src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwNone_GeomW4W4_sse4_1_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 sse4_1_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse4_1_double.h"
  50 #include "kernelutil_x86_sse4_1_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sse4_1_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sse4_1_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              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     int              vdwjidx3A,vdwjidx3B;
  93     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  94     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  95     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  96     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  97     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  98     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  99     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 100     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 101     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 102     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 103     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128d          dummy_mask,cutoff_mask;
 106     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 107     __m128d          one     = _mm_set1_pd(1.0);
 108     __m128d          two     = _mm_set1_pd(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_pd(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_pd(fr->ic->k_rf);
 123     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_pd(fr->ic->c_rf);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 129     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 130     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 131
 132     jq1              = _mm_set1_pd(charge[inr+1]);
 133     jq2              = _mm_set1_pd(charge[inr+2]);
 134     jq3              = _mm_set1_pd(charge[inr+3]);
 135     qq11             = _mm_mul_pd(iq1,jq1);
 136     qq12             = _mm_mul_pd(iq1,jq2);
 137     qq13             = _mm_mul_pd(iq1,jq3);
 138     qq21             = _mm_mul_pd(iq2,jq1);
 139     qq22             = _mm_mul_pd(iq2,jq2);
 140     qq23             = _mm_mul_pd(iq2,jq3);
 141     qq31             = _mm_mul_pd(iq3,jq1);
 142     qq32             = _mm_mul_pd(iq3,jq2);
 143     qq33             = _mm_mul_pd(iq3,jq3);
 144
 145     /* Avoid stupid compiler warnings */
 146     jnrA = jnrB = 0;
 147     j_coord_offsetA = 0;
 148     j_coord_offsetB = 0;
 149
 150     outeriter        = 0;
 151     inneriter        = 0;
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 169                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 170
 171         fix1             = _mm_setzero_pd();
 172         fiy1             = _mm_setzero_pd();
 173         fiz1             = _mm_setzero_pd();
 174         fix2             = _mm_setzero_pd();
 175         fiy2             = _mm_setzero_pd();
 176         fiz2             = _mm_setzero_pd();
 177         fix3             = _mm_setzero_pd();
 178         fiy3             = _mm_setzero_pd();
 179         fiz3             = _mm_setzero_pd();
 180
 181         /* Reset potential sums */
 182         velecsum         = _mm_setzero_pd();
 183
 184         /* Start inner kernel loop */
 185         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 186         {
 187
 188             /* Get j neighbor index, and coordinate index */
 189             jnrA             = jjnr[jidx];
 190             jnrB             = jjnr[jidx+1];
 191             j_coord_offsetA  = DIM*jnrA;
 192             j_coord_offsetB  = DIM*jnrB;
 193
 194             /* load j atom coordinates */
 195             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 196                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 197
 198             /* Calculate displacement vector */
 199             dx11             = _mm_sub_pd(ix1,jx1);
 200             dy11             = _mm_sub_pd(iy1,jy1);
 201             dz11             = _mm_sub_pd(iz1,jz1);
 202             dx12             = _mm_sub_pd(ix1,jx2);
 203             dy12             = _mm_sub_pd(iy1,jy2);
 204             dz12             = _mm_sub_pd(iz1,jz2);
 205             dx13             = _mm_sub_pd(ix1,jx3);
 206             dy13             = _mm_sub_pd(iy1,jy3);
 207             dz13             = _mm_sub_pd(iz1,jz3);
 208             dx21             = _mm_sub_pd(ix2,jx1);
 209             dy21             = _mm_sub_pd(iy2,jy1);
 210             dz21             = _mm_sub_pd(iz2,jz1);
 211             dx22             = _mm_sub_pd(ix2,jx2);
 212             dy22             = _mm_sub_pd(iy2,jy2);
 213             dz22             = _mm_sub_pd(iz2,jz2);
 214             dx23             = _mm_sub_pd(ix2,jx3);
 215             dy23             = _mm_sub_pd(iy2,jy3);
 216             dz23             = _mm_sub_pd(iz2,jz3);
 217             dx31             = _mm_sub_pd(ix3,jx1);
 218             dy31             = _mm_sub_pd(iy3,jy1);
 219             dz31             = _mm_sub_pd(iz3,jz1);
 220             dx32             = _mm_sub_pd(ix3,jx2);
 221             dy32             = _mm_sub_pd(iy3,jy2);
 222             dz32             = _mm_sub_pd(iz3,jz2);
 223             dx33             = _mm_sub_pd(ix3,jx3);
 224             dy33             = _mm_sub_pd(iy3,jy3);
 225             dz33             = _mm_sub_pd(iz3,jz3);
 226
 227             /* Calculate squared distance and things based on it */
 228             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 229             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 230             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 231             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 232             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 233             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 234             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 235             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 236             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 237
 238             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 239             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 240             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 241             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 242             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 243             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 244             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 245             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 246             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 247
 248             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 249             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 250             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 251             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 252             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 253             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 254             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 255             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 256             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 257
 258             fjx1             = _mm_setzero_pd();
 259             fjy1             = _mm_setzero_pd();
 260             fjz1             = _mm_setzero_pd();
 261             fjx2             = _mm_setzero_pd();
 262             fjy2             = _mm_setzero_pd();
 263             fjz2             = _mm_setzero_pd();
 264             fjx3             = _mm_setzero_pd();
 265             fjy3             = _mm_setzero_pd();
 266             fjz3             = _mm_setzero_pd();
 267
 268             /**************************
 269              * CALCULATE INTERACTIONS *
 270              **************************/
 271
 272             /* REACTION-FIELD ELECTROSTATICS */
 273             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 274             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 275
 276             /* Update potential sum for this i atom from the interaction with this j atom. */
 277             velecsum         = _mm_add_pd(velecsum,velec);
 278
 279             fscal            = felec;
 280
 281             /* Calculate temporary vectorial force */
 282             tx               = _mm_mul_pd(fscal,dx11);
 283             ty               = _mm_mul_pd(fscal,dy11);
 284             tz               = _mm_mul_pd(fscal,dz11);
 285
 286             /* Update vectorial force */
 287             fix1             = _mm_add_pd(fix1,tx);
 288             fiy1             = _mm_add_pd(fiy1,ty);
 289             fiz1             = _mm_add_pd(fiz1,tz);
 290
 291             fjx1             = _mm_add_pd(fjx1,tx);
 292             fjy1             = _mm_add_pd(fjy1,ty);
 293             fjz1             = _mm_add_pd(fjz1,tz);
 294
 295             /**************************
 296              * CALCULATE INTERACTIONS *
 297              **************************/
 298
 299             /* REACTION-FIELD ELECTROSTATICS */
 300             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 301             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 302
 303             /* Update potential sum for this i atom from the interaction with this j atom. */
 304             velecsum         = _mm_add_pd(velecsum,velec);
 305
 306             fscal            = felec;
 307
 308             /* Calculate temporary vectorial force */
 309             tx               = _mm_mul_pd(fscal,dx12);
 310             ty               = _mm_mul_pd(fscal,dy12);
 311             tz               = _mm_mul_pd(fscal,dz12);
 312
 313             /* Update vectorial force */
 314             fix1             = _mm_add_pd(fix1,tx);
 315             fiy1             = _mm_add_pd(fiy1,ty);
 316             fiz1             = _mm_add_pd(fiz1,tz);
 317
 318             fjx2             = _mm_add_pd(fjx2,tx);
 319             fjy2             = _mm_add_pd(fjy2,ty);
 320             fjz2             = _mm_add_pd(fjz2,tz);
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             /* REACTION-FIELD ELECTROSTATICS */
 327             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 328             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             velecsum         = _mm_add_pd(velecsum,velec);
 332
 333             fscal            = felec;
 334
 335             /* Calculate temporary vectorial force */
 336             tx               = _mm_mul_pd(fscal,dx13);
 337             ty               = _mm_mul_pd(fscal,dy13);
 338             tz               = _mm_mul_pd(fscal,dz13);
 339
 340             /* Update vectorial force */
 341             fix1             = _mm_add_pd(fix1,tx);
 342             fiy1             = _mm_add_pd(fiy1,ty);
 343             fiz1             = _mm_add_pd(fiz1,tz);
 344
 345             fjx3             = _mm_add_pd(fjx3,tx);
 346             fjy3             = _mm_add_pd(fjy3,ty);
 347             fjz3             = _mm_add_pd(fjz3,tz);
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             /* REACTION-FIELD ELECTROSTATICS */
 354             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 355             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             velecsum         = _mm_add_pd(velecsum,velec);
 359
 360             fscal            = felec;
 361
 362             /* Calculate temporary vectorial force */
 363             tx               = _mm_mul_pd(fscal,dx21);
 364             ty               = _mm_mul_pd(fscal,dy21);
 365             tz               = _mm_mul_pd(fscal,dz21);
 366
 367             /* Update vectorial force */
 368             fix2             = _mm_add_pd(fix2,tx);
 369             fiy2             = _mm_add_pd(fiy2,ty);
 370             fiz2             = _mm_add_pd(fiz2,tz);
 371
 372             fjx1             = _mm_add_pd(fjx1,tx);
 373             fjy1             = _mm_add_pd(fjy1,ty);
 374             fjz1             = _mm_add_pd(fjz1,tz);
 375
 376             /**************************
 377              * CALCULATE INTERACTIONS *
 378              **************************/
 379
 380             /* REACTION-FIELD ELECTROSTATICS */
 381             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 382             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 383
 384             /* Update potential sum for this i atom from the interaction with this j atom. */
 385             velecsum         = _mm_add_pd(velecsum,velec);
 386
 387             fscal            = felec;
 388
 389             /* Calculate temporary vectorial force */
 390             tx               = _mm_mul_pd(fscal,dx22);
 391             ty               = _mm_mul_pd(fscal,dy22);
 392             tz               = _mm_mul_pd(fscal,dz22);
 393
 394             /* Update vectorial force */
 395             fix2             = _mm_add_pd(fix2,tx);
 396             fiy2             = _mm_add_pd(fiy2,ty);
 397             fiz2             = _mm_add_pd(fiz2,tz);
 398
 399             fjx2             = _mm_add_pd(fjx2,tx);
 400             fjy2             = _mm_add_pd(fjy2,ty);
 401             fjz2             = _mm_add_pd(fjz2,tz);
 402
 403             /**************************
 404              * CALCULATE INTERACTIONS *
 405              **************************/
 406
 407             /* REACTION-FIELD ELECTROSTATICS */
 408             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 409             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velecsum         = _mm_add_pd(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             /* Calculate temporary vectorial force */
 417             tx               = _mm_mul_pd(fscal,dx23);
 418             ty               = _mm_mul_pd(fscal,dy23);
 419             tz               = _mm_mul_pd(fscal,dz23);
 420
 421             /* Update vectorial force */
 422             fix2             = _mm_add_pd(fix2,tx);
 423             fiy2             = _mm_add_pd(fiy2,ty);
 424             fiz2             = _mm_add_pd(fiz2,tz);
 425
 426             fjx3             = _mm_add_pd(fjx3,tx);
 427             fjy3             = _mm_add_pd(fjy3,ty);
 428             fjz3             = _mm_add_pd(fjz3,tz);
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             /* REACTION-FIELD ELECTROSTATICS */
 435             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 436             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 437
 438             /* Update potential sum for this i atom from the interaction with this j atom. */
 439             velecsum         = _mm_add_pd(velecsum,velec);
 440
 441             fscal            = felec;
 442
 443             /* Calculate temporary vectorial force */
 444             tx               = _mm_mul_pd(fscal,dx31);
 445             ty               = _mm_mul_pd(fscal,dy31);
 446             tz               = _mm_mul_pd(fscal,dz31);
 447
 448             /* Update vectorial force */
 449             fix3             = _mm_add_pd(fix3,tx);
 450             fiy3             = _mm_add_pd(fiy3,ty);
 451             fiz3             = _mm_add_pd(fiz3,tz);
 452
 453             fjx1             = _mm_add_pd(fjx1,tx);
 454             fjy1             = _mm_add_pd(fjy1,ty);
 455             fjz1             = _mm_add_pd(fjz1,tz);
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             /* REACTION-FIELD ELECTROSTATICS */
 462             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 463             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 464
 465             /* Update potential sum for this i atom from the interaction with this j atom. */
 466             velecsum         = _mm_add_pd(velecsum,velec);
 467
 468             fscal            = felec;
 469
 470             /* Calculate temporary vectorial force */
 471             tx               = _mm_mul_pd(fscal,dx32);
 472             ty               = _mm_mul_pd(fscal,dy32);
 473             tz               = _mm_mul_pd(fscal,dz32);
 474
 475             /* Update vectorial force */
 476             fix3             = _mm_add_pd(fix3,tx);
 477             fiy3             = _mm_add_pd(fiy3,ty);
 478             fiz3             = _mm_add_pd(fiz3,tz);
 479
 480             fjx2             = _mm_add_pd(fjx2,tx);
 481             fjy2             = _mm_add_pd(fjy2,ty);
 482             fjz2             = _mm_add_pd(fjz2,tz);
 483
 484             /**************************
 485              * CALCULATE INTERACTIONS *
 486              **************************/
 487
 488             /* REACTION-FIELD ELECTROSTATICS */
 489             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 490             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 491
 492             /* Update potential sum for this i atom from the interaction with this j atom. */
 493             velecsum         = _mm_add_pd(velecsum,velec);
 494
 495             fscal            = felec;
 496
 497             /* Calculate temporary vectorial force */
 498             tx               = _mm_mul_pd(fscal,dx33);
 499             ty               = _mm_mul_pd(fscal,dy33);
 500             tz               = _mm_mul_pd(fscal,dz33);
 501
 502             /* Update vectorial force */
 503             fix3             = _mm_add_pd(fix3,tx);
 504             fiy3             = _mm_add_pd(fiy3,ty);
 505             fiz3             = _mm_add_pd(fiz3,tz);
 506
 507             fjx3             = _mm_add_pd(fjx3,tx);
 508             fjy3             = _mm_add_pd(fjy3,ty);
 509             fjz3             = _mm_add_pd(fjz3,tz);
 510
 511             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 512
 513             /* Inner loop uses 288 flops */
 514         }
 515
 516         if(jidx<j_index_end)
 517         {
 518
 519             jnrA             = jjnr[jidx];
 520             j_coord_offsetA  = DIM*jnrA;
 521
 522             /* load j atom coordinates */
 523             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 524                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 525
 526             /* Calculate displacement vector */
 527             dx11             = _mm_sub_pd(ix1,jx1);
 528             dy11             = _mm_sub_pd(iy1,jy1);
 529             dz11             = _mm_sub_pd(iz1,jz1);
 530             dx12             = _mm_sub_pd(ix1,jx2);
 531             dy12             = _mm_sub_pd(iy1,jy2);
 532             dz12             = _mm_sub_pd(iz1,jz2);
 533             dx13             = _mm_sub_pd(ix1,jx3);
 534             dy13             = _mm_sub_pd(iy1,jy3);
 535             dz13             = _mm_sub_pd(iz1,jz3);
 536             dx21             = _mm_sub_pd(ix2,jx1);
 537             dy21             = _mm_sub_pd(iy2,jy1);
 538             dz21             = _mm_sub_pd(iz2,jz1);
 539             dx22             = _mm_sub_pd(ix2,jx2);
 540             dy22             = _mm_sub_pd(iy2,jy2);
 541             dz22             = _mm_sub_pd(iz2,jz2);
 542             dx23             = _mm_sub_pd(ix2,jx3);
 543             dy23             = _mm_sub_pd(iy2,jy3);
 544             dz23             = _mm_sub_pd(iz2,jz3);
 545             dx31             = _mm_sub_pd(ix3,jx1);
 546             dy31             = _mm_sub_pd(iy3,jy1);
 547             dz31             = _mm_sub_pd(iz3,jz1);
 548             dx32             = _mm_sub_pd(ix3,jx2);
 549             dy32             = _mm_sub_pd(iy3,jy2);
 550             dz32             = _mm_sub_pd(iz3,jz2);
 551             dx33             = _mm_sub_pd(ix3,jx3);
 552             dy33             = _mm_sub_pd(iy3,jy3);
 553             dz33             = _mm_sub_pd(iz3,jz3);
 554
 555             /* Calculate squared distance and things based on it */
 556             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 557             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 558             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 559             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 560             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 561             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 562             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 563             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 564             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 565
 566             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 567             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 568             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 569             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 570             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 571             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 572             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 573             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 574             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 575
 576             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 577             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 578             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 579             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 580             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 581             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 582             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 583             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 584             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 585
 586             fjx1             = _mm_setzero_pd();
 587             fjy1             = _mm_setzero_pd();
 588             fjz1             = _mm_setzero_pd();
 589             fjx2             = _mm_setzero_pd();
 590             fjy2             = _mm_setzero_pd();
 591             fjz2             = _mm_setzero_pd();
 592             fjx3             = _mm_setzero_pd();
 593             fjy3             = _mm_setzero_pd();
 594             fjz3             = _mm_setzero_pd();
 595
 596             /**************************
 597              * CALCULATE INTERACTIONS *
 598              **************************/
 599
 600             /* REACTION-FIELD ELECTROSTATICS */
 601             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 602             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 603
 604             /* Update potential sum for this i atom from the interaction with this j atom. */
 605             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 606             velecsum         = _mm_add_pd(velecsum,velec);
 607
 608             fscal            = felec;
 609
 610             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 611
 612             /* Calculate temporary vectorial force */
 613             tx               = _mm_mul_pd(fscal,dx11);
 614             ty               = _mm_mul_pd(fscal,dy11);
 615             tz               = _mm_mul_pd(fscal,dz11);
 616
 617             /* Update vectorial force */
 618             fix1             = _mm_add_pd(fix1,tx);
 619             fiy1             = _mm_add_pd(fiy1,ty);
 620             fiz1             = _mm_add_pd(fiz1,tz);
 621
 622             fjx1             = _mm_add_pd(fjx1,tx);
 623             fjy1             = _mm_add_pd(fjy1,ty);
 624             fjz1             = _mm_add_pd(fjz1,tz);
 625
 626             /**************************
 627              * CALCULATE INTERACTIONS *
 628              **************************/
 629
 630             /* REACTION-FIELD ELECTROSTATICS */
 631             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 632             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 633
 634             /* Update potential sum for this i atom from the interaction with this j atom. */
 635             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 636             velecsum         = _mm_add_pd(velecsum,velec);
 637
 638             fscal            = felec;
 639
 640             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 641
 642             /* Calculate temporary vectorial force */
 643             tx               = _mm_mul_pd(fscal,dx12);
 644             ty               = _mm_mul_pd(fscal,dy12);
 645             tz               = _mm_mul_pd(fscal,dz12);
 646
 647             /* Update vectorial force */
 648             fix1             = _mm_add_pd(fix1,tx);
 649             fiy1             = _mm_add_pd(fiy1,ty);
 650             fiz1             = _mm_add_pd(fiz1,tz);
 651
 652             fjx2             = _mm_add_pd(fjx2,tx);
 653             fjy2             = _mm_add_pd(fjy2,ty);
 654             fjz2             = _mm_add_pd(fjz2,tz);
 655
 656             /**************************
 657              * CALCULATE INTERACTIONS *
 658              **************************/
 659
 660             /* REACTION-FIELD ELECTROSTATICS */
 661             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 662             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 663
 664             /* Update potential sum for this i atom from the interaction with this j atom. */
 665             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 666             velecsum         = _mm_add_pd(velecsum,velec);
 667
 668             fscal            = felec;
 669
 670             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 671
 672             /* Calculate temporary vectorial force */
 673             tx               = _mm_mul_pd(fscal,dx13);
 674             ty               = _mm_mul_pd(fscal,dy13);
 675             tz               = _mm_mul_pd(fscal,dz13);
 676
 677             /* Update vectorial force */
 678             fix1             = _mm_add_pd(fix1,tx);
 679             fiy1             = _mm_add_pd(fiy1,ty);
 680             fiz1             = _mm_add_pd(fiz1,tz);
 681
 682             fjx3             = _mm_add_pd(fjx3,tx);
 683             fjy3             = _mm_add_pd(fjy3,ty);
 684             fjz3             = _mm_add_pd(fjz3,tz);
 685
 686             /**************************
 687              * CALCULATE INTERACTIONS *
 688              **************************/
 689
 690             /* REACTION-FIELD ELECTROSTATICS */
 691             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 692             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 693
 694             /* Update potential sum for this i atom from the interaction with this j atom. */
 695             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 696             velecsum         = _mm_add_pd(velecsum,velec);
 697
 698             fscal            = felec;
 699
 700             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 701
 702             /* Calculate temporary vectorial force */
 703             tx               = _mm_mul_pd(fscal,dx21);
 704             ty               = _mm_mul_pd(fscal,dy21);
 705             tz               = _mm_mul_pd(fscal,dz21);
 706
 707             /* Update vectorial force */
 708             fix2             = _mm_add_pd(fix2,tx);
 709             fiy2             = _mm_add_pd(fiy2,ty);
 710             fiz2             = _mm_add_pd(fiz2,tz);
 711
 712             fjx1             = _mm_add_pd(fjx1,tx);
 713             fjy1             = _mm_add_pd(fjy1,ty);
 714             fjz1             = _mm_add_pd(fjz1,tz);
 715
 716             /**************************
 717              * CALCULATE INTERACTIONS *
 718              **************************/
 719
 720             /* REACTION-FIELD ELECTROSTATICS */
 721             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 722             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 723
 724             /* Update potential sum for this i atom from the interaction with this j atom. */
 725             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 726             velecsum         = _mm_add_pd(velecsum,velec);
 727
 728             fscal            = felec;
 729
 730             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 731
 732             /* Calculate temporary vectorial force */
 733             tx               = _mm_mul_pd(fscal,dx22);
 734             ty               = _mm_mul_pd(fscal,dy22);
 735             tz               = _mm_mul_pd(fscal,dz22);
 736
 737             /* Update vectorial force */
 738             fix2             = _mm_add_pd(fix2,tx);
 739             fiy2             = _mm_add_pd(fiy2,ty);
 740             fiz2             = _mm_add_pd(fiz2,tz);
 741
 742             fjx2             = _mm_add_pd(fjx2,tx);
 743             fjy2             = _mm_add_pd(fjy2,ty);
 744             fjz2             = _mm_add_pd(fjz2,tz);
 745
 746             /**************************
 747              * CALCULATE INTERACTIONS *
 748              **************************/
 749
 750             /* REACTION-FIELD ELECTROSTATICS */
 751             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 752             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 753
 754             /* Update potential sum for this i atom from the interaction with this j atom. */
 755             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 756             velecsum         = _mm_add_pd(velecsum,velec);
 757
 758             fscal            = felec;
 759
 760             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 761
 762             /* Calculate temporary vectorial force */
 763             tx               = _mm_mul_pd(fscal,dx23);
 764             ty               = _mm_mul_pd(fscal,dy23);
 765             tz               = _mm_mul_pd(fscal,dz23);
 766
 767             /* Update vectorial force */
 768             fix2             = _mm_add_pd(fix2,tx);
 769             fiy2             = _mm_add_pd(fiy2,ty);
 770             fiz2             = _mm_add_pd(fiz2,tz);
 771
 772             fjx3             = _mm_add_pd(fjx3,tx);
 773             fjy3             = _mm_add_pd(fjy3,ty);
 774             fjz3             = _mm_add_pd(fjz3,tz);
 775
 776             /**************************
 777              * CALCULATE INTERACTIONS *
 778              **************************/
 779
 780             /* REACTION-FIELD ELECTROSTATICS */
 781             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 782             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 783
 784             /* Update potential sum for this i atom from the interaction with this j atom. */
 785             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 786             velecsum         = _mm_add_pd(velecsum,velec);
 787
 788             fscal            = felec;
 789
 790             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 791
 792             /* Calculate temporary vectorial force */
 793             tx               = _mm_mul_pd(fscal,dx31);
 794             ty               = _mm_mul_pd(fscal,dy31);
 795             tz               = _mm_mul_pd(fscal,dz31);
 796
 797             /* Update vectorial force */
 798             fix3             = _mm_add_pd(fix3,tx);
 799             fiy3             = _mm_add_pd(fiy3,ty);
 800             fiz3             = _mm_add_pd(fiz3,tz);
 801
 802             fjx1             = _mm_add_pd(fjx1,tx);
 803             fjy1             = _mm_add_pd(fjy1,ty);
 804             fjz1             = _mm_add_pd(fjz1,tz);
 805
 806             /**************************
 807              * CALCULATE INTERACTIONS *
 808              **************************/
 809
 810             /* REACTION-FIELD ELECTROSTATICS */
 811             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 812             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 813
 814             /* Update potential sum for this i atom from the interaction with this j atom. */
 815             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 816             velecsum         = _mm_add_pd(velecsum,velec);
 817
 818             fscal            = felec;
 819
 820             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 821
 822             /* Calculate temporary vectorial force */
 823             tx               = _mm_mul_pd(fscal,dx32);
 824             ty               = _mm_mul_pd(fscal,dy32);
 825             tz               = _mm_mul_pd(fscal,dz32);
 826
 827             /* Update vectorial force */
 828             fix3             = _mm_add_pd(fix3,tx);
 829             fiy3             = _mm_add_pd(fiy3,ty);
 830             fiz3             = _mm_add_pd(fiz3,tz);
 831
 832             fjx2             = _mm_add_pd(fjx2,tx);
 833             fjy2             = _mm_add_pd(fjy2,ty);
 834             fjz2             = _mm_add_pd(fjz2,tz);
 835
 836             /**************************
 837              * CALCULATE INTERACTIONS *
 838              **************************/
 839
 840             /* REACTION-FIELD ELECTROSTATICS */
 841             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 842             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 843
 844             /* Update potential sum for this i atom from the interaction with this j atom. */
 845             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 846             velecsum         = _mm_add_pd(velecsum,velec);
 847
 848             fscal            = felec;
 849
 850             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 851
 852             /* Calculate temporary vectorial force */
 853             tx               = _mm_mul_pd(fscal,dx33);
 854             ty               = _mm_mul_pd(fscal,dy33);
 855             tz               = _mm_mul_pd(fscal,dz33);
 856
 857             /* Update vectorial force */
 858             fix3             = _mm_add_pd(fix3,tx);
 859             fiy3             = _mm_add_pd(fiy3,ty);
 860             fiz3             = _mm_add_pd(fiz3,tz);
 861
 862             fjx3             = _mm_add_pd(fjx3,tx);
 863             fjy3             = _mm_add_pd(fjy3,ty);
 864             fjz3             = _mm_add_pd(fjz3,tz);
 865
 866             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 867
 868             /* Inner loop uses 288 flops */
 869         }
 870
 871         /* End of innermost loop */
 872
 873         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 874                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 875
 876         ggid                        = gid[iidx];
 877         /* Update potential energies */
 878         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 879
 880         /* Increment number of inner iterations */
 881         inneriter                  += j_index_end - j_index_start;
 882
 883         /* Outer loop uses 19 flops */
 884     }
 885
 886     /* Increment number of outer iterations */
 887     outeriter        += nri;
 888
 889     /* Update outer/inner flops */
 890
 891     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*288);
 892 }
 893 /*
 894  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse4_1_double
 895  * Electrostatics interaction: ReactionField
 896  * VdW interaction:            None
 897  * Geometry:                   Water4-Water4
 898  * Calculate force/pot:        Force
 899  */
 900 void
 901 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse4_1_double
 902                     (t_nblist                    * gmx_restrict       nlist,
 903                      rvec                        * gmx_restrict          xx,
 904                      rvec                        * gmx_restrict          ff,
 905                      t_forcerec                  * gmx_restrict          fr,
 906                      t_mdatoms                   * gmx_restrict     mdatoms,
 907                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 908                      t_nrnb                      * gmx_restrict        nrnb)
 909 {
 910     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 911      * just 0 for non-waters.
 912      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 913      * jnr indices corresponding to data put in the four positions in the SIMD register.
 914      */
 915     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 916     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 917     int              jnrA,jnrB;
 918     int              j_coord_offsetA,j_coord_offsetB;
 919     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 920     real             rcutoff_scalar;
 921     real             *shiftvec,*fshift,*x,*f;
 922     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 923     int              vdwioffset1;
 924     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 925     int              vdwioffset2;
 926     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 927     int              vdwioffset3;
 928     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 929     int              vdwjidx1A,vdwjidx1B;
 930     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 931     int              vdwjidx2A,vdwjidx2B;
 932     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 933     int              vdwjidx3A,vdwjidx3B;
 934     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 935     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 936     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 937     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 938     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 939     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 940     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 941     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 942     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 943     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 944     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 945     real             *charge;
 946     __m128d          dummy_mask,cutoff_mask;
 947     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 948     __m128d          one     = _mm_set1_pd(1.0);
 949     __m128d          two     = _mm_set1_pd(2.0);
 950     x                = xx[0];
 951     f                = ff[0];
 952
 953     nri              = nlist->nri;
 954     iinr             = nlist->iinr;
 955     jindex           = nlist->jindex;
 956     jjnr             = nlist->jjnr;
 957     shiftidx         = nlist->shift;
 958     gid              = nlist->gid;
 959     shiftvec         = fr->shift_vec[0];
 960     fshift           = fr->fshift[0];
 961     facel            = _mm_set1_pd(fr->epsfac);
 962     charge           = mdatoms->chargeA;
 963     krf              = _mm_set1_pd(fr->ic->k_rf);
 964     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 965     crf              = _mm_set1_pd(fr->ic->c_rf);
 966
 967     /* Setup water-specific parameters */
 968     inr              = nlist->iinr[0];
 969     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 970     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 971     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 972
 973     jq1              = _mm_set1_pd(charge[inr+1]);
 974     jq2              = _mm_set1_pd(charge[inr+2]);
 975     jq3              = _mm_set1_pd(charge[inr+3]);
 976     qq11             = _mm_mul_pd(iq1,jq1);
 977     qq12             = _mm_mul_pd(iq1,jq2);
 978     qq13             = _mm_mul_pd(iq1,jq3);
 979     qq21             = _mm_mul_pd(iq2,jq1);
 980     qq22             = _mm_mul_pd(iq2,jq2);
 981     qq23             = _mm_mul_pd(iq2,jq3);
 982     qq31             = _mm_mul_pd(iq3,jq1);
 983     qq32             = _mm_mul_pd(iq3,jq2);
 984     qq33             = _mm_mul_pd(iq3,jq3);
 985
 986     /* Avoid stupid compiler warnings */
 987     jnrA = jnrB = 0;
 988     j_coord_offsetA = 0;
 989     j_coord_offsetB = 0;
 990
 991     outeriter        = 0;
 992     inneriter        = 0;
 993
 994     /* Start outer loop over neighborlists */
 995     for(iidx=0; iidx<nri; iidx++)
 996     {
 997         /* Load shift vector for this list */
 998         i_shift_offset   = DIM*shiftidx[iidx];
 999
1000         /* Load limits for loop over neighbors */
1001         j_index_start    = jindex[iidx];
1002         j_index_end      = jindex[iidx+1];
1003
1004         /* Get outer coordinate index */
1005         inr              = iinr[iidx];
1006         i_coord_offset   = DIM*inr;
1007
1008         /* Load i particle coords and add shift vector */
1009         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1010                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1011
1012         fix1             = _mm_setzero_pd();
1013         fiy1             = _mm_setzero_pd();
1014         fiz1             = _mm_setzero_pd();
1015         fix2             = _mm_setzero_pd();
1016         fiy2             = _mm_setzero_pd();
1017         fiz2             = _mm_setzero_pd();
1018         fix3             = _mm_setzero_pd();
1019         fiy3             = _mm_setzero_pd();
1020         fiz3             = _mm_setzero_pd();
1021
1022         /* Start inner kernel loop */
1023         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1024         {
1025
1026             /* Get j neighbor index, and coordinate index */
1027             jnrA             = jjnr[jidx];
1028             jnrB             = jjnr[jidx+1];
1029             j_coord_offsetA  = DIM*jnrA;
1030             j_coord_offsetB  = DIM*jnrB;
1031
1032             /* load j atom coordinates */
1033             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1034                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1035
1036             /* Calculate displacement vector */
1037             dx11             = _mm_sub_pd(ix1,jx1);
1038             dy11             = _mm_sub_pd(iy1,jy1);
1039             dz11             = _mm_sub_pd(iz1,jz1);
1040             dx12             = _mm_sub_pd(ix1,jx2);
1041             dy12             = _mm_sub_pd(iy1,jy2);
1042             dz12             = _mm_sub_pd(iz1,jz2);
1043             dx13             = _mm_sub_pd(ix1,jx3);
1044             dy13             = _mm_sub_pd(iy1,jy3);
1045             dz13             = _mm_sub_pd(iz1,jz3);
1046             dx21             = _mm_sub_pd(ix2,jx1);
1047             dy21             = _mm_sub_pd(iy2,jy1);
1048             dz21             = _mm_sub_pd(iz2,jz1);
1049             dx22             = _mm_sub_pd(ix2,jx2);
1050             dy22             = _mm_sub_pd(iy2,jy2);
1051             dz22             = _mm_sub_pd(iz2,jz2);
1052             dx23             = _mm_sub_pd(ix2,jx3);
1053             dy23             = _mm_sub_pd(iy2,jy3);
1054             dz23             = _mm_sub_pd(iz2,jz3);
1055             dx31             = _mm_sub_pd(ix3,jx1);
1056             dy31             = _mm_sub_pd(iy3,jy1);
1057             dz31             = _mm_sub_pd(iz3,jz1);
1058             dx32             = _mm_sub_pd(ix3,jx2);
1059             dy32             = _mm_sub_pd(iy3,jy2);
1060             dz32             = _mm_sub_pd(iz3,jz2);
1061             dx33             = _mm_sub_pd(ix3,jx3);
1062             dy33             = _mm_sub_pd(iy3,jy3);
1063             dz33             = _mm_sub_pd(iz3,jz3);
1064
1065             /* Calculate squared distance and things based on it */
1066             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1067             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1068             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1069             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1070             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1071             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1072             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1073             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1074             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1075
1076             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1077             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1078             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1079             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1080             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1081             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1082             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1083             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1084             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1085
1086             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1087             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1088             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1089             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1090             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1091             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1092             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1093             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1094             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1095
1096             fjx1             = _mm_setzero_pd();
1097             fjy1             = _mm_setzero_pd();
1098             fjz1             = _mm_setzero_pd();
1099             fjx2             = _mm_setzero_pd();
1100             fjy2             = _mm_setzero_pd();
1101             fjz2             = _mm_setzero_pd();
1102             fjx3             = _mm_setzero_pd();
1103             fjy3             = _mm_setzero_pd();
1104             fjz3             = _mm_setzero_pd();
1105
1106             /**************************
1107              * CALCULATE INTERACTIONS *
1108              **************************/
1109
1110             /* REACTION-FIELD ELECTROSTATICS */
1111             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1112
1113             fscal            = felec;
1114
1115             /* Calculate temporary vectorial force */
1116             tx               = _mm_mul_pd(fscal,dx11);
1117             ty               = _mm_mul_pd(fscal,dy11);
1118             tz               = _mm_mul_pd(fscal,dz11);
1119
1120             /* Update vectorial force */
1121             fix1             = _mm_add_pd(fix1,tx);
1122             fiy1             = _mm_add_pd(fiy1,ty);
1123             fiz1             = _mm_add_pd(fiz1,tz);
1124
1125             fjx1             = _mm_add_pd(fjx1,tx);
1126             fjy1             = _mm_add_pd(fjy1,ty);
1127             fjz1             = _mm_add_pd(fjz1,tz);
1128
1129             /**************************
1130              * CALCULATE INTERACTIONS *
1131              **************************/
1132
1133             /* REACTION-FIELD ELECTROSTATICS */
1134             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1135
1136             fscal            = felec;
1137
1138             /* Calculate temporary vectorial force */
1139             tx               = _mm_mul_pd(fscal,dx12);
1140             ty               = _mm_mul_pd(fscal,dy12);
1141             tz               = _mm_mul_pd(fscal,dz12);
1142
1143             /* Update vectorial force */
1144             fix1             = _mm_add_pd(fix1,tx);
1145             fiy1             = _mm_add_pd(fiy1,ty);
1146             fiz1             = _mm_add_pd(fiz1,tz);
1147
1148             fjx2             = _mm_add_pd(fjx2,tx);
1149             fjy2             = _mm_add_pd(fjy2,ty);
1150             fjz2             = _mm_add_pd(fjz2,tz);
1151
1152             /**************************
1153              * CALCULATE INTERACTIONS *
1154              **************************/
1155
1156             /* REACTION-FIELD ELECTROSTATICS */
1157             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1158
1159             fscal            = felec;
1160
1161             /* Calculate temporary vectorial force */
1162             tx               = _mm_mul_pd(fscal,dx13);
1163             ty               = _mm_mul_pd(fscal,dy13);
1164             tz               = _mm_mul_pd(fscal,dz13);
1165
1166             /* Update vectorial force */
1167             fix1             = _mm_add_pd(fix1,tx);
1168             fiy1             = _mm_add_pd(fiy1,ty);
1169             fiz1             = _mm_add_pd(fiz1,tz);
1170
1171             fjx3             = _mm_add_pd(fjx3,tx);
1172             fjy3             = _mm_add_pd(fjy3,ty);
1173             fjz3             = _mm_add_pd(fjz3,tz);
1174
1175             /**************************
1176              * CALCULATE INTERACTIONS *
1177              **************************/
1178
1179             /* REACTION-FIELD ELECTROSTATICS */
1180             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1181
1182             fscal            = felec;
1183
1184             /* Calculate temporary vectorial force */
1185             tx               = _mm_mul_pd(fscal,dx21);
1186             ty               = _mm_mul_pd(fscal,dy21);
1187             tz               = _mm_mul_pd(fscal,dz21);
1188
1189             /* Update vectorial force */
1190             fix2             = _mm_add_pd(fix2,tx);
1191             fiy2             = _mm_add_pd(fiy2,ty);
1192             fiz2             = _mm_add_pd(fiz2,tz);
1193
1194             fjx1             = _mm_add_pd(fjx1,tx);
1195             fjy1             = _mm_add_pd(fjy1,ty);
1196             fjz1             = _mm_add_pd(fjz1,tz);
1197
1198             /**************************
1199              * CALCULATE INTERACTIONS *
1200              **************************/
1201
1202             /* REACTION-FIELD ELECTROSTATICS */
1203             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1204
1205             fscal            = felec;
1206
1207             /* Calculate temporary vectorial force */
1208             tx               = _mm_mul_pd(fscal,dx22);
1209             ty               = _mm_mul_pd(fscal,dy22);
1210             tz               = _mm_mul_pd(fscal,dz22);
1211
1212             /* Update vectorial force */
1213             fix2             = _mm_add_pd(fix2,tx);
1214             fiy2             = _mm_add_pd(fiy2,ty);
1215             fiz2             = _mm_add_pd(fiz2,tz);
1216
1217             fjx2             = _mm_add_pd(fjx2,tx);
1218             fjy2             = _mm_add_pd(fjy2,ty);
1219             fjz2             = _mm_add_pd(fjz2,tz);
1220
1221             /**************************
1222              * CALCULATE INTERACTIONS *
1223              **************************/
1224
1225             /* REACTION-FIELD ELECTROSTATICS */
1226             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1227
1228             fscal            = felec;
1229
1230             /* Calculate temporary vectorial force */
1231             tx               = _mm_mul_pd(fscal,dx23);
1232             ty               = _mm_mul_pd(fscal,dy23);
1233             tz               = _mm_mul_pd(fscal,dz23);
1234
1235             /* Update vectorial force */
1236             fix2             = _mm_add_pd(fix2,tx);
1237             fiy2             = _mm_add_pd(fiy2,ty);
1238             fiz2             = _mm_add_pd(fiz2,tz);
1239
1240             fjx3             = _mm_add_pd(fjx3,tx);
1241             fjy3             = _mm_add_pd(fjy3,ty);
1242             fjz3             = _mm_add_pd(fjz3,tz);
1243
1244             /**************************
1245              * CALCULATE INTERACTIONS *
1246              **************************/
1247
1248             /* REACTION-FIELD ELECTROSTATICS */
1249             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1250
1251             fscal            = felec;
1252
1253             /* Calculate temporary vectorial force */
1254             tx               = _mm_mul_pd(fscal,dx31);
1255             ty               = _mm_mul_pd(fscal,dy31);
1256             tz               = _mm_mul_pd(fscal,dz31);
1257
1258             /* Update vectorial force */
1259             fix3             = _mm_add_pd(fix3,tx);
1260             fiy3             = _mm_add_pd(fiy3,ty);
1261             fiz3             = _mm_add_pd(fiz3,tz);
1262
1263             fjx1             = _mm_add_pd(fjx1,tx);
1264             fjy1             = _mm_add_pd(fjy1,ty);
1265             fjz1             = _mm_add_pd(fjz1,tz);
1266
1267             /**************************
1268              * CALCULATE INTERACTIONS *
1269              **************************/
1270
1271             /* REACTION-FIELD ELECTROSTATICS */
1272             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1273
1274             fscal            = felec;
1275
1276             /* Calculate temporary vectorial force */
1277             tx               = _mm_mul_pd(fscal,dx32);
1278             ty               = _mm_mul_pd(fscal,dy32);
1279             tz               = _mm_mul_pd(fscal,dz32);
1280
1281             /* Update vectorial force */
1282             fix3             = _mm_add_pd(fix3,tx);
1283             fiy3             = _mm_add_pd(fiy3,ty);
1284             fiz3             = _mm_add_pd(fiz3,tz);
1285
1286             fjx2             = _mm_add_pd(fjx2,tx);
1287             fjy2             = _mm_add_pd(fjy2,ty);
1288             fjz2             = _mm_add_pd(fjz2,tz);
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             /* REACTION-FIELD ELECTROSTATICS */
1295             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1296
1297             fscal            = felec;
1298
1299             /* Calculate temporary vectorial force */
1300             tx               = _mm_mul_pd(fscal,dx33);
1301             ty               = _mm_mul_pd(fscal,dy33);
1302             tz               = _mm_mul_pd(fscal,dz33);
1303
1304             /* Update vectorial force */
1305             fix3             = _mm_add_pd(fix3,tx);
1306             fiy3             = _mm_add_pd(fiy3,ty);
1307             fiz3             = _mm_add_pd(fiz3,tz);
1308
1309             fjx3             = _mm_add_pd(fjx3,tx);
1310             fjy3             = _mm_add_pd(fjy3,ty);
1311             fjz3             = _mm_add_pd(fjz3,tz);
1312
1313             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1314
1315             /* Inner loop uses 243 flops */
1316         }
1317
1318         if(jidx<j_index_end)
1319         {
1320
1321             jnrA             = jjnr[jidx];
1322             j_coord_offsetA  = DIM*jnrA;
1323
1324             /* load j atom coordinates */
1325             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1326                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1327
1328             /* Calculate displacement vector */
1329             dx11             = _mm_sub_pd(ix1,jx1);
1330             dy11             = _mm_sub_pd(iy1,jy1);
1331             dz11             = _mm_sub_pd(iz1,jz1);
1332             dx12             = _mm_sub_pd(ix1,jx2);
1333             dy12             = _mm_sub_pd(iy1,jy2);
1334             dz12             = _mm_sub_pd(iz1,jz2);
1335             dx13             = _mm_sub_pd(ix1,jx3);
1336             dy13             = _mm_sub_pd(iy1,jy3);
1337             dz13             = _mm_sub_pd(iz1,jz3);
1338             dx21             = _mm_sub_pd(ix2,jx1);
1339             dy21             = _mm_sub_pd(iy2,jy1);
1340             dz21             = _mm_sub_pd(iz2,jz1);
1341             dx22             = _mm_sub_pd(ix2,jx2);
1342             dy22             = _mm_sub_pd(iy2,jy2);
1343             dz22             = _mm_sub_pd(iz2,jz2);
1344             dx23             = _mm_sub_pd(ix2,jx3);
1345             dy23             = _mm_sub_pd(iy2,jy3);
1346             dz23             = _mm_sub_pd(iz2,jz3);
1347             dx31             = _mm_sub_pd(ix3,jx1);
1348             dy31             = _mm_sub_pd(iy3,jy1);
1349             dz31             = _mm_sub_pd(iz3,jz1);
1350             dx32             = _mm_sub_pd(ix3,jx2);
1351             dy32             = _mm_sub_pd(iy3,jy2);
1352             dz32             = _mm_sub_pd(iz3,jz2);
1353             dx33             = _mm_sub_pd(ix3,jx3);
1354             dy33             = _mm_sub_pd(iy3,jy3);
1355             dz33             = _mm_sub_pd(iz3,jz3);
1356
1357             /* Calculate squared distance and things based on it */
1358             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1359             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1360             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1361             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1362             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1363             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1364             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1365             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1366             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1367
1368             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1369             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1370             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1371             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1372             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1373             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1374             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1375             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1376             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1377
1378             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1379             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1380             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1381             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1382             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1383             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1384             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1385             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1386             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1387
1388             fjx1             = _mm_setzero_pd();
1389             fjy1             = _mm_setzero_pd();
1390             fjz1             = _mm_setzero_pd();
1391             fjx2             = _mm_setzero_pd();
1392             fjy2             = _mm_setzero_pd();
1393             fjz2             = _mm_setzero_pd();
1394             fjx3             = _mm_setzero_pd();
1395             fjy3             = _mm_setzero_pd();
1396             fjz3             = _mm_setzero_pd();
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             /* REACTION-FIELD ELECTROSTATICS */
1403             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1404
1405             fscal            = felec;
1406
1407             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1408
1409             /* Calculate temporary vectorial force */
1410             tx               = _mm_mul_pd(fscal,dx11);
1411             ty               = _mm_mul_pd(fscal,dy11);
1412             tz               = _mm_mul_pd(fscal,dz11);
1413
1414             /* Update vectorial force */
1415             fix1             = _mm_add_pd(fix1,tx);
1416             fiy1             = _mm_add_pd(fiy1,ty);
1417             fiz1             = _mm_add_pd(fiz1,tz);
1418
1419             fjx1             = _mm_add_pd(fjx1,tx);
1420             fjy1             = _mm_add_pd(fjy1,ty);
1421             fjz1             = _mm_add_pd(fjz1,tz);
1422
1423             /**************************
1424              * CALCULATE INTERACTIONS *
1425              **************************/
1426
1427             /* REACTION-FIELD ELECTROSTATICS */
1428             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1429
1430             fscal            = felec;
1431
1432             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1433
1434             /* Calculate temporary vectorial force */
1435             tx               = _mm_mul_pd(fscal,dx12);
1436             ty               = _mm_mul_pd(fscal,dy12);
1437             tz               = _mm_mul_pd(fscal,dz12);
1438
1439             /* Update vectorial force */
1440             fix1             = _mm_add_pd(fix1,tx);
1441             fiy1             = _mm_add_pd(fiy1,ty);
1442             fiz1             = _mm_add_pd(fiz1,tz);
1443
1444             fjx2             = _mm_add_pd(fjx2,tx);
1445             fjy2             = _mm_add_pd(fjy2,ty);
1446             fjz2             = _mm_add_pd(fjz2,tz);
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             /* REACTION-FIELD ELECTROSTATICS */
1453             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1454
1455             fscal            = felec;
1456
1457             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1458
1459             /* Calculate temporary vectorial force */
1460             tx               = _mm_mul_pd(fscal,dx13);
1461             ty               = _mm_mul_pd(fscal,dy13);
1462             tz               = _mm_mul_pd(fscal,dz13);
1463
1464             /* Update vectorial force */
1465             fix1             = _mm_add_pd(fix1,tx);
1466             fiy1             = _mm_add_pd(fiy1,ty);
1467             fiz1             = _mm_add_pd(fiz1,tz);
1468
1469             fjx3             = _mm_add_pd(fjx3,tx);
1470             fjy3             = _mm_add_pd(fjy3,ty);
1471             fjz3             = _mm_add_pd(fjz3,tz);
1472
1473             /**************************
1474              * CALCULATE INTERACTIONS *
1475              **************************/
1476
1477             /* REACTION-FIELD ELECTROSTATICS */
1478             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1479
1480             fscal            = felec;
1481
1482             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1483
1484             /* Calculate temporary vectorial force */
1485             tx               = _mm_mul_pd(fscal,dx21);
1486             ty               = _mm_mul_pd(fscal,dy21);
1487             tz               = _mm_mul_pd(fscal,dz21);
1488
1489             /* Update vectorial force */
1490             fix2             = _mm_add_pd(fix2,tx);
1491             fiy2             = _mm_add_pd(fiy2,ty);
1492             fiz2             = _mm_add_pd(fiz2,tz);
1493
1494             fjx1             = _mm_add_pd(fjx1,tx);
1495             fjy1             = _mm_add_pd(fjy1,ty);
1496             fjz1             = _mm_add_pd(fjz1,tz);
1497
1498             /**************************
1499              * CALCULATE INTERACTIONS *
1500              **************************/
1501
1502             /* REACTION-FIELD ELECTROSTATICS */
1503             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1504
1505             fscal            = felec;
1506
1507             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1508
1509             /* Calculate temporary vectorial force */
1510             tx               = _mm_mul_pd(fscal,dx22);
1511             ty               = _mm_mul_pd(fscal,dy22);
1512             tz               = _mm_mul_pd(fscal,dz22);
1513
1514             /* Update vectorial force */
1515             fix2             = _mm_add_pd(fix2,tx);
1516             fiy2             = _mm_add_pd(fiy2,ty);
1517             fiz2             = _mm_add_pd(fiz2,tz);
1518
1519             fjx2             = _mm_add_pd(fjx2,tx);
1520             fjy2             = _mm_add_pd(fjy2,ty);
1521             fjz2             = _mm_add_pd(fjz2,tz);
1522
1523             /**************************
1524              * CALCULATE INTERACTIONS *
1525              **************************/
1526
1527             /* REACTION-FIELD ELECTROSTATICS */
1528             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1529
1530             fscal            = felec;
1531
1532             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1533
1534             /* Calculate temporary vectorial force */
1535             tx               = _mm_mul_pd(fscal,dx23);
1536             ty               = _mm_mul_pd(fscal,dy23);
1537             tz               = _mm_mul_pd(fscal,dz23);
1538
1539             /* Update vectorial force */
1540             fix2             = _mm_add_pd(fix2,tx);
1541             fiy2             = _mm_add_pd(fiy2,ty);
1542             fiz2             = _mm_add_pd(fiz2,tz);
1543
1544             fjx3             = _mm_add_pd(fjx3,tx);
1545             fjy3             = _mm_add_pd(fjy3,ty);
1546             fjz3             = _mm_add_pd(fjz3,tz);
1547
1548             /**************************
1549              * CALCULATE INTERACTIONS *
1550              **************************/
1551
1552             /* REACTION-FIELD ELECTROSTATICS */
1553             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1554
1555             fscal            = felec;
1556
1557             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1558
1559             /* Calculate temporary vectorial force */
1560             tx               = _mm_mul_pd(fscal,dx31);
1561             ty               = _mm_mul_pd(fscal,dy31);
1562             tz               = _mm_mul_pd(fscal,dz31);
1563
1564             /* Update vectorial force */
1565             fix3             = _mm_add_pd(fix3,tx);
1566             fiy3             = _mm_add_pd(fiy3,ty);
1567             fiz3             = _mm_add_pd(fiz3,tz);
1568
1569             fjx1             = _mm_add_pd(fjx1,tx);
1570             fjy1             = _mm_add_pd(fjy1,ty);
1571             fjz1             = _mm_add_pd(fjz1,tz);
1572
1573             /**************************
1574              * CALCULATE INTERACTIONS *
1575              **************************/
1576
1577             /* REACTION-FIELD ELECTROSTATICS */
1578             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1579
1580             fscal            = felec;
1581
1582             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1583
1584             /* Calculate temporary vectorial force */
1585             tx               = _mm_mul_pd(fscal,dx32);
1586             ty               = _mm_mul_pd(fscal,dy32);
1587             tz               = _mm_mul_pd(fscal,dz32);
1588
1589             /* Update vectorial force */
1590             fix3             = _mm_add_pd(fix3,tx);
1591             fiy3             = _mm_add_pd(fiy3,ty);
1592             fiz3             = _mm_add_pd(fiz3,tz);
1593
1594             fjx2             = _mm_add_pd(fjx2,tx);
1595             fjy2             = _mm_add_pd(fjy2,ty);
1596             fjz2             = _mm_add_pd(fjz2,tz);
1597
1598             /**************************
1599              * CALCULATE INTERACTIONS *
1600              **************************/
1601
1602             /* REACTION-FIELD ELECTROSTATICS */
1603             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1604
1605             fscal            = felec;
1606
1607             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1608
1609             /* Calculate temporary vectorial force */
1610             tx               = _mm_mul_pd(fscal,dx33);
1611             ty               = _mm_mul_pd(fscal,dy33);
1612             tz               = _mm_mul_pd(fscal,dz33);
1613
1614             /* Update vectorial force */
1615             fix3             = _mm_add_pd(fix3,tx);
1616             fiy3             = _mm_add_pd(fiy3,ty);
1617             fiz3             = _mm_add_pd(fiz3,tz);
1618
1619             fjx3             = _mm_add_pd(fjx3,tx);
1620             fjy3             = _mm_add_pd(fjy3,ty);
1621             fjz3             = _mm_add_pd(fjz3,tz);
1622
1623             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1624
1625             /* Inner loop uses 243 flops */
1626         }
1627
1628         /* End of innermost loop */
1629
1630         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1631                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1632
1633         /* Increment number of inner iterations */
1634         inneriter                  += j_index_end - j_index_start;
1635
1636         /* Outer loop uses 18 flops */
1637     }
1638
1639     /* Increment number of outer iterations */
1640     outeriter        += nri;
1641
1642     /* Update outer/inner flops */
1643
1644     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1645 }