src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_sse2_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_double.h"
  50 #include "kernelutil_x86_sse2_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_sse2_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_sse2_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 117     real             rswitch_scalar,d_scalar;
 118     __m128d          dummy_mask,cutoff_mask;
 119     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 120     __m128d          one     = _mm_set1_pd(1.0);
 121     __m128d          two     = _mm_set1_pd(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_pd(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     krf              = _mm_set1_pd(fr->ic->k_rf);
 136     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 137     crf              = _mm_set1_pd(fr->ic->c_rf);
 138     nvdwtype         = fr->ntype;
 139     vdwparam         = fr->nbfp;
 140     vdwtype          = mdatoms->typeA;
 141
 142     /* Setup water-specific parameters */
 143     inr              = nlist->iinr[0];
 144     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 145     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 146     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 147     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 148
 149     jq1              = _mm_set1_pd(charge[inr+1]);
 150     jq2              = _mm_set1_pd(charge[inr+2]);
 151     jq3              = _mm_set1_pd(charge[inr+3]);
 152     vdwjidx0A        = 2*vdwtype[inr+0];
 153     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 154     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 155     qq11             = _mm_mul_pd(iq1,jq1);
 156     qq12             = _mm_mul_pd(iq1,jq2);
 157     qq13             = _mm_mul_pd(iq1,jq3);
 158     qq21             = _mm_mul_pd(iq2,jq1);
 159     qq22             = _mm_mul_pd(iq2,jq2);
 160     qq23             = _mm_mul_pd(iq2,jq3);
 161     qq31             = _mm_mul_pd(iq3,jq1);
 162     qq32             = _mm_mul_pd(iq3,jq2);
 163     qq33             = _mm_mul_pd(iq3,jq3);
 164
 165     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 166     rcutoff_scalar   = fr->rcoulomb;
 167     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 168     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 169
 170     rswitch_scalar   = fr->rvdw_switch;
 171     rswitch          = _mm_set1_pd(rswitch_scalar);
 172     /* Setup switch parameters */
 173     d_scalar         = rcutoff_scalar-rswitch_scalar;
 174     d                = _mm_set1_pd(d_scalar);
 175     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 176     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 177     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 178     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 179     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 180     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 181
 182     /* Avoid stupid compiler warnings */
 183     jnrA = jnrB = 0;
 184     j_coord_offsetA = 0;
 185     j_coord_offsetB = 0;
 186
 187     outeriter        = 0;
 188     inneriter        = 0;
 189
 190     /* Start outer loop over neighborlists */
 191     for(iidx=0; iidx<nri; iidx++)
 192     {
 193         /* Load shift vector for this list */
 194         i_shift_offset   = DIM*shiftidx[iidx];
 195
 196         /* Load limits for loop over neighbors */
 197         j_index_start    = jindex[iidx];
 198         j_index_end      = jindex[iidx+1];
 199
 200         /* Get outer coordinate index */
 201         inr              = iinr[iidx];
 202         i_coord_offset   = DIM*inr;
 203
 204         /* Load i particle coords and add shift vector */
 205         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 206                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 207
 208         fix0             = _mm_setzero_pd();
 209         fiy0             = _mm_setzero_pd();
 210         fiz0             = _mm_setzero_pd();
 211         fix1             = _mm_setzero_pd();
 212         fiy1             = _mm_setzero_pd();
 213         fiz1             = _mm_setzero_pd();
 214         fix2             = _mm_setzero_pd();
 215         fiy2             = _mm_setzero_pd();
 216         fiz2             = _mm_setzero_pd();
 217         fix3             = _mm_setzero_pd();
 218         fiy3             = _mm_setzero_pd();
 219         fiz3             = _mm_setzero_pd();
 220
 221         /* Reset potential sums */
 222         velecsum         = _mm_setzero_pd();
 223         vvdwsum          = _mm_setzero_pd();
 224
 225         /* Start inner kernel loop */
 226         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 227         {
 228
 229             /* Get j neighbor index, and coordinate index */
 230             jnrA             = jjnr[jidx];
 231             jnrB             = jjnr[jidx+1];
 232             j_coord_offsetA  = DIM*jnrA;
 233             j_coord_offsetB  = DIM*jnrB;
 234
 235             /* load j atom coordinates */
 236             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 237                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 238                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 239
 240             /* Calculate displacement vector */
 241             dx00             = _mm_sub_pd(ix0,jx0);
 242             dy00             = _mm_sub_pd(iy0,jy0);
 243             dz00             = _mm_sub_pd(iz0,jz0);
 244             dx11             = _mm_sub_pd(ix1,jx1);
 245             dy11             = _mm_sub_pd(iy1,jy1);
 246             dz11             = _mm_sub_pd(iz1,jz1);
 247             dx12             = _mm_sub_pd(ix1,jx2);
 248             dy12             = _mm_sub_pd(iy1,jy2);
 249             dz12             = _mm_sub_pd(iz1,jz2);
 250             dx13             = _mm_sub_pd(ix1,jx3);
 251             dy13             = _mm_sub_pd(iy1,jy3);
 252             dz13             = _mm_sub_pd(iz1,jz3);
 253             dx21             = _mm_sub_pd(ix2,jx1);
 254             dy21             = _mm_sub_pd(iy2,jy1);
 255             dz21             = _mm_sub_pd(iz2,jz1);
 256             dx22             = _mm_sub_pd(ix2,jx2);
 257             dy22             = _mm_sub_pd(iy2,jy2);
 258             dz22             = _mm_sub_pd(iz2,jz2);
 259             dx23             = _mm_sub_pd(ix2,jx3);
 260             dy23             = _mm_sub_pd(iy2,jy3);
 261             dz23             = _mm_sub_pd(iz2,jz3);
 262             dx31             = _mm_sub_pd(ix3,jx1);
 263             dy31             = _mm_sub_pd(iy3,jy1);
 264             dz31             = _mm_sub_pd(iz3,jz1);
 265             dx32             = _mm_sub_pd(ix3,jx2);
 266             dy32             = _mm_sub_pd(iy3,jy2);
 267             dz32             = _mm_sub_pd(iz3,jz2);
 268             dx33             = _mm_sub_pd(ix3,jx3);
 269             dy33             = _mm_sub_pd(iy3,jy3);
 270             dz33             = _mm_sub_pd(iz3,jz3);
 271
 272             /* Calculate squared distance and things based on it */
 273             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 274             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 275             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 276             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 277             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 278             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 279             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 280             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 281             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 282             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 283
 284             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 285             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 286             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 287             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 288             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 289             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 290             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 291             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 292             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 293             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 294
 295             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 296             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 297             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 298             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 299             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 300             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 301             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 302             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 303             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 304             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 305
 306             fjx0             = _mm_setzero_pd();
 307             fjy0             = _mm_setzero_pd();
 308             fjz0             = _mm_setzero_pd();
 309             fjx1             = _mm_setzero_pd();
 310             fjy1             = _mm_setzero_pd();
 311             fjz1             = _mm_setzero_pd();
 312             fjx2             = _mm_setzero_pd();
 313             fjy2             = _mm_setzero_pd();
 314             fjz2             = _mm_setzero_pd();
 315             fjx3             = _mm_setzero_pd();
 316             fjy3             = _mm_setzero_pd();
 317             fjz3             = _mm_setzero_pd();
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             if (gmx_mm_any_lt(rsq00,rcutoff2))
 324             {
 325
 326             r00              = _mm_mul_pd(rsq00,rinv00);
 327
 328             /* LENNARD-JONES DISPERSION/REPULSION */
 329
 330             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 331             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 332             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 333             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 334             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 335
 336             d                = _mm_sub_pd(r00,rswitch);
 337             d                = _mm_max_pd(d,_mm_setzero_pd());
 338             d2               = _mm_mul_pd(d,d);
 339             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)))))));
 340
 341             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
 342
 343             /* Evaluate switch function */
 344             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 345             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 346             vvdw             = _mm_mul_pd(vvdw,sw);
 347             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 351             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 352
 353             fscal            = fvdw;
 354
 355             fscal            = _mm_and_pd(fscal,cutoff_mask);
 356
 357             /* Calculate temporary vectorial force */
 358             tx               = _mm_mul_pd(fscal,dx00);
 359             ty               = _mm_mul_pd(fscal,dy00);
 360             tz               = _mm_mul_pd(fscal,dz00);
 361
 362             /* Update vectorial force */
 363             fix0             = _mm_add_pd(fix0,tx);
 364             fiy0             = _mm_add_pd(fiy0,ty);
 365             fiz0             = _mm_add_pd(fiz0,tz);
 366
 367             fjx0             = _mm_add_pd(fjx0,tx);
 368             fjy0             = _mm_add_pd(fjy0,ty);
 369             fjz0             = _mm_add_pd(fjz0,tz);
 370
 371             }
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             if (gmx_mm_any_lt(rsq11,rcutoff2))
 378             {
 379
 380             /* REACTION-FIELD ELECTROSTATICS */
 381             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 382             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 383
 384             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velec            = _mm_and_pd(velec,cutoff_mask);
 388             velecsum         = _mm_add_pd(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392             fscal            = _mm_and_pd(fscal,cutoff_mask);
 393
 394             /* Calculate temporary vectorial force */
 395             tx               = _mm_mul_pd(fscal,dx11);
 396             ty               = _mm_mul_pd(fscal,dy11);
 397             tz               = _mm_mul_pd(fscal,dz11);
 398
 399             /* Update vectorial force */
 400             fix1             = _mm_add_pd(fix1,tx);
 401             fiy1             = _mm_add_pd(fiy1,ty);
 402             fiz1             = _mm_add_pd(fiz1,tz);
 403
 404             fjx1             = _mm_add_pd(fjx1,tx);
 405             fjy1             = _mm_add_pd(fjy1,ty);
 406             fjz1             = _mm_add_pd(fjz1,tz);
 407
 408             }
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             if (gmx_mm_any_lt(rsq12,rcutoff2))
 415             {
 416
 417             /* REACTION-FIELD ELECTROSTATICS */
 418             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 419             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 420
 421             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 422
 423             /* Update potential sum for this i atom from the interaction with this j atom. */
 424             velec            = _mm_and_pd(velec,cutoff_mask);
 425             velecsum         = _mm_add_pd(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             fscal            = _mm_and_pd(fscal,cutoff_mask);
 430
 431             /* Calculate temporary vectorial force */
 432             tx               = _mm_mul_pd(fscal,dx12);
 433             ty               = _mm_mul_pd(fscal,dy12);
 434             tz               = _mm_mul_pd(fscal,dz12);
 435
 436             /* Update vectorial force */
 437             fix1             = _mm_add_pd(fix1,tx);
 438             fiy1             = _mm_add_pd(fiy1,ty);
 439             fiz1             = _mm_add_pd(fiz1,tz);
 440
 441             fjx2             = _mm_add_pd(fjx2,tx);
 442             fjy2             = _mm_add_pd(fjy2,ty);
 443             fjz2             = _mm_add_pd(fjz2,tz);
 444
 445             }
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             if (gmx_mm_any_lt(rsq13,rcutoff2))
 452             {
 453
 454             /* REACTION-FIELD ELECTROSTATICS */
 455             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 456             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 457
 458             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velec            = _mm_and_pd(velec,cutoff_mask);
 462             velecsum         = _mm_add_pd(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466             fscal            = _mm_and_pd(fscal,cutoff_mask);
 467
 468             /* Calculate temporary vectorial force */
 469             tx               = _mm_mul_pd(fscal,dx13);
 470             ty               = _mm_mul_pd(fscal,dy13);
 471             tz               = _mm_mul_pd(fscal,dz13);
 472
 473             /* Update vectorial force */
 474             fix1             = _mm_add_pd(fix1,tx);
 475             fiy1             = _mm_add_pd(fiy1,ty);
 476             fiz1             = _mm_add_pd(fiz1,tz);
 477
 478             fjx3             = _mm_add_pd(fjx3,tx);
 479             fjy3             = _mm_add_pd(fjy3,ty);
 480             fjz3             = _mm_add_pd(fjz3,tz);
 481
 482             }
 483
 484             /**************************
 485              * CALCULATE INTERACTIONS *
 486              **************************/
 487
 488             if (gmx_mm_any_lt(rsq21,rcutoff2))
 489             {
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 493             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 494
 495             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velec            = _mm_and_pd(velec,cutoff_mask);
 499             velecsum         = _mm_add_pd(velecsum,velec);
 500
 501             fscal            = felec;
 502
 503             fscal            = _mm_and_pd(fscal,cutoff_mask);
 504
 505             /* Calculate temporary vectorial force */
 506             tx               = _mm_mul_pd(fscal,dx21);
 507             ty               = _mm_mul_pd(fscal,dy21);
 508             tz               = _mm_mul_pd(fscal,dz21);
 509
 510             /* Update vectorial force */
 511             fix2             = _mm_add_pd(fix2,tx);
 512             fiy2             = _mm_add_pd(fiy2,ty);
 513             fiz2             = _mm_add_pd(fiz2,tz);
 514
 515             fjx1             = _mm_add_pd(fjx1,tx);
 516             fjy1             = _mm_add_pd(fjy1,ty);
 517             fjz1             = _mm_add_pd(fjz1,tz);
 518
 519             }
 520
 521             /**************************
 522              * CALCULATE INTERACTIONS *
 523              **************************/
 524
 525             if (gmx_mm_any_lt(rsq22,rcutoff2))
 526             {
 527
 528             /* REACTION-FIELD ELECTROSTATICS */
 529             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 530             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 531
 532             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velec            = _mm_and_pd(velec,cutoff_mask);
 536             velecsum         = _mm_add_pd(velecsum,velec);
 537
 538             fscal            = felec;
 539
 540             fscal            = _mm_and_pd(fscal,cutoff_mask);
 541
 542             /* Calculate temporary vectorial force */
 543             tx               = _mm_mul_pd(fscal,dx22);
 544             ty               = _mm_mul_pd(fscal,dy22);
 545             tz               = _mm_mul_pd(fscal,dz22);
 546
 547             /* Update vectorial force */
 548             fix2             = _mm_add_pd(fix2,tx);
 549             fiy2             = _mm_add_pd(fiy2,ty);
 550             fiz2             = _mm_add_pd(fiz2,tz);
 551
 552             fjx2             = _mm_add_pd(fjx2,tx);
 553             fjy2             = _mm_add_pd(fjy2,ty);
 554             fjz2             = _mm_add_pd(fjz2,tz);
 555
 556             }
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             if (gmx_mm_any_lt(rsq23,rcutoff2))
 563             {
 564
 565             /* REACTION-FIELD ELECTROSTATICS */
 566             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 567             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 568
 569             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 570
 571             /* Update potential sum for this i atom from the interaction with this j atom. */
 572             velec            = _mm_and_pd(velec,cutoff_mask);
 573             velecsum         = _mm_add_pd(velecsum,velec);
 574
 575             fscal            = felec;
 576
 577             fscal            = _mm_and_pd(fscal,cutoff_mask);
 578
 579             /* Calculate temporary vectorial force */
 580             tx               = _mm_mul_pd(fscal,dx23);
 581             ty               = _mm_mul_pd(fscal,dy23);
 582             tz               = _mm_mul_pd(fscal,dz23);
 583
 584             /* Update vectorial force */
 585             fix2             = _mm_add_pd(fix2,tx);
 586             fiy2             = _mm_add_pd(fiy2,ty);
 587             fiz2             = _mm_add_pd(fiz2,tz);
 588
 589             fjx3             = _mm_add_pd(fjx3,tx);
 590             fjy3             = _mm_add_pd(fjy3,ty);
 591             fjz3             = _mm_add_pd(fjz3,tz);
 592
 593             }
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             if (gmx_mm_any_lt(rsq31,rcutoff2))
 600             {
 601
 602             /* REACTION-FIELD ELECTROSTATICS */
 603             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 604             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 605
 606             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 607
 608             /* Update potential sum for this i atom from the interaction with this j atom. */
 609             velec            = _mm_and_pd(velec,cutoff_mask);
 610             velecsum         = _mm_add_pd(velecsum,velec);
 611
 612             fscal            = felec;
 613
 614             fscal            = _mm_and_pd(fscal,cutoff_mask);
 615
 616             /* Calculate temporary vectorial force */
 617             tx               = _mm_mul_pd(fscal,dx31);
 618             ty               = _mm_mul_pd(fscal,dy31);
 619             tz               = _mm_mul_pd(fscal,dz31);
 620
 621             /* Update vectorial force */
 622             fix3             = _mm_add_pd(fix3,tx);
 623             fiy3             = _mm_add_pd(fiy3,ty);
 624             fiz3             = _mm_add_pd(fiz3,tz);
 625
 626             fjx1             = _mm_add_pd(fjx1,tx);
 627             fjy1             = _mm_add_pd(fjy1,ty);
 628             fjz1             = _mm_add_pd(fjz1,tz);
 629
 630             }
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             if (gmx_mm_any_lt(rsq32,rcutoff2))
 637             {
 638
 639             /* REACTION-FIELD ELECTROSTATICS */
 640             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 641             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 642
 643             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 644
 645             /* Update potential sum for this i atom from the interaction with this j atom. */
 646             velec            = _mm_and_pd(velec,cutoff_mask);
 647             velecsum         = _mm_add_pd(velecsum,velec);
 648
 649             fscal            = felec;
 650
 651             fscal            = _mm_and_pd(fscal,cutoff_mask);
 652
 653             /* Calculate temporary vectorial force */
 654             tx               = _mm_mul_pd(fscal,dx32);
 655             ty               = _mm_mul_pd(fscal,dy32);
 656             tz               = _mm_mul_pd(fscal,dz32);
 657
 658             /* Update vectorial force */
 659             fix3             = _mm_add_pd(fix3,tx);
 660             fiy3             = _mm_add_pd(fiy3,ty);
 661             fiz3             = _mm_add_pd(fiz3,tz);
 662
 663             fjx2             = _mm_add_pd(fjx2,tx);
 664             fjy2             = _mm_add_pd(fjy2,ty);
 665             fjz2             = _mm_add_pd(fjz2,tz);
 666
 667             }
 668
 669             /**************************
 670              * CALCULATE INTERACTIONS *
 671              **************************/
 672
 673             if (gmx_mm_any_lt(rsq33,rcutoff2))
 674             {
 675
 676             /* REACTION-FIELD ELECTROSTATICS */
 677             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 678             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 679
 680             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 681
 682             /* Update potential sum for this i atom from the interaction with this j atom. */
 683             velec            = _mm_and_pd(velec,cutoff_mask);
 684             velecsum         = _mm_add_pd(velecsum,velec);
 685
 686             fscal            = felec;
 687
 688             fscal            = _mm_and_pd(fscal,cutoff_mask);
 689
 690             /* Calculate temporary vectorial force */
 691             tx               = _mm_mul_pd(fscal,dx33);
 692             ty               = _mm_mul_pd(fscal,dy33);
 693             tz               = _mm_mul_pd(fscal,dz33);
 694
 695             /* Update vectorial force */
 696             fix3             = _mm_add_pd(fix3,tx);
 697             fiy3             = _mm_add_pd(fiy3,ty);
 698             fiz3             = _mm_add_pd(fiz3,tz);
 699
 700             fjx3             = _mm_add_pd(fjx3,tx);
 701             fjy3             = _mm_add_pd(fjy3,ty);
 702             fjz3             = _mm_add_pd(fjz3,tz);
 703
 704             }
 705
 706             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);
 707
 708             /* Inner loop uses 386 flops */
 709         }
 710
 711         if(jidx<j_index_end)
 712         {
 713
 714             jnrA             = jjnr[jidx];
 715             j_coord_offsetA  = DIM*jnrA;
 716
 717             /* load j atom coordinates */
 718             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 719                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 720                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 721
 722             /* Calculate displacement vector */
 723             dx00             = _mm_sub_pd(ix0,jx0);
 724             dy00             = _mm_sub_pd(iy0,jy0);
 725             dz00             = _mm_sub_pd(iz0,jz0);
 726             dx11             = _mm_sub_pd(ix1,jx1);
 727             dy11             = _mm_sub_pd(iy1,jy1);
 728             dz11             = _mm_sub_pd(iz1,jz1);
 729             dx12             = _mm_sub_pd(ix1,jx2);
 730             dy12             = _mm_sub_pd(iy1,jy2);
 731             dz12             = _mm_sub_pd(iz1,jz2);
 732             dx13             = _mm_sub_pd(ix1,jx3);
 733             dy13             = _mm_sub_pd(iy1,jy3);
 734             dz13             = _mm_sub_pd(iz1,jz3);
 735             dx21             = _mm_sub_pd(ix2,jx1);
 736             dy21             = _mm_sub_pd(iy2,jy1);
 737             dz21             = _mm_sub_pd(iz2,jz1);
 738             dx22             = _mm_sub_pd(ix2,jx2);
 739             dy22             = _mm_sub_pd(iy2,jy2);
 740             dz22             = _mm_sub_pd(iz2,jz2);
 741             dx23             = _mm_sub_pd(ix2,jx3);
 742             dy23             = _mm_sub_pd(iy2,jy3);
 743             dz23             = _mm_sub_pd(iz2,jz3);
 744             dx31             = _mm_sub_pd(ix3,jx1);
 745             dy31             = _mm_sub_pd(iy3,jy1);
 746             dz31             = _mm_sub_pd(iz3,jz1);
 747             dx32             = _mm_sub_pd(ix3,jx2);
 748             dy32             = _mm_sub_pd(iy3,jy2);
 749             dz32             = _mm_sub_pd(iz3,jz2);
 750             dx33             = _mm_sub_pd(ix3,jx3);
 751             dy33             = _mm_sub_pd(iy3,jy3);
 752             dz33             = _mm_sub_pd(iz3,jz3);
 753
 754             /* Calculate squared distance and things based on it */
 755             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 756             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 757             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 758             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 759             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 760             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 761             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 762             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 763             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 764             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 765
 766             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 767             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 768             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 769             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 770             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 771             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 772             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 773             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 774             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 775             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 776
 777             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 778             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 779             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 780             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 781             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 782             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 783             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 784             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 785             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 786             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 787
 788             fjx0             = _mm_setzero_pd();
 789             fjy0             = _mm_setzero_pd();
 790             fjz0             = _mm_setzero_pd();
 791             fjx1             = _mm_setzero_pd();
 792             fjy1             = _mm_setzero_pd();
 793             fjz1             = _mm_setzero_pd();
 794             fjx2             = _mm_setzero_pd();
 795             fjy2             = _mm_setzero_pd();
 796             fjz2             = _mm_setzero_pd();
 797             fjx3             = _mm_setzero_pd();
 798             fjy3             = _mm_setzero_pd();
 799             fjz3             = _mm_setzero_pd();
 800
 801             /**************************
 802              * CALCULATE INTERACTIONS *
 803              **************************/
 804
 805             if (gmx_mm_any_lt(rsq00,rcutoff2))
 806             {
 807
 808             r00              = _mm_mul_pd(rsq00,rinv00);
 809
 810             /* LENNARD-JONES DISPERSION/REPULSION */
 811
 812             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 813             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 814             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 815             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 816             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 817
 818             d                = _mm_sub_pd(r00,rswitch);
 819             d                = _mm_max_pd(d,_mm_setzero_pd());
 820             d2               = _mm_mul_pd(d,d);
 821             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)))))));
 822
 823             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
 824
 825             /* Evaluate switch function */
 826             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 827             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 828             vvdw             = _mm_mul_pd(vvdw,sw);
 829             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 830
 831             /* Update potential sum for this i atom from the interaction with this j atom. */
 832             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 833             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 834             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 835
 836             fscal            = fvdw;
 837
 838             fscal            = _mm_and_pd(fscal,cutoff_mask);
 839
 840             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 841
 842             /* Calculate temporary vectorial force */
 843             tx               = _mm_mul_pd(fscal,dx00);
 844             ty               = _mm_mul_pd(fscal,dy00);
 845             tz               = _mm_mul_pd(fscal,dz00);
 846
 847             /* Update vectorial force */
 848             fix0             = _mm_add_pd(fix0,tx);
 849             fiy0             = _mm_add_pd(fiy0,ty);
 850             fiz0             = _mm_add_pd(fiz0,tz);
 851
 852             fjx0             = _mm_add_pd(fjx0,tx);
 853             fjy0             = _mm_add_pd(fjy0,ty);
 854             fjz0             = _mm_add_pd(fjz0,tz);
 855
 856             }
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             if (gmx_mm_any_lt(rsq11,rcutoff2))
 863             {
 864
 865             /* REACTION-FIELD ELECTROSTATICS */
 866             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 867             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 868
 869             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 870
 871             /* Update potential sum for this i atom from the interaction with this j atom. */
 872             velec            = _mm_and_pd(velec,cutoff_mask);
 873             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 874             velecsum         = _mm_add_pd(velecsum,velec);
 875
 876             fscal            = felec;
 877
 878             fscal            = _mm_and_pd(fscal,cutoff_mask);
 879
 880             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 881
 882             /* Calculate temporary vectorial force */
 883             tx               = _mm_mul_pd(fscal,dx11);
 884             ty               = _mm_mul_pd(fscal,dy11);
 885             tz               = _mm_mul_pd(fscal,dz11);
 886
 887             /* Update vectorial force */
 888             fix1             = _mm_add_pd(fix1,tx);
 889             fiy1             = _mm_add_pd(fiy1,ty);
 890             fiz1             = _mm_add_pd(fiz1,tz);
 891
 892             fjx1             = _mm_add_pd(fjx1,tx);
 893             fjy1             = _mm_add_pd(fjy1,ty);
 894             fjz1             = _mm_add_pd(fjz1,tz);
 895
 896             }
 897
 898             /**************************
 899              * CALCULATE INTERACTIONS *
 900              **************************/
 901
 902             if (gmx_mm_any_lt(rsq12,rcutoff2))
 903             {
 904
 905             /* REACTION-FIELD ELECTROSTATICS */
 906             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 907             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 908
 909             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 910
 911             /* Update potential sum for this i atom from the interaction with this j atom. */
 912             velec            = _mm_and_pd(velec,cutoff_mask);
 913             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 914             velecsum         = _mm_add_pd(velecsum,velec);
 915
 916             fscal            = felec;
 917
 918             fscal            = _mm_and_pd(fscal,cutoff_mask);
 919
 920             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 921
 922             /* Calculate temporary vectorial force */
 923             tx               = _mm_mul_pd(fscal,dx12);
 924             ty               = _mm_mul_pd(fscal,dy12);
 925             tz               = _mm_mul_pd(fscal,dz12);
 926
 927             /* Update vectorial force */
 928             fix1             = _mm_add_pd(fix1,tx);
 929             fiy1             = _mm_add_pd(fiy1,ty);
 930             fiz1             = _mm_add_pd(fiz1,tz);
 931
 932             fjx2             = _mm_add_pd(fjx2,tx);
 933             fjy2             = _mm_add_pd(fjy2,ty);
 934             fjz2             = _mm_add_pd(fjz2,tz);
 935
 936             }
 937
 938             /**************************
 939              * CALCULATE INTERACTIONS *
 940              **************************/
 941
 942             if (gmx_mm_any_lt(rsq13,rcutoff2))
 943             {
 944
 945             /* REACTION-FIELD ELECTROSTATICS */
 946             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 947             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 948
 949             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 950
 951             /* Update potential sum for this i atom from the interaction with this j atom. */
 952             velec            = _mm_and_pd(velec,cutoff_mask);
 953             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 954             velecsum         = _mm_add_pd(velecsum,velec);
 955
 956             fscal            = felec;
 957
 958             fscal            = _mm_and_pd(fscal,cutoff_mask);
 959
 960             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 961
 962             /* Calculate temporary vectorial force */
 963             tx               = _mm_mul_pd(fscal,dx13);
 964             ty               = _mm_mul_pd(fscal,dy13);
 965             tz               = _mm_mul_pd(fscal,dz13);
 966
 967             /* Update vectorial force */
 968             fix1             = _mm_add_pd(fix1,tx);
 969             fiy1             = _mm_add_pd(fiy1,ty);
 970             fiz1             = _mm_add_pd(fiz1,tz);
 971
 972             fjx3             = _mm_add_pd(fjx3,tx);
 973             fjy3             = _mm_add_pd(fjy3,ty);
 974             fjz3             = _mm_add_pd(fjz3,tz);
 975
 976             }
 977
 978             /**************************
 979              * CALCULATE INTERACTIONS *
 980              **************************/
 981
 982             if (gmx_mm_any_lt(rsq21,rcutoff2))
 983             {
 984
 985             /* REACTION-FIELD ELECTROSTATICS */
 986             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 987             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 988
 989             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 990
 991             /* Update potential sum for this i atom from the interaction with this j atom. */
 992             velec            = _mm_and_pd(velec,cutoff_mask);
 993             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 994             velecsum         = _mm_add_pd(velecsum,velec);
 995
 996             fscal            = felec;
 997
 998             fscal            = _mm_and_pd(fscal,cutoff_mask);
 999
1000             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1001
1002             /* Calculate temporary vectorial force */
1003             tx               = _mm_mul_pd(fscal,dx21);
1004             ty               = _mm_mul_pd(fscal,dy21);
1005             tz               = _mm_mul_pd(fscal,dz21);
1006
1007             /* Update vectorial force */
1008             fix2             = _mm_add_pd(fix2,tx);
1009             fiy2             = _mm_add_pd(fiy2,ty);
1010             fiz2             = _mm_add_pd(fiz2,tz);
1011
1012             fjx1             = _mm_add_pd(fjx1,tx);
1013             fjy1             = _mm_add_pd(fjy1,ty);
1014             fjz1             = _mm_add_pd(fjz1,tz);
1015
1016             }
1017
1018             /**************************
1019              * CALCULATE INTERACTIONS *
1020              **************************/
1021
1022             if (gmx_mm_any_lt(rsq22,rcutoff2))
1023             {
1024
1025             /* REACTION-FIELD ELECTROSTATICS */
1026             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1027             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1028
1029             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1030
1031             /* Update potential sum for this i atom from the interaction with this j atom. */
1032             velec            = _mm_and_pd(velec,cutoff_mask);
1033             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1034             velecsum         = _mm_add_pd(velecsum,velec);
1035
1036             fscal            = felec;
1037
1038             fscal            = _mm_and_pd(fscal,cutoff_mask);
1039
1040             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1041
1042             /* Calculate temporary vectorial force */
1043             tx               = _mm_mul_pd(fscal,dx22);
1044             ty               = _mm_mul_pd(fscal,dy22);
1045             tz               = _mm_mul_pd(fscal,dz22);
1046
1047             /* Update vectorial force */
1048             fix2             = _mm_add_pd(fix2,tx);
1049             fiy2             = _mm_add_pd(fiy2,ty);
1050             fiz2             = _mm_add_pd(fiz2,tz);
1051
1052             fjx2             = _mm_add_pd(fjx2,tx);
1053             fjy2             = _mm_add_pd(fjy2,ty);
1054             fjz2             = _mm_add_pd(fjz2,tz);
1055
1056             }
1057
1058             /**************************
1059              * CALCULATE INTERACTIONS *
1060              **************************/
1061
1062             if (gmx_mm_any_lt(rsq23,rcutoff2))
1063             {
1064
1065             /* REACTION-FIELD ELECTROSTATICS */
1066             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
1067             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1068
1069             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1070
1071             /* Update potential sum for this i atom from the interaction with this j atom. */
1072             velec            = _mm_and_pd(velec,cutoff_mask);
1073             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1074             velecsum         = _mm_add_pd(velecsum,velec);
1075
1076             fscal            = felec;
1077
1078             fscal            = _mm_and_pd(fscal,cutoff_mask);
1079
1080             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1081
1082             /* Calculate temporary vectorial force */
1083             tx               = _mm_mul_pd(fscal,dx23);
1084             ty               = _mm_mul_pd(fscal,dy23);
1085             tz               = _mm_mul_pd(fscal,dz23);
1086
1087             /* Update vectorial force */
1088             fix2             = _mm_add_pd(fix2,tx);
1089             fiy2             = _mm_add_pd(fiy2,ty);
1090             fiz2             = _mm_add_pd(fiz2,tz);
1091
1092             fjx3             = _mm_add_pd(fjx3,tx);
1093             fjy3             = _mm_add_pd(fjy3,ty);
1094             fjz3             = _mm_add_pd(fjz3,tz);
1095
1096             }
1097
1098             /**************************
1099              * CALCULATE INTERACTIONS *
1100              **************************/
1101
1102             if (gmx_mm_any_lt(rsq31,rcutoff2))
1103             {
1104
1105             /* REACTION-FIELD ELECTROSTATICS */
1106             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
1107             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1108
1109             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1110
1111             /* Update potential sum for this i atom from the interaction with this j atom. */
1112             velec            = _mm_and_pd(velec,cutoff_mask);
1113             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1114             velecsum         = _mm_add_pd(velecsum,velec);
1115
1116             fscal            = felec;
1117
1118             fscal            = _mm_and_pd(fscal,cutoff_mask);
1119
1120             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1121
1122             /* Calculate temporary vectorial force */
1123             tx               = _mm_mul_pd(fscal,dx31);
1124             ty               = _mm_mul_pd(fscal,dy31);
1125             tz               = _mm_mul_pd(fscal,dz31);
1126
1127             /* Update vectorial force */
1128             fix3             = _mm_add_pd(fix3,tx);
1129             fiy3             = _mm_add_pd(fiy3,ty);
1130             fiz3             = _mm_add_pd(fiz3,tz);
1131
1132             fjx1             = _mm_add_pd(fjx1,tx);
1133             fjy1             = _mm_add_pd(fjy1,ty);
1134             fjz1             = _mm_add_pd(fjz1,tz);
1135
1136             }
1137
1138             /**************************
1139              * CALCULATE INTERACTIONS *
1140              **************************/
1141
1142             if (gmx_mm_any_lt(rsq32,rcutoff2))
1143             {
1144
1145             /* REACTION-FIELD ELECTROSTATICS */
1146             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
1147             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1148
1149             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1150
1151             /* Update potential sum for this i atom from the interaction with this j atom. */
1152             velec            = _mm_and_pd(velec,cutoff_mask);
1153             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1154             velecsum         = _mm_add_pd(velecsum,velec);
1155
1156             fscal            = felec;
1157
1158             fscal            = _mm_and_pd(fscal,cutoff_mask);
1159
1160             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1161
1162             /* Calculate temporary vectorial force */
1163             tx               = _mm_mul_pd(fscal,dx32);
1164             ty               = _mm_mul_pd(fscal,dy32);
1165             tz               = _mm_mul_pd(fscal,dz32);
1166
1167             /* Update vectorial force */
1168             fix3             = _mm_add_pd(fix3,tx);
1169             fiy3             = _mm_add_pd(fiy3,ty);
1170             fiz3             = _mm_add_pd(fiz3,tz);
1171
1172             fjx2             = _mm_add_pd(fjx2,tx);
1173             fjy2             = _mm_add_pd(fjy2,ty);
1174             fjz2             = _mm_add_pd(fjz2,tz);
1175
1176             }
1177
1178             /**************************
1179              * CALCULATE INTERACTIONS *
1180              **************************/
1181
1182             if (gmx_mm_any_lt(rsq33,rcutoff2))
1183             {
1184
1185             /* REACTION-FIELD ELECTROSTATICS */
1186             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1187             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1188
1189             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1190
1191             /* Update potential sum for this i atom from the interaction with this j atom. */
1192             velec            = _mm_and_pd(velec,cutoff_mask);
1193             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1194             velecsum         = _mm_add_pd(velecsum,velec);
1195
1196             fscal            = felec;
1197
1198             fscal            = _mm_and_pd(fscal,cutoff_mask);
1199
1200             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1201
1202             /* Calculate temporary vectorial force */
1203             tx               = _mm_mul_pd(fscal,dx33);
1204             ty               = _mm_mul_pd(fscal,dy33);
1205             tz               = _mm_mul_pd(fscal,dz33);
1206
1207             /* Update vectorial force */
1208             fix3             = _mm_add_pd(fix3,tx);
1209             fiy3             = _mm_add_pd(fiy3,ty);
1210             fiz3             = _mm_add_pd(fiz3,tz);
1211
1212             fjx3             = _mm_add_pd(fjx3,tx);
1213             fjy3             = _mm_add_pd(fjy3,ty);
1214             fjz3             = _mm_add_pd(fjz3,tz);
1215
1216             }
1217
1218             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1219
1220             /* Inner loop uses 386 flops */
1221         }
1222
1223         /* End of innermost loop */
1224
1225         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1226                                               f+i_coord_offset,fshift+i_shift_offset);
1227
1228         ggid                        = gid[iidx];
1229         /* Update potential energies */
1230         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1231         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1232
1233         /* Increment number of inner iterations */
1234         inneriter                  += j_index_end - j_index_start;
1235
1236         /* Outer loop uses 26 flops */
1237     }
1238
1239     /* Increment number of outer iterations */
1240     outeriter        += nri;
1241
1242     /* Update outer/inner flops */
1243
1244     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*386);
1245 }
1246 /*
1247  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sse2_double
1248  * Electrostatics interaction: ReactionField
1249  * VdW interaction:            LennardJones
1250  * Geometry:                   Water4-Water4
1251  * Calculate force/pot:        Force
1252  */
1253 void
1254 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_sse2_double
1255                     (t_nblist                    * gmx_restrict       nlist,
1256                      rvec                        * gmx_restrict          xx,
1257                      rvec                        * gmx_restrict          ff,
1258                      t_forcerec                  * gmx_restrict          fr,
1259                      t_mdatoms                   * gmx_restrict     mdatoms,
1260                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1261                      t_nrnb                      * gmx_restrict        nrnb)
1262 {
1263     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1264      * just 0 for non-waters.
1265      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1266      * jnr indices corresponding to data put in the four positions in the SIMD register.
1267      */
1268     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1269     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1270     int              jnrA,jnrB;
1271     int              j_coord_offsetA,j_coord_offsetB;
1272     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1273     real             rcutoff_scalar;
1274     real             *shiftvec,*fshift,*x,*f;
1275     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1276     int              vdwioffset0;
1277     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1278     int              vdwioffset1;
1279     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1280     int              vdwioffset2;
1281     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1282     int              vdwioffset3;
1283     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1284     int              vdwjidx0A,vdwjidx0B;
1285     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1286     int              vdwjidx1A,vdwjidx1B;
1287     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1288     int              vdwjidx2A,vdwjidx2B;
1289     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1290     int              vdwjidx3A,vdwjidx3B;
1291     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1292     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1293     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1294     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1295     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1296     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1297     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1298     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1299     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1300     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1301     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1302     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1303     real             *charge;
1304     int              nvdwtype;
1305     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1306     int              *vdwtype;
1307     real             *vdwparam;
1308     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1309     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1310     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1311     real             rswitch_scalar,d_scalar;
1312     __m128d          dummy_mask,cutoff_mask;
1313     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1314     __m128d          one     = _mm_set1_pd(1.0);
1315     __m128d          two     = _mm_set1_pd(2.0);
1316     x                = xx[0];
1317     f                = ff[0];
1318
1319     nri              = nlist->nri;
1320     iinr             = nlist->iinr;
1321     jindex           = nlist->jindex;
1322     jjnr             = nlist->jjnr;
1323     shiftidx         = nlist->shift;
1324     gid              = nlist->gid;
1325     shiftvec         = fr->shift_vec[0];
1326     fshift           = fr->fshift[0];
1327     facel            = _mm_set1_pd(fr->epsfac);
1328     charge           = mdatoms->chargeA;
1329     krf              = _mm_set1_pd(fr->ic->k_rf);
1330     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1331     crf              = _mm_set1_pd(fr->ic->c_rf);
1332     nvdwtype         = fr->ntype;
1333     vdwparam         = fr->nbfp;
1334     vdwtype          = mdatoms->typeA;
1335
1336     /* Setup water-specific parameters */
1337     inr              = nlist->iinr[0];
1338     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1339     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1340     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1341     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1342
1343     jq1              = _mm_set1_pd(charge[inr+1]);
1344     jq2              = _mm_set1_pd(charge[inr+2]);
1345     jq3              = _mm_set1_pd(charge[inr+3]);
1346     vdwjidx0A        = 2*vdwtype[inr+0];
1347     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1348     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1349     qq11             = _mm_mul_pd(iq1,jq1);
1350     qq12             = _mm_mul_pd(iq1,jq2);
1351     qq13             = _mm_mul_pd(iq1,jq3);
1352     qq21             = _mm_mul_pd(iq2,jq1);
1353     qq22             = _mm_mul_pd(iq2,jq2);
1354     qq23             = _mm_mul_pd(iq2,jq3);
1355     qq31             = _mm_mul_pd(iq3,jq1);
1356     qq32             = _mm_mul_pd(iq3,jq2);
1357     qq33             = _mm_mul_pd(iq3,jq3);
1358
1359     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1360     rcutoff_scalar   = fr->rcoulomb;
1361     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1362     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1363
1364     rswitch_scalar   = fr->rvdw_switch;
1365     rswitch          = _mm_set1_pd(rswitch_scalar);
1366     /* Setup switch parameters */
1367     d_scalar         = rcutoff_scalar-rswitch_scalar;
1368     d                = _mm_set1_pd(d_scalar);
1369     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1370     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1371     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1372     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1373     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1374     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1375
1376     /* Avoid stupid compiler warnings */
1377     jnrA = jnrB = 0;
1378     j_coord_offsetA = 0;
1379     j_coord_offsetB = 0;
1380
1381     outeriter        = 0;
1382     inneriter        = 0;
1383
1384     /* Start outer loop over neighborlists */
1385     for(iidx=0; iidx<nri; iidx++)
1386     {
1387         /* Load shift vector for this list */
1388         i_shift_offset   = DIM*shiftidx[iidx];
1389
1390         /* Load limits for loop over neighbors */
1391         j_index_start    = jindex[iidx];
1392         j_index_end      = jindex[iidx+1];
1393
1394         /* Get outer coordinate index */
1395         inr              = iinr[iidx];
1396         i_coord_offset   = DIM*inr;
1397
1398         /* Load i particle coords and add shift vector */
1399         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1400                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1401
1402         fix0             = _mm_setzero_pd();
1403         fiy0             = _mm_setzero_pd();
1404         fiz0             = _mm_setzero_pd();
1405         fix1             = _mm_setzero_pd();
1406         fiy1             = _mm_setzero_pd();
1407         fiz1             = _mm_setzero_pd();
1408         fix2             = _mm_setzero_pd();
1409         fiy2             = _mm_setzero_pd();
1410         fiz2             = _mm_setzero_pd();
1411         fix3             = _mm_setzero_pd();
1412         fiy3             = _mm_setzero_pd();
1413         fiz3             = _mm_setzero_pd();
1414
1415         /* Start inner kernel loop */
1416         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1417         {
1418
1419             /* Get j neighbor index, and coordinate index */
1420             jnrA             = jjnr[jidx];
1421             jnrB             = jjnr[jidx+1];
1422             j_coord_offsetA  = DIM*jnrA;
1423             j_coord_offsetB  = DIM*jnrB;
1424
1425             /* load j atom coordinates */
1426             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1427                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1428                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1429
1430             /* Calculate displacement vector */
1431             dx00             = _mm_sub_pd(ix0,jx0);
1432             dy00             = _mm_sub_pd(iy0,jy0);
1433             dz00             = _mm_sub_pd(iz0,jz0);
1434             dx11             = _mm_sub_pd(ix1,jx1);
1435             dy11             = _mm_sub_pd(iy1,jy1);
1436             dz11             = _mm_sub_pd(iz1,jz1);
1437             dx12             = _mm_sub_pd(ix1,jx2);
1438             dy12             = _mm_sub_pd(iy1,jy2);
1439             dz12             = _mm_sub_pd(iz1,jz2);
1440             dx13             = _mm_sub_pd(ix1,jx3);
1441             dy13             = _mm_sub_pd(iy1,jy3);
1442             dz13             = _mm_sub_pd(iz1,jz3);
1443             dx21             = _mm_sub_pd(ix2,jx1);
1444             dy21             = _mm_sub_pd(iy2,jy1);
1445             dz21             = _mm_sub_pd(iz2,jz1);
1446             dx22             = _mm_sub_pd(ix2,jx2);
1447             dy22             = _mm_sub_pd(iy2,jy2);
1448             dz22             = _mm_sub_pd(iz2,jz2);
1449             dx23             = _mm_sub_pd(ix2,jx3);
1450             dy23             = _mm_sub_pd(iy2,jy3);
1451             dz23             = _mm_sub_pd(iz2,jz3);
1452             dx31             = _mm_sub_pd(ix3,jx1);
1453             dy31             = _mm_sub_pd(iy3,jy1);
1454             dz31             = _mm_sub_pd(iz3,jz1);
1455             dx32             = _mm_sub_pd(ix3,jx2);
1456             dy32             = _mm_sub_pd(iy3,jy2);
1457             dz32             = _mm_sub_pd(iz3,jz2);
1458             dx33             = _mm_sub_pd(ix3,jx3);
1459             dy33             = _mm_sub_pd(iy3,jy3);
1460             dz33             = _mm_sub_pd(iz3,jz3);
1461
1462             /* Calculate squared distance and things based on it */
1463             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1464             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1465             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1466             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1467             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1468             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1469             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1470             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1471             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1472             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1473
1474             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1475             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1476             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1477             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1478             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1479             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1480             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1481             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1482             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1483             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1484
1485             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1486             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1487             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1488             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1489             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1490             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1491             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1492             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1493             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1494             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1495
1496             fjx0             = _mm_setzero_pd();
1497             fjy0             = _mm_setzero_pd();
1498             fjz0             = _mm_setzero_pd();
1499             fjx1             = _mm_setzero_pd();
1500             fjy1             = _mm_setzero_pd();
1501             fjz1             = _mm_setzero_pd();
1502             fjx2             = _mm_setzero_pd();
1503             fjy2             = _mm_setzero_pd();
1504             fjz2             = _mm_setzero_pd();
1505             fjx3             = _mm_setzero_pd();
1506             fjy3             = _mm_setzero_pd();
1507             fjz3             = _mm_setzero_pd();
1508
1509             /**************************
1510              * CALCULATE INTERACTIONS *
1511              **************************/
1512
1513             if (gmx_mm_any_lt(rsq00,rcutoff2))
1514             {
1515
1516             r00              = _mm_mul_pd(rsq00,rinv00);
1517
1518             /* LENNARD-JONES DISPERSION/REPULSION */
1519
1520             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1521             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1522             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1523             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
1524             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1525
1526             d                = _mm_sub_pd(r00,rswitch);
1527             d                = _mm_max_pd(d,_mm_setzero_pd());
1528             d2               = _mm_mul_pd(d,d);
1529             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)))))));
1530
1531             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
1532
1533             /* Evaluate switch function */
1534             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1535             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1536             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1537
1538             fscal            = fvdw;
1539
1540             fscal            = _mm_and_pd(fscal,cutoff_mask);
1541
1542             /* Calculate temporary vectorial force */
1543             tx               = _mm_mul_pd(fscal,dx00);
1544             ty               = _mm_mul_pd(fscal,dy00);
1545             tz               = _mm_mul_pd(fscal,dz00);
1546
1547             /* Update vectorial force */
1548             fix0             = _mm_add_pd(fix0,tx);
1549             fiy0             = _mm_add_pd(fiy0,ty);
1550             fiz0             = _mm_add_pd(fiz0,tz);
1551
1552             fjx0             = _mm_add_pd(fjx0,tx);
1553             fjy0             = _mm_add_pd(fjy0,ty);
1554             fjz0             = _mm_add_pd(fjz0,tz);
1555
1556             }
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             if (gmx_mm_any_lt(rsq11,rcutoff2))
1563             {
1564
1565             /* REACTION-FIELD ELECTROSTATICS */
1566             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1567
1568             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1569
1570             fscal            = felec;
1571
1572             fscal            = _mm_and_pd(fscal,cutoff_mask);
1573
1574             /* Calculate temporary vectorial force */
1575             tx               = _mm_mul_pd(fscal,dx11);
1576             ty               = _mm_mul_pd(fscal,dy11);
1577             tz               = _mm_mul_pd(fscal,dz11);
1578
1579             /* Update vectorial force */
1580             fix1             = _mm_add_pd(fix1,tx);
1581             fiy1             = _mm_add_pd(fiy1,ty);
1582             fiz1             = _mm_add_pd(fiz1,tz);
1583
1584             fjx1             = _mm_add_pd(fjx1,tx);
1585             fjy1             = _mm_add_pd(fjy1,ty);
1586             fjz1             = _mm_add_pd(fjz1,tz);
1587
1588             }
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             if (gmx_mm_any_lt(rsq12,rcutoff2))
1595             {
1596
1597             /* REACTION-FIELD ELECTROSTATICS */
1598             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1599
1600             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1601
1602             fscal            = felec;
1603
1604             fscal            = _mm_and_pd(fscal,cutoff_mask);
1605
1606             /* Calculate temporary vectorial force */
1607             tx               = _mm_mul_pd(fscal,dx12);
1608             ty               = _mm_mul_pd(fscal,dy12);
1609             tz               = _mm_mul_pd(fscal,dz12);
1610
1611             /* Update vectorial force */
1612             fix1             = _mm_add_pd(fix1,tx);
1613             fiy1             = _mm_add_pd(fiy1,ty);
1614             fiz1             = _mm_add_pd(fiz1,tz);
1615
1616             fjx2             = _mm_add_pd(fjx2,tx);
1617             fjy2             = _mm_add_pd(fjy2,ty);
1618             fjz2             = _mm_add_pd(fjz2,tz);
1619
1620             }
1621
1622             /**************************
1623              * CALCULATE INTERACTIONS *
1624              **************************/
1625
1626             if (gmx_mm_any_lt(rsq13,rcutoff2))
1627             {
1628
1629             /* REACTION-FIELD ELECTROSTATICS */
1630             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1631
1632             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1633
1634             fscal            = felec;
1635
1636             fscal            = _mm_and_pd(fscal,cutoff_mask);
1637
1638             /* Calculate temporary vectorial force */
1639             tx               = _mm_mul_pd(fscal,dx13);
1640             ty               = _mm_mul_pd(fscal,dy13);
1641             tz               = _mm_mul_pd(fscal,dz13);
1642
1643             /* Update vectorial force */
1644             fix1             = _mm_add_pd(fix1,tx);
1645             fiy1             = _mm_add_pd(fiy1,ty);
1646             fiz1             = _mm_add_pd(fiz1,tz);
1647
1648             fjx3             = _mm_add_pd(fjx3,tx);
1649             fjy3             = _mm_add_pd(fjy3,ty);
1650             fjz3             = _mm_add_pd(fjz3,tz);
1651
1652             }
1653
1654             /**************************
1655              * CALCULATE INTERACTIONS *
1656              **************************/
1657
1658             if (gmx_mm_any_lt(rsq21,rcutoff2))
1659             {
1660
1661             /* REACTION-FIELD ELECTROSTATICS */
1662             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1663
1664             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1665
1666             fscal            = felec;
1667
1668             fscal            = _mm_and_pd(fscal,cutoff_mask);
1669
1670             /* Calculate temporary vectorial force */
1671             tx               = _mm_mul_pd(fscal,dx21);
1672             ty               = _mm_mul_pd(fscal,dy21);
1673             tz               = _mm_mul_pd(fscal,dz21);
1674
1675             /* Update vectorial force */
1676             fix2             = _mm_add_pd(fix2,tx);
1677             fiy2             = _mm_add_pd(fiy2,ty);
1678             fiz2             = _mm_add_pd(fiz2,tz);
1679
1680             fjx1             = _mm_add_pd(fjx1,tx);
1681             fjy1             = _mm_add_pd(fjy1,ty);
1682             fjz1             = _mm_add_pd(fjz1,tz);
1683
1684             }
1685
1686             /**************************
1687              * CALCULATE INTERACTIONS *
1688              **************************/
1689
1690             if (gmx_mm_any_lt(rsq22,rcutoff2))
1691             {
1692
1693             /* REACTION-FIELD ELECTROSTATICS */
1694             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1695
1696             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1697
1698             fscal            = felec;
1699
1700             fscal            = _mm_and_pd(fscal,cutoff_mask);
1701
1702             /* Calculate temporary vectorial force */
1703             tx               = _mm_mul_pd(fscal,dx22);
1704             ty               = _mm_mul_pd(fscal,dy22);
1705             tz               = _mm_mul_pd(fscal,dz22);
1706
1707             /* Update vectorial force */
1708             fix2             = _mm_add_pd(fix2,tx);
1709             fiy2             = _mm_add_pd(fiy2,ty);
1710             fiz2             = _mm_add_pd(fiz2,tz);
1711
1712             fjx2             = _mm_add_pd(fjx2,tx);
1713             fjy2             = _mm_add_pd(fjy2,ty);
1714             fjz2             = _mm_add_pd(fjz2,tz);
1715
1716             }
1717
1718             /**************************
1719              * CALCULATE INTERACTIONS *
1720              **************************/
1721
1722             if (gmx_mm_any_lt(rsq23,rcutoff2))
1723             {
1724
1725             /* REACTION-FIELD ELECTROSTATICS */
1726             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1727
1728             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1729
1730             fscal            = felec;
1731
1732             fscal            = _mm_and_pd(fscal,cutoff_mask);
1733
1734             /* Calculate temporary vectorial force */
1735             tx               = _mm_mul_pd(fscal,dx23);
1736             ty               = _mm_mul_pd(fscal,dy23);
1737             tz               = _mm_mul_pd(fscal,dz23);
1738
1739             /* Update vectorial force */
1740             fix2             = _mm_add_pd(fix2,tx);
1741             fiy2             = _mm_add_pd(fiy2,ty);
1742             fiz2             = _mm_add_pd(fiz2,tz);
1743
1744             fjx3             = _mm_add_pd(fjx3,tx);
1745             fjy3             = _mm_add_pd(fjy3,ty);
1746             fjz3             = _mm_add_pd(fjz3,tz);
1747
1748             }
1749
1750             /**************************
1751              * CALCULATE INTERACTIONS *
1752              **************************/
1753
1754             if (gmx_mm_any_lt(rsq31,rcutoff2))
1755             {
1756
1757             /* REACTION-FIELD ELECTROSTATICS */
1758             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1759
1760             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1761
1762             fscal            = felec;
1763
1764             fscal            = _mm_and_pd(fscal,cutoff_mask);
1765
1766             /* Calculate temporary vectorial force */
1767             tx               = _mm_mul_pd(fscal,dx31);
1768             ty               = _mm_mul_pd(fscal,dy31);
1769             tz               = _mm_mul_pd(fscal,dz31);
1770
1771             /* Update vectorial force */
1772             fix3             = _mm_add_pd(fix3,tx);
1773             fiy3             = _mm_add_pd(fiy3,ty);
1774             fiz3             = _mm_add_pd(fiz3,tz);
1775
1776             fjx1             = _mm_add_pd(fjx1,tx);
1777             fjy1             = _mm_add_pd(fjy1,ty);
1778             fjz1             = _mm_add_pd(fjz1,tz);
1779
1780             }
1781
1782             /**************************
1783              * CALCULATE INTERACTIONS *
1784              **************************/
1785
1786             if (gmx_mm_any_lt(rsq32,rcutoff2))
1787             {
1788
1789             /* REACTION-FIELD ELECTROSTATICS */
1790             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1791
1792             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1793
1794             fscal            = felec;
1795
1796             fscal            = _mm_and_pd(fscal,cutoff_mask);
1797
1798             /* Calculate temporary vectorial force */
1799             tx               = _mm_mul_pd(fscal,dx32);
1800             ty               = _mm_mul_pd(fscal,dy32);
1801             tz               = _mm_mul_pd(fscal,dz32);
1802
1803             /* Update vectorial force */
1804             fix3             = _mm_add_pd(fix3,tx);
1805             fiy3             = _mm_add_pd(fiy3,ty);
1806             fiz3             = _mm_add_pd(fiz3,tz);
1807
1808             fjx2             = _mm_add_pd(fjx2,tx);
1809             fjy2             = _mm_add_pd(fjy2,ty);
1810             fjz2             = _mm_add_pd(fjz2,tz);
1811
1812             }
1813
1814             /**************************
1815              * CALCULATE INTERACTIONS *
1816              **************************/
1817
1818             if (gmx_mm_any_lt(rsq33,rcutoff2))
1819             {
1820
1821             /* REACTION-FIELD ELECTROSTATICS */
1822             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1823
1824             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1825
1826             fscal            = felec;
1827
1828             fscal            = _mm_and_pd(fscal,cutoff_mask);
1829
1830             /* Calculate temporary vectorial force */
1831             tx               = _mm_mul_pd(fscal,dx33);
1832             ty               = _mm_mul_pd(fscal,dy33);
1833             tz               = _mm_mul_pd(fscal,dz33);
1834
1835             /* Update vectorial force */
1836             fix3             = _mm_add_pd(fix3,tx);
1837             fiy3             = _mm_add_pd(fiy3,ty);
1838             fiz3             = _mm_add_pd(fiz3,tz);
1839
1840             fjx3             = _mm_add_pd(fjx3,tx);
1841             fjy3             = _mm_add_pd(fjy3,ty);
1842             fjz3             = _mm_add_pd(fjz3,tz);
1843
1844             }
1845
1846             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);
1847
1848             /* Inner loop uses 329 flops */
1849         }
1850
1851         if(jidx<j_index_end)
1852         {
1853
1854             jnrA             = jjnr[jidx];
1855             j_coord_offsetA  = DIM*jnrA;
1856
1857             /* load j atom coordinates */
1858             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1859                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1860                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1861
1862             /* Calculate displacement vector */
1863             dx00             = _mm_sub_pd(ix0,jx0);
1864             dy00             = _mm_sub_pd(iy0,jy0);
1865             dz00             = _mm_sub_pd(iz0,jz0);
1866             dx11             = _mm_sub_pd(ix1,jx1);
1867             dy11             = _mm_sub_pd(iy1,jy1);
1868             dz11             = _mm_sub_pd(iz1,jz1);
1869             dx12             = _mm_sub_pd(ix1,jx2);
1870             dy12             = _mm_sub_pd(iy1,jy2);
1871             dz12             = _mm_sub_pd(iz1,jz2);
1872             dx13             = _mm_sub_pd(ix1,jx3);
1873             dy13             = _mm_sub_pd(iy1,jy3);
1874             dz13             = _mm_sub_pd(iz1,jz3);
1875             dx21             = _mm_sub_pd(ix2,jx1);
1876             dy21             = _mm_sub_pd(iy2,jy1);
1877             dz21             = _mm_sub_pd(iz2,jz1);
1878             dx22             = _mm_sub_pd(ix2,jx2);
1879             dy22             = _mm_sub_pd(iy2,jy2);
1880             dz22             = _mm_sub_pd(iz2,jz2);
1881             dx23             = _mm_sub_pd(ix2,jx3);
1882             dy23             = _mm_sub_pd(iy2,jy3);
1883             dz23             = _mm_sub_pd(iz2,jz3);
1884             dx31             = _mm_sub_pd(ix3,jx1);
1885             dy31             = _mm_sub_pd(iy3,jy1);
1886             dz31             = _mm_sub_pd(iz3,jz1);
1887             dx32             = _mm_sub_pd(ix3,jx2);
1888             dy32             = _mm_sub_pd(iy3,jy2);
1889             dz32             = _mm_sub_pd(iz3,jz2);
1890             dx33             = _mm_sub_pd(ix3,jx3);
1891             dy33             = _mm_sub_pd(iy3,jy3);
1892             dz33             = _mm_sub_pd(iz3,jz3);
1893
1894             /* Calculate squared distance and things based on it */
1895             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1896             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1897             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1898             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1899             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1900             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1901             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1902             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1903             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1904             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1905
1906             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1907             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1908             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1909             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1910             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1911             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1912             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1913             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1914             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1915             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1916
1917             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1918             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1919             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1920             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1921             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1922             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1923             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1924             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1925             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1926             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1927
1928             fjx0             = _mm_setzero_pd();
1929             fjy0             = _mm_setzero_pd();
1930             fjz0             = _mm_setzero_pd();
1931             fjx1             = _mm_setzero_pd();
1932             fjy1             = _mm_setzero_pd();
1933             fjz1             = _mm_setzero_pd();
1934             fjx2             = _mm_setzero_pd();
1935             fjy2             = _mm_setzero_pd();
1936             fjz2             = _mm_setzero_pd();
1937             fjx3             = _mm_setzero_pd();
1938             fjy3             = _mm_setzero_pd();
1939             fjz3             = _mm_setzero_pd();
1940
1941             /**************************
1942              * CALCULATE INTERACTIONS *
1943              **************************/
1944
1945             if (gmx_mm_any_lt(rsq00,rcutoff2))
1946             {
1947
1948             r00              = _mm_mul_pd(rsq00,rinv00);
1949
1950             /* LENNARD-JONES DISPERSION/REPULSION */
1951
1952             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1953             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1954             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1955             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
1956             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1957
1958             d                = _mm_sub_pd(r00,rswitch);
1959             d                = _mm_max_pd(d,_mm_setzero_pd());
1960             d2               = _mm_mul_pd(d,d);
1961             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)))))));
1962
1963             dsw              = _mm_mul_pd(d2,_mm_add_pd(swF2,_mm_mul_pd(d,_mm_add_pd(swF3,_mm_mul_pd(d,swF4)))));
1964
1965             /* Evaluate switch function */
1966             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1967             fvdw             = _mm_sub_pd( _mm_mul_pd(fvdw,sw) , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1968             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1969
1970             fscal            = fvdw;
1971
1972             fscal            = _mm_and_pd(fscal,cutoff_mask);
1973
1974             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1975
1976             /* Calculate temporary vectorial force */
1977             tx               = _mm_mul_pd(fscal,dx00);
1978             ty               = _mm_mul_pd(fscal,dy00);
1979             tz               = _mm_mul_pd(fscal,dz00);
1980
1981             /* Update vectorial force */
1982             fix0             = _mm_add_pd(fix0,tx);
1983             fiy0             = _mm_add_pd(fiy0,ty);
1984             fiz0             = _mm_add_pd(fiz0,tz);
1985
1986             fjx0             = _mm_add_pd(fjx0,tx);
1987             fjy0             = _mm_add_pd(fjy0,ty);
1988             fjz0             = _mm_add_pd(fjz0,tz);
1989
1990             }
1991
1992             /**************************
1993              * CALCULATE INTERACTIONS *
1994              **************************/
1995
1996             if (gmx_mm_any_lt(rsq11,rcutoff2))
1997             {
1998
1999             /* REACTION-FIELD ELECTROSTATICS */
2000             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
2001
2002             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
2003
2004             fscal            = felec;
2005
2006             fscal            = _mm_and_pd(fscal,cutoff_mask);
2007
2008             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2009
2010             /* Calculate temporary vectorial force */
2011             tx               = _mm_mul_pd(fscal,dx11);
2012             ty               = _mm_mul_pd(fscal,dy11);
2013             tz               = _mm_mul_pd(fscal,dz11);
2014
2015             /* Update vectorial force */
2016             fix1             = _mm_add_pd(fix1,tx);
2017             fiy1             = _mm_add_pd(fiy1,ty);
2018             fiz1             = _mm_add_pd(fiz1,tz);
2019
2020             fjx1             = _mm_add_pd(fjx1,tx);
2021             fjy1             = _mm_add_pd(fjy1,ty);
2022             fjz1             = _mm_add_pd(fjz1,tz);
2023
2024             }
2025
2026             /**************************
2027              * CALCULATE INTERACTIONS *
2028              **************************/
2029
2030             if (gmx_mm_any_lt(rsq12,rcutoff2))
2031             {
2032
2033             /* REACTION-FIELD ELECTROSTATICS */
2034             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
2035
2036             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2037
2038             fscal            = felec;
2039
2040             fscal            = _mm_and_pd(fscal,cutoff_mask);
2041
2042             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2043
2044             /* Calculate temporary vectorial force */
2045             tx               = _mm_mul_pd(fscal,dx12);
2046             ty               = _mm_mul_pd(fscal,dy12);
2047             tz               = _mm_mul_pd(fscal,dz12);
2048
2049             /* Update vectorial force */
2050             fix1             = _mm_add_pd(fix1,tx);
2051             fiy1             = _mm_add_pd(fiy1,ty);
2052             fiz1             = _mm_add_pd(fiz1,tz);
2053
2054             fjx2             = _mm_add_pd(fjx2,tx);
2055             fjy2             = _mm_add_pd(fjy2,ty);
2056             fjz2             = _mm_add_pd(fjz2,tz);
2057
2058             }
2059
2060             /**************************
2061              * CALCULATE INTERACTIONS *
2062              **************************/
2063
2064             if (gmx_mm_any_lt(rsq13,rcutoff2))
2065             {
2066
2067             /* REACTION-FIELD ELECTROSTATICS */
2068             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
2069
2070             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
2071
2072             fscal            = felec;
2073
2074             fscal            = _mm_and_pd(fscal,cutoff_mask);
2075
2076             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2077
2078             /* Calculate temporary vectorial force */
2079             tx               = _mm_mul_pd(fscal,dx13);
2080             ty               = _mm_mul_pd(fscal,dy13);
2081             tz               = _mm_mul_pd(fscal,dz13);
2082
2083             /* Update vectorial force */
2084             fix1             = _mm_add_pd(fix1,tx);
2085             fiy1             = _mm_add_pd(fiy1,ty);
2086             fiz1             = _mm_add_pd(fiz1,tz);
2087
2088             fjx3             = _mm_add_pd(fjx3,tx);
2089             fjy3             = _mm_add_pd(fjy3,ty);
2090             fjz3             = _mm_add_pd(fjz3,tz);
2091
2092             }
2093
2094             /**************************
2095              * CALCULATE INTERACTIONS *
2096              **************************/
2097
2098             if (gmx_mm_any_lt(rsq21,rcutoff2))
2099             {
2100
2101             /* REACTION-FIELD ELECTROSTATICS */
2102             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2103
2104             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2105
2106             fscal            = felec;
2107
2108             fscal            = _mm_and_pd(fscal,cutoff_mask);
2109
2110             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2111
2112             /* Calculate temporary vectorial force */
2113             tx               = _mm_mul_pd(fscal,dx21);
2114             ty               = _mm_mul_pd(fscal,dy21);
2115             tz               = _mm_mul_pd(fscal,dz21);
2116
2117             /* Update vectorial force */
2118             fix2             = _mm_add_pd(fix2,tx);
2119             fiy2             = _mm_add_pd(fiy2,ty);
2120             fiz2             = _mm_add_pd(fiz2,tz);
2121
2122             fjx1             = _mm_add_pd(fjx1,tx);
2123             fjy1             = _mm_add_pd(fjy1,ty);
2124             fjz1             = _mm_add_pd(fjz1,tz);
2125
2126             }
2127
2128             /**************************
2129              * CALCULATE INTERACTIONS *
2130              **************************/
2131
2132             if (gmx_mm_any_lt(rsq22,rcutoff2))
2133             {
2134
2135             /* REACTION-FIELD ELECTROSTATICS */
2136             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2137
2138             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2139
2140             fscal            = felec;
2141
2142             fscal            = _mm_and_pd(fscal,cutoff_mask);
2143
2144             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2145
2146             /* Calculate temporary vectorial force */
2147             tx               = _mm_mul_pd(fscal,dx22);
2148             ty               = _mm_mul_pd(fscal,dy22);
2149             tz               = _mm_mul_pd(fscal,dz22);
2150
2151             /* Update vectorial force */
2152             fix2             = _mm_add_pd(fix2,tx);
2153             fiy2             = _mm_add_pd(fiy2,ty);
2154             fiz2             = _mm_add_pd(fiz2,tz);
2155
2156             fjx2             = _mm_add_pd(fjx2,tx);
2157             fjy2             = _mm_add_pd(fjy2,ty);
2158             fjz2             = _mm_add_pd(fjz2,tz);
2159
2160             }
2161
2162             /**************************
2163              * CALCULATE INTERACTIONS *
2164              **************************/
2165
2166             if (gmx_mm_any_lt(rsq23,rcutoff2))
2167             {
2168
2169             /* REACTION-FIELD ELECTROSTATICS */
2170             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
2171
2172             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2173
2174             fscal            = felec;
2175
2176             fscal            = _mm_and_pd(fscal,cutoff_mask);
2177
2178             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2179
2180             /* Calculate temporary vectorial force */
2181             tx               = _mm_mul_pd(fscal,dx23);
2182             ty               = _mm_mul_pd(fscal,dy23);
2183             tz               = _mm_mul_pd(fscal,dz23);
2184
2185             /* Update vectorial force */
2186             fix2             = _mm_add_pd(fix2,tx);
2187             fiy2             = _mm_add_pd(fiy2,ty);
2188             fiz2             = _mm_add_pd(fiz2,tz);
2189
2190             fjx3             = _mm_add_pd(fjx3,tx);
2191             fjy3             = _mm_add_pd(fjy3,ty);
2192             fjz3             = _mm_add_pd(fjz3,tz);
2193
2194             }
2195
2196             /**************************
2197              * CALCULATE INTERACTIONS *
2198              **************************/
2199
2200             if (gmx_mm_any_lt(rsq31,rcutoff2))
2201             {
2202
2203             /* REACTION-FIELD ELECTROSTATICS */
2204             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
2205
2206             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2207
2208             fscal            = felec;
2209
2210             fscal            = _mm_and_pd(fscal,cutoff_mask);
2211
2212             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2213
2214             /* Calculate temporary vectorial force */
2215             tx               = _mm_mul_pd(fscal,dx31);
2216             ty               = _mm_mul_pd(fscal,dy31);
2217             tz               = _mm_mul_pd(fscal,dz31);
2218
2219             /* Update vectorial force */
2220             fix3             = _mm_add_pd(fix3,tx);
2221             fiy3             = _mm_add_pd(fiy3,ty);
2222             fiz3             = _mm_add_pd(fiz3,tz);
2223
2224             fjx1             = _mm_add_pd(fjx1,tx);
2225             fjy1             = _mm_add_pd(fjy1,ty);
2226             fjz1             = _mm_add_pd(fjz1,tz);
2227
2228             }
2229
2230             /**************************
2231              * CALCULATE INTERACTIONS *
2232              **************************/
2233
2234             if (gmx_mm_any_lt(rsq32,rcutoff2))
2235             {
2236
2237             /* REACTION-FIELD ELECTROSTATICS */
2238             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
2239
2240             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2241
2242             fscal            = felec;
2243
2244             fscal            = _mm_and_pd(fscal,cutoff_mask);
2245
2246             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2247
2248             /* Calculate temporary vectorial force */
2249             tx               = _mm_mul_pd(fscal,dx32);
2250             ty               = _mm_mul_pd(fscal,dy32);
2251             tz               = _mm_mul_pd(fscal,dz32);
2252
2253             /* Update vectorial force */
2254             fix3             = _mm_add_pd(fix3,tx);
2255             fiy3             = _mm_add_pd(fiy3,ty);
2256             fiz3             = _mm_add_pd(fiz3,tz);
2257
2258             fjx2             = _mm_add_pd(fjx2,tx);
2259             fjy2             = _mm_add_pd(fjy2,ty);
2260             fjz2             = _mm_add_pd(fjz2,tz);
2261
2262             }
2263
2264             /**************************
2265              * CALCULATE INTERACTIONS *
2266              **************************/
2267
2268             if (gmx_mm_any_lt(rsq33,rcutoff2))
2269             {
2270
2271             /* REACTION-FIELD ELECTROSTATICS */
2272             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
2273
2274             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2275
2276             fscal            = felec;
2277
2278             fscal            = _mm_and_pd(fscal,cutoff_mask);
2279
2280             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2281
2282             /* Calculate temporary vectorial force */
2283             tx               = _mm_mul_pd(fscal,dx33);
2284             ty               = _mm_mul_pd(fscal,dy33);
2285             tz               = _mm_mul_pd(fscal,dz33);
2286
2287             /* Update vectorial force */
2288             fix3             = _mm_add_pd(fix3,tx);
2289             fiy3             = _mm_add_pd(fiy3,ty);
2290             fiz3             = _mm_add_pd(fiz3,tz);
2291
2292             fjx3             = _mm_add_pd(fjx3,tx);
2293             fjy3             = _mm_add_pd(fjy3,ty);
2294             fjz3             = _mm_add_pd(fjz3,tz);
2295
2296             }
2297
2298             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2299
2300             /* Inner loop uses 329 flops */
2301         }
2302
2303         /* End of innermost loop */
2304
2305         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2306                                               f+i_coord_offset,fshift+i_shift_offset);
2307
2308         /* Increment number of inner iterations */
2309         inneriter                  += j_index_end - j_index_start;
2310
2311         /* Outer loop uses 24 flops */
2312     }
2313
2314     /* Increment number of outer iterations */
2315     outeriter        += nri;
2316
2317     /* Update outer/inner flops */
2318
2319     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*329);
2320 }