);
1155
1156             fjx2             = _mm_add_pd(fjx2,tx);
1157             fjy2             = _mm_add_pd(fjy2,ty);
1158             fjz2             = _mm_add_pd(fjz2,tz);
1159
1160             }
1161
1162             /**************************
1163              * CALCULATE INTERACTIONS *
1164              **************************/
1165
1166             if (gmx_mm_any_lt(rsq33,rcutoff2))
1167             {
1168
1169             /* REACTION-FIELD ELECTROSTATICS */
1170             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1171             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1172
1173             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1174
1175             /* Update potential sum for this i atom from the interaction with this j atom. */
1176             velec            = _mm_and_pd(velec,cutoff_mask);
1177             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1178             velecsum         = _mm_add_pd(velecsum,velec);
1179
1180             fscal            = felec;
1181
1182             fscal            = _mm_and_pd(fscal,cutoff_mask);
1183
1184             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1185
1186             /* Calculate temporary vectorial force */
1187             tx               = _mm_mul_pd(fscal,dx33);
1188             ty               = _mm_mul_pd(fscal,dy33);
1189             tz               = _mm_mul_pd(fscal,dz33);
1190
1191             /* Update vectorial force */
1192             fix3             = _mm_add_pd(fix3,tx);
1193             fiy3             = _mm_add_pd(fiy3,ty);
1194             fiz3             = _mm_add_pd(fiz3,tz);
1195
1196             fjx3             = _mm_add_pd(fjx3,tx);
1197             fjy3             = _mm_add_pd(fjy3,ty);
1198             fjz3             = _mm_add_pd(fjz3,tz);
1199
1200             }
1201
1202             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1203
1204             /* Inner loop uses 386 flops */
1205         }
1206
1207         /* End of innermost loop */
1208
1209         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1210                                               f+i_coord_offset,fshift+i_shift_offset);
1211
1212         ggid                        = gid[iidx];
1213         /* Update potential energies */
1214         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1215         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1216
1217         /* Increment number of inner iterations */
1218         inneriter                  += j_index_end - j_index_start;
1219
1220         /* Outer loop uses 26 flops */
1221     }
1222
1223     /* Increment number of outer iterations */
1224     outeriter        += nri;
1225
1226     /* Update outer/inner flops */
1227
1228     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*386);
1229 }
1230 /*
1231  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sse4_1_double
1232  * Electrostatics interaction: ReactionField
1233  * VdW interaction:            LennardJones
1234  * Geometry:                   Water4-Water4
1235  * Calculate force/pot:        Force
1236  */
1237 void
1238 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sse4_1_double
1239                     (t_nblist * gmx_restrict                nlist,
1240                      rvec * gmx_restrict                    xx,
1241                      rvec * gmx_restrict                    ff,
1242                      t_forcerec * gmx_restrict              fr,
1243                      t_mdatoms * gmx_restrict               mdatoms,
1244                      nb_kernel_data_t * gmx_restrict        kernel_data,
1245                      t_nrnb * gmx_restrict                  nrnb)
1246 {
1247     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1248      * just 0 for non-waters.
1249      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1250      * jnr indices corresponding to data put in the four positions in the SIMD register.
1251      */
1252     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1253     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1254     int              jnrA,jnrB;
1255     int              j_coord_offsetA,j_coord_offsetB;
1256     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1257     real             rcutoff_scalar;
1258     real             *shiftvec,*fshift,*x,*f;
1259     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1260     int              vdwioffset0;
1261     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1262     int              vdwioffset1;
1263     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1264     int              vdwioffset2;
1265     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1266     int              vdwioffset3;
1267     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1268     int              vdwjidx0A,vdwjidx0B;
1269     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1270     int              vdwjidx1A,vdwjidx1B;
1271     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1272     int              vdwjidx2A,vdwjidx2B;
1273     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1274     int              vdwjidx3A,vdwjidx3B;
1275     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1276     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1277     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1278     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1279     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1280     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1281     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1282     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1283     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1284     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1285     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1286     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1287     real             *charge;
1288     int              nvdwtype;
1289     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1290     int              *vdwtype;
1291     real             *vdwparam;
1292     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1293     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1294     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1295     real             rswitch_scalar,d_scalar;
1296     __m128d          dummy_mask,cutoff_mask;
1297     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1298     __m128d          one     = _mm_set1_pd(1.0);
1299     __m128d          two     = _mm_set1_pd(2.0);
1300     x                = xx[0];
1301     f                = ff[0];
1302
1303     nri              = nlist->nri;
1304     iinr             = nlist->iinr;
1305     jindex           = nlist->jindex;
1306     jjnr             = nlist->jjnr;
1307     shiftidx         = nlist->shift;
1308     gid              = nlist->gid;
1309     shiftvec         = fr->shift_vec[0];
1310     fshift           = fr->fshift[0];
1311     facel            = _mm_set1_pd(fr->epsfac);
1312     charge           = mdatoms->chargeA;
1313     krf              = _mm_set1_pd(fr->ic->k_rf);
1314     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1315     crf              = _mm_set1_pd(fr->ic->c_rf);
1316     nvdwtype         = fr->ntype;
1317     vdwparam         = fr->nbfp;
1318     vdwtype          = mdatoms->typeA;
1319
1320     /* Setup water-specific parameters */
1321     inr              = nlist->iinr[0];
1322     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1323     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1324     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1325     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1326
1327     jq1              = _mm_set1_pd(charge[inr+1]);
1328     jq2              = _mm_set1_pd(charge[inr+2]);
1329     jq3              = _mm_set1_pd(charge[inr+3]);
1330     vdwjidx0A        = 2*vdwtype[inr+0];
1331     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1332     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1333     qq11             = _mm_mul_pd(iq1,jq1);
1334     qq12             = _mm_mul_pd(iq1,jq2);
1335     qq13             = _mm_mul_pd(iq1,jq3);
1336     qq21             = _mm_mul_pd(iq2,jq1);
1337     qq22             = _mm_mul_pd(iq2,jq2);
1338     qq23             = _mm_mul_pd(iq2,jq3);
1339     qq31             = _mm_mul_pd(iq3,jq1);
1340     qq32             = _mm_mul_pd(iq3,jq2);
1341     qq33             = _mm_mul_pd(iq3,jq3);
1342
1343     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1344     rcutoff_scalar   = fr->rcoulomb;
1345     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1346     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1347
1348     rswitch_scalar   = fr->rvdw_switch;
1349     rswitch          = _mm_set1_pd(rswitch_scalar);
1350     /* Setup switch parameters */
1351     d_scalar         = rcutoff_scalar-rswitch_scalar;
1352     d                = _mm_set1_pd(d_scalar);
1353     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1354     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1355     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1356     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1357     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1358     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1359
1360     /* Avoid stupid compiler warnings */
1361     jnrA = jnrB = 0;
1362     j_coord_offsetA = 0;
1363     j_coord_offsetB = 0;
1364
1365     outeriter        = 0;
1366     inneriter        = 0;
1367
1368     /* Start outer loop over neighborlists */
1369     for(iidx=0; iidx<nri; iidx++)
1370     {
1371         /* Load shift vector for this list */
1372         i_shift_offset   = DIM*shiftidx[iidx];
1373
1374         /* Load limits for loop over neighbors */
1375         j_index_start    = jindex[iidx];
1376         j_index_end      = jindex[iidx+1];
1377
1378         /* Get outer coordinate index */
1379         inr              = iinr[iidx];
1380         i_coord_offset   = DIM*inr;
1381
1382         /* Load i particle coords and add shift vector */
1383         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1384                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1385
1386         fix0             = _mm_setzero_pd();
1387         fiy0             = _mm_setzero_pd();
1388         fiz0             = _mm_setzero_pd();
1389         fix1             = _mm_setzero_pd();
1390         fiy1             = _mm_setzero_pd();
1391         fiz1             = _mm_setzero_pd();
1392         fix2             = _mm_setzero_pd();
1393         fiy2             = _mm_setzero_pd();
1394         fiz2             = _mm_setzero_pd();
1395         fix3             = _mm_setzero_pd();
1396         fiy3             = _mm_setzero_pd();
1397         fiz3             = _mm_setzero_pd();
1398
1399         /* Start inner kernel loop */
1400         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1401         {
1402
1403             /* Get j neighbor index, and coordinate index */
1404             jnrA             = jjnr[jidx];
1405             jnrB             = jjnr[jidx+1];
1406             j_coord_offsetA  = DIM*jnrA;
1407             j_coord_offsetB  = DIM*jnrB;
1408
1409             /* load j atom coordinates */
1410             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1411                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1412                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1413
1414             /* Calculate displacement vector */
1415             dx00             = _mm_sub_pd(ix0,jx0);
1416             dy00             = _mm_sub_pd(iy0,jy0);
1417             dz00             = _mm_sub_pd(iz0,jz0);
1418             dx11             = _mm_sub_pd(ix1,jx1);
1419             dy11             = _mm_sub_pd(iy1,jy1);
1420             dz11             = _mm_sub_pd(iz1,jz1);
1421             dx12             = _mm_sub_pd(ix1,jx2);
1422             dy12             = _mm_sub_pd(iy1,jy2);
1423             dz12             = _mm_sub_pd(iz1,jz2);
1424             dx13             = _mm_sub_pd(ix1,jx3);
1425             dy13             = _mm_sub_pd(iy1,jy3);
1426             dz13             = _mm_sub_pd(iz1,jz3);
1427             dx21             = _mm_sub_pd(ix2,jx1);
1428             dy21             = _mm_sub_pd(iy2,jy1);
1429             dz21             = _mm_sub_pd(iz2,jz1);
1430             dx22             = _mm_sub_pd(ix2,jx2);
1431             dy22             = _mm_sub_pd(iy2,jy2);
1432             dz22             = _mm_sub_pd(iz2,jz2);
1433             dx23             = _mm_sub_pd(ix2,jx3);
1434             dy23             = _mm_sub_pd(iy2,jy3);
1435             dz23             = _mm_sub_pd(iz2,jz3);
1436             dx31             = _mm_sub_pd(ix3,jx1);
1437             dy31             = _mm_sub_pd(iy3,jy1);
1438             dz31             = _mm_sub_pd(iz3,jz1);
1439             dx32             = _mm_sub_pd(ix3,jx2);
1440             dy32             = _mm_sub_pd(iy3,jy2);
1441             dz32             = _mm_sub_pd(iz3,jz2);
1442             dx33             = _mm_sub_pd(ix3,jx3);
1443             dy33             = _mm_sub_pd(iy3,jy3);
1444             dz33             = _mm_sub_pd(iz3,jz3);
1445
1446             /* Calculate squared distance and things based on it */
1447             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1448             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1449             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1450             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1451             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1452             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1453             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1454             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1455             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1456             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1457
1458             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1459             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1460             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1461             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1462             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1463             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1464             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1465             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1466             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1467             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1468
1469             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1470             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1471             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1472             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1473             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1474             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1475             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1476             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1477             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1478             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1479
1480             fjx0             = _mm_setzero_pd();
1481             fjy0             = _mm_setzero_pd();
1482             fjz0             = _mm_setzero_pd();
1483             fjx1             = _mm_setzero_pd();
1484             fjy1             = _mm_setzero_pd();
1485             fjz1             = _mm_setzero_pd();
1486             fjx2             = _mm_setzero_pd();
1487             fjy2             = _mm_setzero_pd();
1488             fjz2             = _mm_setzero_pd();
1489             fjx3             = _mm_setzero_pd();
1490             fjy3             = _mm_setzero_pd();
1491             fjz3             = _mm_setzero_pd();
1492
1493             /**************************
1494              * CALCULATE INTERACTIONS *
1495              **************************/
1496
1497             if (gmx_mm_any_lt(rsq00,rcutoff2))
1498             {
1499
1500             r00              = _mm_mul_pd(rsq00,rinv00);
1501
1502             /* LENNARD-JONES DISPERSION/REPULSION */
1503
1504             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1505             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1506             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1507             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
1508             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1509
1510             d                = _mm_sub_pd(r00,rswitch);
1511             d                = _mm_max_pd(d,_mm_setzero_pd());
1512             d2               = _mm_mul_pd(d,d);
1513             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)))))));
1514
1515             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
1516
1517             /* Evaluate switch function */
1518             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1519             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1520             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1521
1522             fscal            = fvdw;
1523
1524             fscal            = _mm_and_pd(fscal,cutoff_mask);
1525
1526             /* Calculate temporary vectorial force */
1527             tx               = _mm_mul_pd(fscal,dx00);
1528             ty               = _mm_mul_pd(fscal,dy00);
1529             tz               = _mm_mul_pd(fscal,dz00);
1530
1531             /* Update vectorial force */
1532             fix0             = _mm_add_pd(fix0,tx);
1533             fiy0             = _mm_add_pd(fiy0,ty);
1534             fiz0             = _mm_add_pd(fiz0,tz);
1535
1536             fjx0             = _mm_add_pd(fjx0,tx);
1537             fjy0             = _mm_add_pd(fjy0,ty);
1538             fjz0             = _mm_add_pd(fjz0,tz);
1539
1540             }
1541
1542             /**************************
1543              * CALCULATE INTERACTIONS *
1544              **************************/
1545
1546             if (gmx_mm_any_lt(rsq11,rcutoff2))
1547             {
1548
1549             /* REACTION-FIELD ELECTROSTATICS */
1550             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1551
1552             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1553
1554             fscal            = felec;
1555
1556             fscal            = _mm_and_pd(fscal,cutoff_mask);
1557
1558             /* Calculate temporary vectorial force */
1559             tx               = _mm_mul_pd(fscal,dx11);
1560             ty               = _mm_mul_pd(fscal,dy11);
1561             tz               = _mm_mul_pd(fscal,dz11);
1562
1563             /* Update vectorial force */
1564             fix1             = _mm_add_pd(fix1,tx);
1565             fiy1             = _mm_add_pd(fiy1,ty);
1566             fiz1             = _mm_add_pd(fiz1,tz);
1567
1568             fjx1             = _mm_add_pd(fjx1,tx);
1569             fjy1             = _mm_add_pd(fjy1,ty);
1570             fjz1             = _mm_add_pd(fjz1,tz);
1571
1572             }
1573
1574             /**************************
1575              * CALCULATE INTERACTIONS *
1576              **************************/
1577
1578             if (gmx_mm_any_lt(rsq12,rcutoff2))
1579             {
1580
1581             /* REACTION-FIELD ELECTROSTATICS */
1582             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1583
1584             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1585
1586             fscal            = felec;
1587
1588             fscal            = _mm_and_pd(fscal,cutoff_mask);
1589
1590             /* Calculate temporary vectorial force */
1591             tx               = _mm_mul_pd(fscal,dx12);
1592             ty               = _mm_mul_pd(fscal,dy12);
1593             tz               = _mm_mul_pd(fscal,dz12);
1594
1595             /* Update vectorial force */
1596             fix1             = _mm_add_pd(fix1,tx);
1597             fiy1             = _mm_add_pd(fiy1,ty);
1598             fiz1             = _mm_add_pd(fiz1,tz);
1599
1600             fjx2             = _mm_add_pd(fjx2,tx);
1601             fjy2             = _mm_add_pd(fjy2,ty);
1602             fjz2             = _mm_add_pd(fjz2,tz);
1603
1604             }
1605
1606             /**************************
1607              * CALCULATE INTERACTIONS *
1608              **************************/
1609
1610             if (gmx_mm_any_lt(rsq13,rcutoff2))
1611             {
1612
1613             /* REACTION-FIELD ELECTROSTATICS */
1614             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1615
1616             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1617
1618             fscal            = felec;
1619
1620             fscal            = _mm_and_pd(fscal,cutoff_mask);
1621
1622             /* Calculate temporary vectorial force */
1623             tx               = _mm_mul_pd(fscal,dx13);
1624             ty               = _mm_mul_pd(fscal,dy13);
1625             tz               = _mm_mul_pd(fscal,dz13);
1626
1627             /* Update vectorial force */
1628             fix1             = _mm_add_pd(fix1,tx);
1629             fiy1             = _mm_add_pd(fiy1,ty);
1630             fiz1             = _mm_add_pd(fiz1,tz);
1631
1632             fjx3             = _mm_add_pd(fjx3,tx);
1633             fjy3             = _mm_add_pd(fjy3,ty);
1634             fjz3             = _mm_add_pd(fjz3,tz);
1635
1636             }
1637
1638             /**************************
1639              * CALCULATE INTERACTIONS *
1640              **************************/
1641
1642             if (gmx_mm_any_lt(rsq21,rcutoff2))
1643             {
1644
1645             /* REACTION-FIELD ELECTROSTATICS */
1646             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1647
1648             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1649
1650             fscal            = felec;
1651
1652             fscal            = _mm_and_pd(fscal,cutoff_mask);
1653
1654             /* Calculate temporary vectorial force */
1655             tx               = _mm_mul_pd(fscal,dx21);
1656             ty               = _mm_mul_pd(fscal,dy21);
1657             tz               = _mm_mul_pd(fscal,dz21);
1658
1659             /* Update vectorial force */
1660             fix2             = _mm_add_pd(fix2,tx);
1661             fiy2             = _mm_add_pd(fiy2,ty);
1662             fiz2             = _mm_add_pd(fiz2,tz);
1663
1664             fjx1             = _mm_add_pd(fjx1,tx);
1665             fjy1             = _mm_add_pd(fjy1,ty);
1666             fjz1             = _mm_add_pd(fjz1,tz);
1667
1668             }
1669
1670             /**************************
1671              * CALCULATE INTERACTIONS *
1672              **************************/
1673
1674             if (gmx_mm_any_lt(rsq22,rcutoff2))
1675             {
1676
1677             /* REACTION-FIELD ELECTROSTATICS */
1678             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1679
1680             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1681
1682             fscal            = felec;
1683
1684             fscal            = _mm_and_pd(fscal,cutoff_mask);
1685
1686             /* Calculate temporary vectorial force */
1687             tx               = _mm_mul_pd(fscal,dx22);
1688             ty               = _mm_mul_pd(fscal,dy22);
1689             tz               = _mm_mul_pd(fscal,dz22);
1690
1691             /* Update vectorial force */
1692             fix2             = _mm_add_pd(fix2,tx);
1693             fiy2             = _mm_add_pd(fiy2,ty);
1694             fiz2             = _mm_add_pd(fiz2,tz);
1695
1696             fjx2             = _mm_add_pd(fjx2,tx);
1697             fjy2             = _mm_add_pd(fjy2,ty);
1698             fjz2             = _mm_add_pd(fjz2,tz);
1699
1700             }
1701
1702             /**************************
1703              * CALCULATE INTERACTIONS *
1704              **************************/
1705
1706             if (gmx_mm_any_lt(rsq23,rcutoff2))
1707             {
1708
1709             /* REACTION-FIELD ELECTROSTATICS */
1710             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1711
1712             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1713
1714             fscal            = felec;
1715
1716             fscal            = _mm_and_pd(fscal,cutoff_mask);
1717
1718             /* Calculate temporary vectorial force */
1719             tx               = _mm_mul_pd(fscal,dx23);
1720             ty               = _mm_mul_pd(fscal,dy23);
1721             tz               = _mm_mul_pd(fscal,dz23);
1722
1723             /* Update vectorial force */
1724             fix2             = _mm_add_pd(fix2,tx);
1725             fiy2             = _mm_add_pd(fiy2,ty);
1726             fiz2             = _mm_add_pd(fiz2,tz);
1727
1728             fjx3             = _mm_add_pd(fjx3,tx);
1729             fjy3             = _mm_add_pd(fjy3,ty);
1730             fjz3             = _mm_add_pd(fjz3,tz);
1731
1732             }
1733
1734             /**************************
1735              * CALCULATE INTERACTIONS *
1736              **************************/
1737
1738             if (gmx_mm_any_lt(rsq31,rcutoff2))
1739             {
1740
1741             /* REACTION-FIELD ELECTROSTATICS */
1742             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1743
1744             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1745
1746             fscal            = felec;
1747
1748             fscal            = _mm_and_pd(fscal,cutoff_mask);
1749
1750             /* Calculate temporary vectorial force */
1751             tx               = _mm_mul_pd(fscal,dx31);
1752             ty               = _mm_mul_pd(fscal,dy31);
1753             tz               = _mm_mul_pd(fscal,dz31);
1754
1755             /* Update vectorial force */
1756             fix3             = _mm_add_pd(fix3,tx);
1757             fiy3             = _mm_add_pd(fiy3,ty);
1758             fiz3             = _mm_add_pd(fiz3,tz);
1759
1760             fjx1             = _mm_add_pd(fjx1,tx);
1761             fjy1             = _mm_add_pd(fjy1,ty);
1762             fjz1             = _mm_add_pd(fjz1,tz);
1763
1764             }
1765
1766             /**************************
1767              * CALCULATE INTERACTIONS *
1768              **************************/
1769
1770             if (gmx_mm_any_lt(rsq32,rcutoff2))
1771             {
1772
1773             /* REACTION-FIELD ELECTROSTATICS */
1774             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1775
1776             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1777
1778             fscal            = felec;
1779
1780             fscal            = _mm_and_pd(fscal,cutoff_mask);
1781
1782             /* Calculate temporary vectorial force */
1783             tx               = _mm_mul_pd(fscal,dx32);
1784             ty               = _mm_mul_pd(fscal,dy32);
1785             tz               = _mm_mul_pd(fscal,dz32);
1786
1787             /* Update vectorial force */
1788             fix3             = _mm_add_pd(fix3,tx);
1789             fiy3             = _mm_add_pd(fiy3,ty);
1790             fiz3             = _mm_add_pd(fiz3,tz);
1791
1792             fjx2             = _mm_add_pd(fjx2,tx);
1793             fjy2             = _mm_add_pd(fjy2,ty);
1794             fjz2             = _mm_add_pd(fjz2,tz);
1795
1796             }
1797
1798             /**************************
1799              * CALCULATE INTERACTIONS *
1800              **************************/
1801
1802             if (gmx_mm_any_lt(rsq33,rcutoff2))
1803             {
1804
1805             /* REACTION-FIELD ELECTROSTATICS */
1806             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1807
1808             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1809
1810             fscal            = felec;
1811
1812             fscal            = _mm_and_pd(fscal,cutoff_mask);
1813
1814             /* Calculate temporary vectorial force */
1815             tx               = _mm_mul_pd(fscal,dx33);
1816             ty               = _mm_mul_pd(fscal,dy33);
1817             tz               = _mm_mul_pd(fscal,dz33);
1818
1819             /* Update vectorial force */
1820             fix3             = _mm_add_pd(fix3,tx);
1821             fiy3             = _mm_add_pd(fiy3,ty);
1822             fiz3             = _mm_add_pd(fiz3,tz);
1823
1824             fjx3             = _mm_add_pd(fjx3,tx);
1825             fjy3             = _mm_add_pd(fjy3,ty);
1826             fjz3             = _mm_add_pd(fjz3,tz);
1827
1828             }
1829
1830             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1831
1832             /* Inner loop uses 329 flops */
1833         }
1834
1835         if(jidx<j_index_end)
1836         {
1837
1838             jnrA             = jjnr[jidx];
1839             j_coord_offsetA  = DIM*jnrA;
1840
1841             /* load j atom coordinates */
1842             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1843                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1844                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1845
1846             /* Calculate displacement vector */
1847             dx00             = _mm_sub_pd(ix0,jx0);
1848             dy00             = _mm_sub_pd(iy0,jy0);
1849             dz00             = _mm_sub_pd(iz0,jz0);
1850             dx11             = _mm_sub_pd(ix1,jx1);
1851             dy11             = _mm_sub_pd(iy1,jy1);
1852             dz11             = _mm_sub_pd(iz1,jz1);
1853             dx12             = _mm_sub_pd(ix1,jx2);
1854             dy12             = _mm_sub_pd(iy1,jy2);
1855             dz12             = _mm_sub_pd(iz1,jz2);
1856             dx13             = _mm_sub_pd(ix1,jx3);
1857             dy13             = _mm_sub_pd(iy1,jy3);
1858             dz13             = _mm_sub_pd(iz1,jz3);
1859             dx21             = _mm_sub_pd(ix2,jx1);
1860             dy21             = _mm_sub_pd(iy2,jy1);
1861             dz21             = _mm_sub_pd(iz2,jz1);
1862             dx22             = _mm_sub_pd(ix2,jx2);
1863             dy22             = _mm_sub_pd(iy2,jy2);
1864             dz22             = _mm_sub_pd(iz2,jz2);
1865             dx23             = _mm_sub_pd(ix2,jx3);
1866             dy23             = _mm_sub_pd(iy2,jy3);
1867             dz23             = _mm_sub_pd(iz2,jz3);
1868             dx31             = _mm_sub_pd(ix3,jx1);
1869             dy31             = _mm_sub_pd(iy3,jy1);
1870             dz31             = _mm_sub_pd(iz3,jz1);
1871             dx32             = _mm_sub_pd(ix3,jx2);
1872             dy32             = _mm_sub_pd(iy3,jy2);
1873             dz32             = _mm_sub_pd(iz3,jz2);
1874             dx33             = _mm_sub_pd(ix3,jx3);
1875             dy33             = _mm_sub_pd(iy3,jy3);
1876             dz33             = _mm_sub_pd(iz3,jz3);
1877
1878             /* Calculate squared distance and things based on it */
1879             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1880             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1881             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1882             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1883             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1884             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1885             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1886             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1887             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1888             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1889
1890             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1891             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1892             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1893             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1894             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1895             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1896             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1897             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1898             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1899             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1900
1901             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1902             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1903             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1904             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1905             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1906             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1907             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1908             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1909             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1910             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1911
1912             fjx0             = _mm_setzero_pd();
1913             fjy0             = _mm_setzero_pd();
1914             fjz0             = _mm_setzero_pd();
1915             fjx1             = _mm_setzero_pd();
1916             fjy1             = _mm_setzero_pd();
1917             fjz1             = _mm_setzero_pd();
1918             fjx2             = _mm_setzero_pd();
1919             fjy2             = _mm_setzero_pd();
1920             fjz2             = _mm_setzero_pd();
1921             fjx3             = _mm_setzero_pd();
1922             fjy3             = _mm_setzero_pd();
1923             fjz3             = _mm_setzero_pd();
1924
1925             /**************************
1926              * CALCULATE INTERACTIONS *
1927              **************************/
1928
1929             if (gmx_mm_any_lt(rsq00,rcutoff2))
1930             {
1931
1932             r00              = _mm_mul_pd(rsq00,rinv00);
1933
1934             /* LENNARD-JONES DISPERSION/REPULSION */
1935
1936             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1937             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1938             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1939             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
1940             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1941
1942             d                = _mm_sub_pd(r00,rswitch);
1943             d                = _mm_max_pd(d,_mm_setzero_pd());
1944             d2               = _mm_mul_pd(d,d);
1945             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)))))));
1946
1947             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
1948
1949             /* Evaluate switch function */
1950             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1951             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1952             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1953
1954             fscal            = fvdw;
1955
1956             fscal            = _mm_and_pd(fscal,cutoff_mask);
1957
1958             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1959
1960             /* Calculate temporary vectorial force */
1961             tx               = _mm_mul_pd(fscal,dx00);
1962             ty               = _mm_mul_pd(fscal,dy00);
1963             tz               = _mm_mul_pd(fscal,dz00);
1964
1965             /* Update vectorial force */
1966             fix0             = _mm_add_pd(fix0,tx);
1967             fiy0             = _mm_add_pd(fiy0,ty);
1968             fiz0             = _mm_add_pd(fiz0,tz);
1969
1970             fjx0             = _mm_add_pd(fjx0,tx);
1971             fjy0             = _mm_add_pd(fjy0,ty);
1972             fjz0             = _mm_add_pd(fjz0,tz);
1973
1974             }
1975
1976             /**************************
1977              * CALCULATE INTERACTIONS *
1978              **************************/
1979
1980             if (gmx_mm_any_lt(rsq11,rcutoff2))
1981             {
1982
1983             /* REACTION-FIELD ELECTROSTATICS */
1984             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1985
1986             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1987
1988             fscal            = felec;
1989
1990             fscal            = _mm_and_pd(fscal,cutoff_mask);
1991
1992             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1993
1994             /* Calculate temporary vectorial force */
1995             tx               = _mm_mul_pd(fscal,dx11);
1996             ty               = _mm_mul_pd(fscal,dy11);
1997             tz               = _mm_mul_pd(fscal,dz11);
1998
1999             /* Update vectorial force */
2000             fix1             = _mm_add_pd(fix1,tx);
2001             fiy1             = _mm_add_pd(fiy1,ty);
2002             fiz1             = _mm_add_pd(fiz1,tz);
2003
2004             fjx1             = _mm_add_pd(fjx1,tx);
2005             fjy1             = _mm_add_pd(fjy1,ty);
2006             fjz1             = _mm_add_pd(fjz1,tz);
2007
2008             }
2009
2010             /**************************
2011              * CALCULATE INTERACTIONS *
2012              **************************/
2013
2014             if (gmx_mm_any_lt(rsq12,rcutoff2))
2015             {
2016
2017             /* REACTION-FIELD ELECTROSTATICS */
2018             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
2019
2020             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2021
2022             fscal            = felec;
2023
2024             fscal            = _mm_and_pd(fscal,cutoff_mask);
2025
2026             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2027
2028             /* Calculate temporary vectorial force */
2029             tx               = _mm_mul_pd(fscal,dx12);
2030             ty               = _mm_mul_pd(fscal,dy12);
2031             tz               = _mm_mul_pd(fscal,dz12);
2032
2033             /* Update vectorial force */
2034             fix1             = _mm_add_pd(fix1,tx);
2035             fiy1             = _mm_add_pd(fiy1,ty);
2036             fiz1             = _mm_add_pd(fiz1,tz);
2037
2038             fjx2             = _mm_add_pd(fjx2,tx);
2039             fjy2             = _mm_add_pd(fjy2,ty);
2040             fjz2             = _mm_add_pd(fjz2,tz);
2041
2042             }
2043
2044             /**************************
2045              * CALCULATE INTERACTIONS *
2046              **************************/
2047
2048             if (gmx_mm_any_lt(rsq13,rcutoff2))
2049             {
2050
2051             /* REACTION-FIELD ELECTROSTATICS */
2052             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
2053
2054             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
2055
2056             fscal            = felec;
2057
2058             fscal            = _mm_and_pd(fscal,cutoff_mask);
2059
2060             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2061
2062             /* Calculate temporary vectorial force */
2063             tx               = _mm_mul_pd(fscal,dx13);
2064             ty               = _mm_mul_pd(fscal,dy13);
2065             tz               = _mm_mul_pd(fscal,dz13);
2066
2067             /* Update vectorial force */
2068             fix1             = _mm_add_pd(fix1,tx);
2069             fiy1             = _mm_add_pd(fiy1,ty);
2070             fiz1             = _mm_add_pd(fiz1,tz);
2071
2072             fjx3             = _mm_add_pd(fjx3,tx);
2073             fjy3             = _mm_add_pd(fjy3,ty);
2074             fjz3             = _mm_add_pd(fjz3,tz);
2075
2076             }
2077
2078             /**************************
2079              * CALCULATE INTERACTIONS *
2080              **************************/
2081
2082             if (gmx_mm_any_lt(rsq21,rcutoff2))
2083             {
2084
2085             /* REACTION-FIELD ELECTROSTATICS */
2086             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2087
2088             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2089
2090             fscal            = felec;
2091
2092             fscal            = _mm_and_pd(fscal,cutoff_mask);
2093
2094             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2095
2096             /* Calculate temporary vectorial force */
2097             tx               = _mm_mul_pd(fscal,dx21);
2098             ty               = _mm_mul_pd(fscal,dy21);
2099             tz               = _mm_mul_pd(fscal,dz21);
2100
2101             /* Update vectorial force */
2102             fix2             = _mm_add_pd(fix2,tx);
2103             fiy2             = _mm_add_pd(fiy2,ty);
2104             fiz2             = _mm_add_pd(fiz2,tz);
2105
2106             fjx1             = _mm_add_pd(fjx1,tx);
2107             fjy1             = _mm_add_pd(fjy1,ty);
2108             fjz1             = _mm_add_pd(fjz1,tz);
2109
2110             }
2111
2112             /**************************
2113              * CALCULATE INTERACTIONS *
2114              **************************/
2115
2116             if (gmx_mm_any_lt(rsq22,rcutoff2))
2117             {
2118
2119             /* REACTION-FIELD ELECTROSTATICS */
2120             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2121
2122             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2123
2124             fscal            = felec;
2125
2126             fscal            = _mm_and_pd(fscal,cutoff_mask);
2127
2128             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2129
2130             /* Calculate temporary vectorial force */
2131             tx               = _mm_mul_pd(fscal,dx22);
2132             ty               = _mm_mul_pd(fscal,dy22);
2133             tz               = _mm_mul_pd(fscal,dz22);
2134
2135             /* Update vectorial force */
2136             fix2             = _mm_add_pd(fix2,tx);
2137             fiy2             = _mm_add_pd(fiy2,ty);
2138             fiz2             = _mm_add_pd(fiz2,tz);
2139
2140             fjx2             = _mm_add_pd(fjx2,tx);
2141             fjy2             = _mm_add_pd(fjy2,ty);
2142             fjz2             = _mm_add_pd(fjz2,tz);
2143
2144             }
2145
2146             /**************************
2147              * CALCULATE INTERACTIONS *
2148              **************************/
2149
2150             if (gmx_mm_any_lt(rsq23,rcutoff2))
2151             {
2152
2153             /* REACTION-FIELD ELECTROSTATICS */
2154             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
2155
2156             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2157
2158             fscal            = felec;
2159
2160             fscal            = _mm_and_pd(fscal,cutoff_mask);
2161
2162             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2163
2164             /* Calculate temporary vectorial force */
2165             tx               = _mm_mul_pd(fscal,dx23);
2166             ty               = _mm_mul_pd(fscal,dy23);
2167             tz               = _mm_mul_pd(fscal,dz23);
2168
2169             /* Update vectorial force */
2170             fix2             = _mm_add_pd(fix2,tx);
2171             fiy2             = _mm_add_pd(fiy2,ty);
2172             fiz2             = _mm_add_pd(fiz2,tz);
2173
2174             fjx3             = _mm_add_pd(fjx3,tx);
2175             fjy3             = _mm_add_pd(fjy3,ty);
2176             fjz3             = _mm_add_pd(fjz3,tz);
2177
2178             }
2179
2180             /**************************
2181              * CALCULATE INTERACTIONS *
2182              **************************/
2183
2184             if (gmx_mm_any_lt(rsq31,rcutoff2))
2185             {
2186
2187             /* REACTION-FIELD ELECTROSTATICS */
2188             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
2189
2190             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2191
2192             fscal            = felec;
2193
2194             fscal            = _mm_and_pd(fscal,cutoff_mask);
2195
2196             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2197
2198             /* Calculate temporary vectorial force */
2199             tx               = _mm_mul_pd(fscal,dx31);
2200             ty               = _mm_mul_pd(fscal,dy31);
2201             tz               = _mm_mul_pd(fscal,dz31);
2202
2203             /* Update vectorial force */
2204             fix3             = _mm_add_pd(fix3,tx);
2205             fiy3             = _mm_add_pd(fiy3,ty);
2206             fiz3             = _mm_add_pd(fiz3,tz);
2207
2208             fjx1             = _mm_add_pd(fjx1,tx);
2209             fjy1             = _mm_add_pd(fjy1,ty);
2210             fjz1             = _mm_add_pd(fjz1,tz);
2211
2212             }
2213
2214             /**************************
2215              * CALCULATE INTERACTIONS *
2216              **************************/
2217
2218             if (gmx_mm_any_lt(rsq32,rcutoff2))
2219             {
2220
2221             /* REACTION-FIELD ELECTROSTATICS */
2222             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
2223
2224             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2225
2226             fscal            = felec;
2227
2228             fscal            = _mm_and_pd(fscal,cutoff_mask);
2229
2230             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2231
2232             /* Calculate temporary vectorial force */
2233             tx               = _mm_mul_pd(fscal,dx32);
2234             ty               = _mm_mul_pd(fscal,dy32);
2235             tz               = _mm_mul_pd(fscal,dz32);
2236
2237             /* Update vectorial force */
2238             fix3             = _mm_add_pd(fix3,tx);
2239             fiy3             = _mm_add_pd(fiy3,ty);
2240             fiz3             = _mm_add_pd(fiz3,tz);
2241
2242             fjx2             = _mm_add_pd(fjx2,tx);
2243             fjy2             = _mm_add_pd(fjy2,ty);
2244             fjz2             = _mm_add_pd(fjz2,tz);
2245
2246             }
2247
2248             /**************************
2249              * CALCULATE INTERACTIONS *
2250              **************************/
2251
2252             if (gmx_mm_any_lt(rsq33,rcutoff2))
2253             {
2254
2255             /* REACTION-FIELD ELECTROSTATICS */
2256             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
2257
2258             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2259
2260             fscal            = felec;
2261
2262             fscal            = _mm_and_pd(fscal,cutoff_mask);
2263
2264             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2265
2266             /* Calculate temporary vectorial force */
2267             tx               = _mm_mul_pd(fscal,dx33);
2268             ty               = _mm_mul_pd(fscal,dy33);
2269             tz               = _mm_mul_pd(fscal,dz33);
2270
2271             /* Update vectorial force */
2272             fix3             = _mm_add_pd(fix3,tx);
2273             fiy3             = _mm_add_pd(fiy3,ty);
2274             fiz3             = _mm_add_pd(fiz3,tz);
2275
2276             fjx3             = _mm_add_pd(fjx3,tx);
2277             fjy3             = _mm_add_pd(fjy3,ty);
2278             fjz3             = _mm_add_pd(fjz3,tz);
2279
2280             }
2281
2282             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2283
2284             /* Inner loop uses 329 flops */
2285         }
2286
2287         /* End of innermost loop */
2288
2289         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2290                                               f+i_coord_offset,fshift+i_shift_offset);
2291
2292         /* Increment number of inner iterations */
2293         inneriter                  += j_index_end - j_index_start;
2294
2295         /* Outer loop uses 24 flops */
2296     }
2297
2298     /* Increment number of outer iterations */
2299     outeriter        += nri;
2300
2301     /* Update outer/inner flops */
2302
2303     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*329);
2304 }