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) 2012,2013,2014, 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 sse4_1_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse4_1_double.h"
  48 #include "kernelutil_x86_sse4_1_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sse4_1_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sse4_1_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset1;
  81     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  82     int              vdwioffset2;
  83     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  84     int              vdwioffset3;
  85     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  86     int              vdwjidx1A,vdwjidx1B;
  87     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  88     int              vdwjidx2A,vdwjidx2B;
  89     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  90     int              vdwjidx3A,vdwjidx3B;
  91     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  92     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  93     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  94     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  95     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  96     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  97     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  98     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  99     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 100     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     __m128d          dummy_mask,cutoff_mask;
 104     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 105     __m128d          one     = _mm_set1_pd(1.0);
 106     __m128d          two     = _mm_set1_pd(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_pd(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     krf              = _mm_set1_pd(fr->ic->k_rf);
 121     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 122     crf              = _mm_set1_pd(fr->ic->c_rf);
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 129
 130     jq1              = _mm_set1_pd(charge[inr+1]);
 131     jq2              = _mm_set1_pd(charge[inr+2]);
 132     jq3              = _mm_set1_pd(charge[inr+3]);
 133     qq11             = _mm_mul_pd(iq1,jq1);
 134     qq12             = _mm_mul_pd(iq1,jq2);
 135     qq13             = _mm_mul_pd(iq1,jq3);
 136     qq21             = _mm_mul_pd(iq2,jq1);
 137     qq22             = _mm_mul_pd(iq2,jq2);
 138     qq23             = _mm_mul_pd(iq2,jq3);
 139     qq31             = _mm_mul_pd(iq3,jq1);
 140     qq32             = _mm_mul_pd(iq3,jq2);
 141     qq33             = _mm_mul_pd(iq3,jq3);
 142
 143     /* Avoid stupid compiler warnings */
 144     jnrA = jnrB = 0;
 145     j_coord_offsetA = 0;
 146     j_coord_offsetB = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 167                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 168
 169         fix1             = _mm_setzero_pd();
 170         fiy1             = _mm_setzero_pd();
 171         fiz1             = _mm_setzero_pd();
 172         fix2             = _mm_setzero_pd();
 173         fiy2             = _mm_setzero_pd();
 174         fiz2             = _mm_setzero_pd();
 175         fix3             = _mm_setzero_pd();
 176         fiy3             = _mm_setzero_pd();
 177         fiz3             = _mm_setzero_pd();
 178
 179         /* Reset potential sums */
 180         velecsum         = _mm_setzero_pd();
 181
 182         /* Start inner kernel loop */
 183         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 184         {
 185
 186             /* Get j neighbor index, and coordinate index */
 187             jnrA             = jjnr[jidx];
 188             jnrB             = jjnr[jidx+1];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 194                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 195
 196             /* Calculate displacement vector */
 197             dx11             = _mm_sub_pd(ix1,jx1);
 198             dy11             = _mm_sub_pd(iy1,jy1);
 199             dz11             = _mm_sub_pd(iz1,jz1);
 200             dx12             = _mm_sub_pd(ix1,jx2);
 201             dy12             = _mm_sub_pd(iy1,jy2);
 202             dz12             = _mm_sub_pd(iz1,jz2);
 203             dx13             = _mm_sub_pd(ix1,jx3);
 204             dy13             = _mm_sub_pd(iy1,jy3);
 205             dz13             = _mm_sub_pd(iz1,jz3);
 206             dx21             = _mm_sub_pd(ix2,jx1);
 207             dy21             = _mm_sub_pd(iy2,jy1);
 208             dz21             = _mm_sub_pd(iz2,jz1);
 209             dx22             = _mm_sub_pd(ix2,jx2);
 210             dy22             = _mm_sub_pd(iy2,jy2);
 211             dz22             = _mm_sub_pd(iz2,jz2);
 212             dx23             = _mm_sub_pd(ix2,jx3);
 213             dy23             = _mm_sub_pd(iy2,jy3);
 214             dz23             = _mm_sub_pd(iz2,jz3);
 215             dx31             = _mm_sub_pd(ix3,jx1);
 216             dy31             = _mm_sub_pd(iy3,jy1);
 217             dz31             = _mm_sub_pd(iz3,jz1);
 218             dx32             = _mm_sub_pd(ix3,jx2);
 219             dy32             = _mm_sub_pd(iy3,jy2);
 220             dz32             = _mm_sub_pd(iz3,jz2);
 221             dx33             = _mm_sub_pd(ix3,jx3);
 222             dy33             = _mm_sub_pd(iy3,jy3);
 223             dz33             = _mm_sub_pd(iz3,jz3);
 224
 225             /* Calculate squared distance and things based on it */
 226             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 227             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 228             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 229             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 230             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 231             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 232             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 233             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 234             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 235
 236             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 237             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 238             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 239             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 240             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 241             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 242             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 243             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 244             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 245
 246             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 247             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 248             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 249             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 250             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 251             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 252             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 253             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 254             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 255
 256             fjx1             = _mm_setzero_pd();
 257             fjy1             = _mm_setzero_pd();
 258             fjz1             = _mm_setzero_pd();
 259             fjx2             = _mm_setzero_pd();
 260             fjy2             = _mm_setzero_pd();
 261             fjz2             = _mm_setzero_pd();
 262             fjx3             = _mm_setzero_pd();
 263             fjy3             = _mm_setzero_pd();
 264             fjz3             = _mm_setzero_pd();
 265
 266             /**************************
 267              * CALCULATE INTERACTIONS *
 268              **************************/
 269
 270             /* REACTION-FIELD ELECTROSTATICS */
 271             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 272             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 273
 274             /* Update potential sum for this i atom from the interaction with this j atom. */
 275             velecsum         = _mm_add_pd(velecsum,velec);
 276
 277             fscal            = felec;
 278
 279             /* Calculate temporary vectorial force */
 280             tx               = _mm_mul_pd(fscal,dx11);
 281             ty               = _mm_mul_pd(fscal,dy11);
 282             tz               = _mm_mul_pd(fscal,dz11);
 283
 284             /* Update vectorial force */
 285             fix1             = _mm_add_pd(fix1,tx);
 286             fiy1             = _mm_add_pd(fiy1,ty);
 287             fiz1             = _mm_add_pd(fiz1,tz);
 288
 289             fjx1             = _mm_add_pd(fjx1,tx);
 290             fjy1             = _mm_add_pd(fjy1,ty);
 291             fjz1             = _mm_add_pd(fjz1,tz);
 292
 293             /**************************
 294              * CALCULATE INTERACTIONS *
 295              **************************/
 296
 297             /* REACTION-FIELD ELECTROSTATICS */
 298             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 299             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 300
 301             /* Update potential sum for this i atom from the interaction with this j atom. */
 302             velecsum         = _mm_add_pd(velecsum,velec);
 303
 304             fscal            = felec;
 305
 306             /* Calculate temporary vectorial force */
 307             tx               = _mm_mul_pd(fscal,dx12);
 308             ty               = _mm_mul_pd(fscal,dy12);
 309             tz               = _mm_mul_pd(fscal,dz12);
 310
 311             /* Update vectorial force */
 312             fix1             = _mm_add_pd(fix1,tx);
 313             fiy1             = _mm_add_pd(fiy1,ty);
 314             fiz1             = _mm_add_pd(fiz1,tz);
 315
 316             fjx2             = _mm_add_pd(fjx2,tx);
 317             fjy2             = _mm_add_pd(fjy2,ty);
 318             fjz2             = _mm_add_pd(fjz2,tz);
 319
 320             /**************************
 321              * CALCULATE INTERACTIONS *
 322              **************************/
 323
 324             /* REACTION-FIELD ELECTROSTATICS */
 325             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 326             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 327
 328             /* Update potential sum for this i atom from the interaction with this j atom. */
 329             velecsum         = _mm_add_pd(velecsum,velec);
 330
 331             fscal            = felec;
 332
 333             /* Calculate temporary vectorial force */
 334             tx               = _mm_mul_pd(fscal,dx13);
 335             ty               = _mm_mul_pd(fscal,dy13);
 336             tz               = _mm_mul_pd(fscal,dz13);
 337
 338             /* Update vectorial force */
 339             fix1             = _mm_add_pd(fix1,tx);
 340             fiy1             = _mm_add_pd(fiy1,ty);
 341             fiz1             = _mm_add_pd(fiz1,tz);
 342
 343             fjx3             = _mm_add_pd(fjx3,tx);
 344             fjy3             = _mm_add_pd(fjy3,ty);
 345             fjz3             = _mm_add_pd(fjz3,tz);
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             /* REACTION-FIELD ELECTROSTATICS */
 352             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 353             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velecsum         = _mm_add_pd(velecsum,velec);
 357
 358             fscal            = felec;
 359
 360             /* Calculate temporary vectorial force */
 361             tx               = _mm_mul_pd(fscal,dx21);
 362             ty               = _mm_mul_pd(fscal,dy21);
 363             tz               = _mm_mul_pd(fscal,dz21);
 364
 365             /* Update vectorial force */
 366             fix2             = _mm_add_pd(fix2,tx);
 367             fiy2             = _mm_add_pd(fiy2,ty);
 368             fiz2             = _mm_add_pd(fiz2,tz);
 369
 370             fjx1             = _mm_add_pd(fjx1,tx);
 371             fjy1             = _mm_add_pd(fjy1,ty);
 372             fjz1             = _mm_add_pd(fjz1,tz);
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             /* REACTION-FIELD ELECTROSTATICS */
 379             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 380             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velecsum         = _mm_add_pd(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             /* Calculate temporary vectorial force */
 388             tx               = _mm_mul_pd(fscal,dx22);
 389             ty               = _mm_mul_pd(fscal,dy22);
 390             tz               = _mm_mul_pd(fscal,dz22);
 391
 392             /* Update vectorial force */
 393             fix2             = _mm_add_pd(fix2,tx);
 394             fiy2             = _mm_add_pd(fiy2,ty);
 395             fiz2             = _mm_add_pd(fiz2,tz);
 396
 397             fjx2             = _mm_add_pd(fjx2,tx);
 398             fjy2             = _mm_add_pd(fjy2,ty);
 399             fjz2             = _mm_add_pd(fjz2,tz);
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 407             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 408
 409             /* Update potential sum for this i atom from the interaction with this j atom. */
 410             velecsum         = _mm_add_pd(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414             /* Calculate temporary vectorial force */
 415             tx               = _mm_mul_pd(fscal,dx23);
 416             ty               = _mm_mul_pd(fscal,dy23);
 417             tz               = _mm_mul_pd(fscal,dz23);
 418
 419             /* Update vectorial force */
 420             fix2             = _mm_add_pd(fix2,tx);
 421             fiy2             = _mm_add_pd(fiy2,ty);
 422             fiz2             = _mm_add_pd(fiz2,tz);
 423
 424             fjx3             = _mm_add_pd(fjx3,tx);
 425             fjy3             = _mm_add_pd(fjy3,ty);
 426             fjz3             = _mm_add_pd(fjz3,tz);
 427
 428             /**************************
 429              * CALCULATE INTERACTIONS *
 430              **************************/
 431
 432             /* REACTION-FIELD ELECTROSTATICS */
 433             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 434             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 435
 436             /* Update potential sum for this i atom from the interaction with this j atom. */
 437             velecsum         = _mm_add_pd(velecsum,velec);
 438
 439             fscal            = felec;
 440
 441             /* Calculate temporary vectorial force */
 442             tx               = _mm_mul_pd(fscal,dx31);
 443             ty               = _mm_mul_pd(fscal,dy31);
 444             tz               = _mm_mul_pd(fscal,dz31);
 445
 446             /* Update vectorial force */
 447             fix3             = _mm_add_pd(fix3,tx);
 448             fiy3             = _mm_add_pd(fiy3,ty);
 449             fiz3             = _mm_add_pd(fiz3,tz);
 450
 451             fjx1             = _mm_add_pd(fjx1,tx);
 452             fjy1             = _mm_add_pd(fjy1,ty);
 453             fjz1             = _mm_add_pd(fjz1,tz);
 454
 455             /**************************
 456              * CALCULATE INTERACTIONS *
 457              **************************/
 458
 459             /* REACTION-FIELD ELECTROSTATICS */
 460             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 461             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 462
 463             /* Update potential sum for this i atom from the interaction with this j atom. */
 464             velecsum         = _mm_add_pd(velecsum,velec);
 465
 466             fscal            = felec;
 467
 468             /* Calculate temporary vectorial force */
 469             tx               = _mm_mul_pd(fscal,dx32);
 470             ty               = _mm_mul_pd(fscal,dy32);
 471             tz               = _mm_mul_pd(fscal,dz32);
 472
 473             /* Update vectorial force */
 474             fix3             = _mm_add_pd(fix3,tx);
 475             fiy3             = _mm_add_pd(fiy3,ty);
 476             fiz3             = _mm_add_pd(fiz3,tz);
 477
 478             fjx2             = _mm_add_pd(fjx2,tx);
 479             fjy2             = _mm_add_pd(fjy2,ty);
 480             fjz2             = _mm_add_pd(fjz2,tz);
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             /* REACTION-FIELD ELECTROSTATICS */
 487             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 488             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 489
 490             /* Update potential sum for this i atom from the interaction with this j atom. */
 491             velecsum         = _mm_add_pd(velecsum,velec);
 492
 493             fscal            = felec;
 494
 495             /* Calculate temporary vectorial force */
 496             tx               = _mm_mul_pd(fscal,dx33);
 497             ty               = _mm_mul_pd(fscal,dy33);
 498             tz               = _mm_mul_pd(fscal,dz33);
 499
 500             /* Update vectorial force */
 501             fix3             = _mm_add_pd(fix3,tx);
 502             fiy3             = _mm_add_pd(fiy3,ty);
 503             fiz3             = _mm_add_pd(fiz3,tz);
 504
 505             fjx3             = _mm_add_pd(fjx3,tx);
 506             fjy3             = _mm_add_pd(fjy3,ty);
 507             fjz3             = _mm_add_pd(fjz3,tz);
 508
 509             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);
 510
 511             /* Inner loop uses 288 flops */
 512         }
 513
 514         if(jidx<j_index_end)
 515         {
 516
 517             jnrA             = jjnr[jidx];
 518             j_coord_offsetA  = DIM*jnrA;
 519
 520             /* load j atom coordinates */
 521             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 522                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 523
 524             /* Calculate displacement vector */
 525             dx11             = _mm_sub_pd(ix1,jx1);
 526             dy11             = _mm_sub_pd(iy1,jy1);
 527             dz11             = _mm_sub_pd(iz1,jz1);
 528             dx12             = _mm_sub_pd(ix1,jx2);
 529             dy12             = _mm_sub_pd(iy1,jy2);
 530             dz12             = _mm_sub_pd(iz1,jz2);
 531             dx13             = _mm_sub_pd(ix1,jx3);
 532             dy13             = _mm_sub_pd(iy1,jy3);
 533             dz13             = _mm_sub_pd(iz1,jz3);
 534             dx21             = _mm_sub_pd(ix2,jx1);
 535             dy21             = _mm_sub_pd(iy2,jy1);
 536             dz21             = _mm_sub_pd(iz2,jz1);
 537             dx22             = _mm_sub_pd(ix2,jx2);
 538             dy22             = _mm_sub_pd(iy2,jy2);
 539             dz22             = _mm_sub_pd(iz2,jz2);
 540             dx23             = _mm_sub_pd(ix2,jx3);
 541             dy23             = _mm_sub_pd(iy2,jy3);
 542             dz23             = _mm_sub_pd(iz2,jz3);
 543             dx31             = _mm_sub_pd(ix3,jx1);
 544             dy31             = _mm_sub_pd(iy3,jy1);
 545             dz31             = _mm_sub_pd(iz3,jz1);
 546             dx32             = _mm_sub_pd(ix3,jx2);
 547             dy32             = _mm_sub_pd(iy3,jy2);
 548             dz32             = _mm_sub_pd(iz3,jz2);
 549             dx33             = _mm_sub_pd(ix3,jx3);
 550             dy33             = _mm_sub_pd(iy3,jy3);
 551             dz33             = _mm_sub_pd(iz3,jz3);
 552
 553             /* Calculate squared distance and things based on it */
 554             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 555             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 556             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 557             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 558             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 559             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 560             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 561             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 562             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 563
 564             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 565             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 566             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 567             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 568             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 569             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 570             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 571             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 572             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 573
 574             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 575             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 576             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 577             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 578             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 579             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 580             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 581             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 582             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 583
 584             fjx1             = _mm_setzero_pd();
 585             fjy1             = _mm_setzero_pd();
 586             fjz1             = _mm_setzero_pd();
 587             fjx2             = _mm_setzero_pd();
 588             fjy2             = _mm_setzero_pd();
 589             fjz2             = _mm_setzero_pd();
 590             fjx3             = _mm_setzero_pd();
 591             fjy3             = _mm_setzero_pd();
 592             fjz3             = _mm_setzero_pd();
 593
 594             /**************************
 595              * CALCULATE INTERACTIONS *
 596              **************************/
 597
 598             /* REACTION-FIELD ELECTROSTATICS */
 599             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 600             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 601
 602             /* Update potential sum for this i atom from the interaction with this j atom. */
 603             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 604             velecsum         = _mm_add_pd(velecsum,velec);
 605
 606             fscal            = felec;
 607
 608             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 609
 610             /* Calculate temporary vectorial force */
 611             tx               = _mm_mul_pd(fscal,dx11);
 612             ty               = _mm_mul_pd(fscal,dy11);
 613             tz               = _mm_mul_pd(fscal,dz11);
 614
 615             /* Update vectorial force */
 616             fix1             = _mm_add_pd(fix1,tx);
 617             fiy1             = _mm_add_pd(fiy1,ty);
 618             fiz1             = _mm_add_pd(fiz1,tz);
 619
 620             fjx1             = _mm_add_pd(fjx1,tx);
 621             fjy1             = _mm_add_pd(fjy1,ty);
 622             fjz1             = _mm_add_pd(fjz1,tz);
 623
 624             /**************************
 625              * CALCULATE INTERACTIONS *
 626              **************************/
 627
 628             /* REACTION-FIELD ELECTROSTATICS */
 629             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 630             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 631
 632             /* Update potential sum for this i atom from the interaction with this j atom. */
 633             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 634             velecsum         = _mm_add_pd(velecsum,velec);
 635
 636             fscal            = felec;
 637
 638             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 639
 640             /* Calculate temporary vectorial force */
 641             tx               = _mm_mul_pd(fscal,dx12);
 642             ty               = _mm_mul_pd(fscal,dy12);
 643             tz               = _mm_mul_pd(fscal,dz12);
 644
 645             /* Update vectorial force */
 646             fix1             = _mm_add_pd(fix1,tx);
 647             fiy1             = _mm_add_pd(fiy1,ty);
 648             fiz1             = _mm_add_pd(fiz1,tz);
 649
 650             fjx2             = _mm_add_pd(fjx2,tx);
 651             fjy2             = _mm_add_pd(fjy2,ty);
 652             fjz2             = _mm_add_pd(fjz2,tz);
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             /* REACTION-FIELD ELECTROSTATICS */
 659             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 660             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 661
 662             /* Update potential sum for this i atom from the interaction with this j atom. */
 663             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 664             velecsum         = _mm_add_pd(velecsum,velec);
 665
 666             fscal            = felec;
 667
 668             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 669
 670             /* Calculate temporary vectorial force */
 671             tx               = _mm_mul_pd(fscal,dx13);
 672             ty               = _mm_mul_pd(fscal,dy13);
 673             tz               = _mm_mul_pd(fscal,dz13);
 674
 675             /* Update vectorial force */
 676             fix1             = _mm_add_pd(fix1,tx);
 677             fiy1             = _mm_add_pd(fiy1,ty);
 678             fiz1             = _mm_add_pd(fiz1,tz);
 679
 680             fjx3             = _mm_add_pd(fjx3,tx);
 681             fjy3             = _mm_add_pd(fjy3,ty);
 682             fjz3             = _mm_add_pd(fjz3,tz);
 683
 684             /**************************
 685              * CALCULATE INTERACTIONS *
 686              **************************/
 687
 688             /* REACTION-FIELD ELECTROSTATICS */
 689             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 690             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 691
 692             /* Update potential sum for this i atom from the interaction with this j atom. */
 693             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 694             velecsum         = _mm_add_pd(velecsum,velec);
 695
 696             fscal            = felec;
 697
 698             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 699
 700             /* Calculate temporary vectorial force */
 701             tx               = _mm_mul_pd(fscal,dx21);
 702             ty               = _mm_mul_pd(fscal,dy21);
 703             tz               = _mm_mul_pd(fscal,dz21);
 704
 705             /* Update vectorial force */
 706             fix2             = _mm_add_pd(fix2,tx);
 707             fiy2             = _mm_add_pd(fiy2,ty);
 708             fiz2             = _mm_add_pd(fiz2,tz);
 709
 710             fjx1             = _mm_add_pd(fjx1,tx);
 711             fjy1             = _mm_add_pd(fjy1,ty);
 712             fjz1             = _mm_add_pd(fjz1,tz);
 713
 714             /**************************
 715              * CALCULATE INTERACTIONS *
 716              **************************/
 717
 718             /* REACTION-FIELD ELECTROSTATICS */
 719             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 720             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 721
 722             /* Update potential sum for this i atom from the interaction with this j atom. */
 723             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 724             velecsum         = _mm_add_pd(velecsum,velec);
 725
 726             fscal            = felec;
 727
 728             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 729
 730             /* Calculate temporary vectorial force */
 731             tx               = _mm_mul_pd(fscal,dx22);
 732             ty               = _mm_mul_pd(fscal,dy22);
 733             tz               = _mm_mul_pd(fscal,dz22);
 734
 735             /* Update vectorial force */
 736             fix2             = _mm_add_pd(fix2,tx);
 737             fiy2             = _mm_add_pd(fiy2,ty);
 738             fiz2             = _mm_add_pd(fiz2,tz);
 739
 740             fjx2             = _mm_add_pd(fjx2,tx);
 741             fjy2             = _mm_add_pd(fjy2,ty);
 742             fjz2             = _mm_add_pd(fjz2,tz);
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             /* REACTION-FIELD ELECTROSTATICS */
 749             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 750             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 751
 752             /* Update potential sum for this i atom from the interaction with this j atom. */
 753             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 754             velecsum         = _mm_add_pd(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 759
 760             /* Calculate temporary vectorial force */
 761             tx               = _mm_mul_pd(fscal,dx23);
 762             ty               = _mm_mul_pd(fscal,dy23);
 763             tz               = _mm_mul_pd(fscal,dz23);
 764
 765             /* Update vectorial force */
 766             fix2             = _mm_add_pd(fix2,tx);
 767             fiy2             = _mm_add_pd(fiy2,ty);
 768             fiz2             = _mm_add_pd(fiz2,tz);
 769
 770             fjx3             = _mm_add_pd(fjx3,tx);
 771             fjy3             = _mm_add_pd(fjy3,ty);
 772             fjz3             = _mm_add_pd(fjz3,tz);
 773
 774             /**************************
 775              * CALCULATE INTERACTIONS *
 776              **************************/
 777
 778             /* REACTION-FIELD ELECTROSTATICS */
 779             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 780             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 781
 782             /* Update potential sum for this i atom from the interaction with this j atom. */
 783             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 784             velecsum         = _mm_add_pd(velecsum,velec);
 785
 786             fscal            = felec;
 787
 788             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 789
 790             /* Calculate temporary vectorial force */
 791             tx               = _mm_mul_pd(fscal,dx31);
 792             ty               = _mm_mul_pd(fscal,dy31);
 793             tz               = _mm_mul_pd(fscal,dz31);
 794
 795             /* Update vectorial force */
 796             fix3             = _mm_add_pd(fix3,tx);
 797             fiy3             = _mm_add_pd(fiy3,ty);
 798             fiz3             = _mm_add_pd(fiz3,tz);
 799
 800             fjx1             = _mm_add_pd(fjx1,tx);
 801             fjy1             = _mm_add_pd(fjy1,ty);
 802             fjz1             = _mm_add_pd(fjz1,tz);
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             /* REACTION-FIELD ELECTROSTATICS */
 809             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 810             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 811
 812             /* Update potential sum for this i atom from the interaction with this j atom. */
 813             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 814             velecsum         = _mm_add_pd(velecsum,velec);
 815
 816             fscal            = felec;
 817
 818             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 819
 820             /* Calculate temporary vectorial force */
 821             tx               = _mm_mul_pd(fscal,dx32);
 822             ty               = _mm_mul_pd(fscal,dy32);
 823             tz               = _mm_mul_pd(fscal,dz32);
 824
 825             /* Update vectorial force */
 826             fix3             = _mm_add_pd(fix3,tx);
 827             fiy3             = _mm_add_pd(fiy3,ty);
 828             fiz3             = _mm_add_pd(fiz3,tz);
 829
 830             fjx2             = _mm_add_pd(fjx2,tx);
 831             fjy2             = _mm_add_pd(fjy2,ty);
 832             fjz2             = _mm_add_pd(fjz2,tz);
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             /* REACTION-FIELD ELECTROSTATICS */
 839             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 840             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 841
 842             /* Update potential sum for this i atom from the interaction with this j atom. */
 843             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 844             velecsum         = _mm_add_pd(velecsum,velec);
 845
 846             fscal            = felec;
 847
 848             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 849
 850             /* Calculate temporary vectorial force */
 851             tx               = _mm_mul_pd(fscal,dx33);
 852             ty               = _mm_mul_pd(fscal,dy33);
 853             tz               = _mm_mul_pd(fscal,dz33);
 854
 855             /* Update vectorial force */
 856             fix3             = _mm_add_pd(fix3,tx);
 857             fiy3             = _mm_add_pd(fiy3,ty);
 858             fiz3             = _mm_add_pd(fiz3,tz);
 859
 860             fjx3             = _mm_add_pd(fjx3,tx);
 861             fjy3             = _mm_add_pd(fjy3,ty);
 862             fjz3             = _mm_add_pd(fjz3,tz);
 863
 864             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 865
 866             /* Inner loop uses 288 flops */
 867         }
 868
 869         /* End of innermost loop */
 870
 871         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 872                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 873
 874         ggid                        = gid[iidx];
 875         /* Update potential energies */
 876         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 877
 878         /* Increment number of inner iterations */
 879         inneriter                  += j_index_end - j_index_start;
 880
 881         /* Outer loop uses 19 flops */
 882     }
 883
 884     /* Increment number of outer iterations */
 885     outeriter        += nri;
 886
 887     /* Update outer/inner flops */
 888
 889     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*288);
 890 }
 891 /*
 892  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse4_1_double
 893  * Electrostatics interaction: ReactionField
 894  * VdW interaction:            None
 895  * Geometry:                   Water4-Water4
 896  * Calculate force/pot:        Force
 897  */
 898 void
 899 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse4_1_double
 900                     (t_nblist                    * gmx_restrict       nlist,
 901                      rvec                        * gmx_restrict          xx,
 902                      rvec                        * gmx_restrict          ff,
 903                      t_forcerec                  * gmx_restrict          fr,
 904                      t_mdatoms                   * gmx_restrict     mdatoms,
 905                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 906                      t_nrnb                      * gmx_restrict        nrnb)
 907 {
 908     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 909      * just 0 for non-waters.
 910      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 911      * jnr indices corresponding to data put in the four positions in the SIMD register.
 912      */
 913     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 914     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 915     int              jnrA,jnrB;
 916     int              j_coord_offsetA,j_coord_offsetB;
 917     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 918     real             rcutoff_scalar;
 919     real             *shiftvec,*fshift,*x,*f;
 920     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 921     int              vdwioffset1;
 922     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 923     int              vdwioffset2;
 924     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 925     int              vdwioffset3;
 926     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 927     int              vdwjidx1A,vdwjidx1B;
 928     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 929     int              vdwjidx2A,vdwjidx2B;
 930     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 931     int              vdwjidx3A,vdwjidx3B;
 932     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 933     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 934     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 935     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 936     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 937     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 938     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 939     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 940     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 941     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 942     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 943     real             *charge;
 944     __m128d          dummy_mask,cutoff_mask;
 945     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 946     __m128d          one     = _mm_set1_pd(1.0);
 947     __m128d          two     = _mm_set1_pd(2.0);
 948     x                = xx[0];
 949     f                = ff[0];
 950
 951     nri              = nlist->nri;
 952     iinr             = nlist->iinr;
 953     jindex           = nlist->jindex;
 954     jjnr             = nlist->jjnr;
 955     shiftidx         = nlist->shift;
 956     gid              = nlist->gid;
 957     shiftvec         = fr->shift_vec[0];
 958     fshift           = fr->fshift[0];
 959     facel            = _mm_set1_pd(fr->epsfac);
 960     charge           = mdatoms->chargeA;
 961     krf              = _mm_set1_pd(fr->ic->k_rf);
 962     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 963     crf              = _mm_set1_pd(fr->ic->c_rf);
 964
 965     /* Setup water-specific parameters */
 966     inr              = nlist->iinr[0];
 967     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 968     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 969     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 970
 971     jq1              = _mm_set1_pd(charge[inr+1]);
 972     jq2              = _mm_set1_pd(charge[inr+2]);
 973     jq3              = _mm_set1_pd(charge[inr+3]);
 974     qq11             = _mm_mul_pd(iq1,jq1);
 975     qq12             = _mm_mul_pd(iq1,jq2);
 976     qq13             = _mm_mul_pd(iq1,jq3);
 977     qq21             = _mm_mul_pd(iq2,jq1);
 978     qq22             = _mm_mul_pd(iq2,jq2);
 979     qq23             = _mm_mul_pd(iq2,jq3);
 980     qq31             = _mm_mul_pd(iq3,jq1);
 981     qq32             = _mm_mul_pd(iq3,jq2);
 982     qq33             = _mm_mul_pd(iq3,jq3);
 983
 984     /* Avoid stupid compiler warnings */
 985     jnrA = jnrB = 0;
 986     j_coord_offsetA = 0;
 987     j_coord_offsetB = 0;
 988
 989     outeriter        = 0;
 990     inneriter        = 0;
 991
 992     /* Start outer loop over neighborlists */
 993     for(iidx=0; iidx<nri; iidx++)
 994     {
 995         /* Load shift vector for this list */
 996         i_shift_offset   = DIM*shiftidx[iidx];
 997
 998         /* Load limits for loop over neighbors */
 999         j_index_start    = jindex[iidx];
1000         j_index_end      = jindex[iidx+1];
1001
1002         /* Get outer coordinate index */
1003         inr              = iinr[iidx];
1004         i_coord_offset   = DIM*inr;
1005
1006         /* Load i particle coords and add shift vector */
1007         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1008                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1009
1010         fix1             = _mm_setzero_pd();
1011         fiy1             = _mm_setzero_pd();
1012         fiz1             = _mm_setzero_pd();
1013         fix2             = _mm_setzero_pd();
1014         fiy2             = _mm_setzero_pd();
1015         fiz2             = _mm_setzero_pd();
1016         fix3             = _mm_setzero_pd();
1017         fiy3             = _mm_setzero_pd();
1018         fiz3             = _mm_setzero_pd();
1019
1020         /* Start inner kernel loop */
1021         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1022         {
1023
1024             /* Get j neighbor index, and coordinate index */
1025             jnrA             = jjnr[jidx];
1026             jnrB             = jjnr[jidx+1];
1027             j_coord_offsetA  = DIM*jnrA;
1028             j_coord_offsetB  = DIM*jnrB;
1029
1030             /* load j atom coordinates */
1031             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1032                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1033
1034             /* Calculate displacement vector */
1035             dx11             = _mm_sub_pd(ix1,jx1);
1036             dy11             = _mm_sub_pd(iy1,jy1);
1037             dz11             = _mm_sub_pd(iz1,jz1);
1038             dx12             = _mm_sub_pd(ix1,jx2);
1039             dy12             = _mm_sub_pd(iy1,jy2);
1040             dz12             = _mm_sub_pd(iz1,jz2);
1041             dx13             = _mm_sub_pd(ix1,jx3);
1042             dy13             = _mm_sub_pd(iy1,jy3);
1043             dz13             = _mm_sub_pd(iz1,jz3);
1044             dx21             = _mm_sub_pd(ix2,jx1);
1045             dy21             = _mm_sub_pd(iy2,jy1);
1046             dz21             = _mm_sub_pd(iz2,jz1);
1047             dx22             = _mm_sub_pd(ix2,jx2);
1048             dy22             = _mm_sub_pd(iy2,jy2);
1049             dz22             = _mm_sub_pd(iz2,jz2);
1050             dx23             = _mm_sub_pd(ix2,jx3);
1051             dy23             = _mm_sub_pd(iy2,jy3);
1052             dz23             = _mm_sub_pd(iz2,jz3);
1053             dx31             = _mm_sub_pd(ix3,jx1);
1054             dy31             = _mm_sub_pd(iy3,jy1);
1055             dz31             = _mm_sub_pd(iz3,jz1);
1056             dx32             = _mm_sub_pd(ix3,jx2);
1057             dy32             = _mm_sub_pd(iy3,jy2);
1058             dz32             = _mm_sub_pd(iz3,jz2);
1059             dx33             = _mm_sub_pd(ix3,jx3);
1060             dy33             = _mm_sub_pd(iy3,jy3);
1061             dz33             = _mm_sub_pd(iz3,jz3);
1062
1063             /* Calculate squared distance and things based on it */
1064             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1065             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1066             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1067             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1068             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1069             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1070             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1071             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1072             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1073
1074             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1075             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1076             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1077             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1078             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1079             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1080             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1081             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1082             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1083
1084             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1085             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1086             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1087             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1088             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1089             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1090             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1091             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1092             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1093
1094             fjx1             = _mm_setzero_pd();
1095             fjy1             = _mm_setzero_pd();
1096             fjz1             = _mm_setzero_pd();
1097             fjx2             = _mm_setzero_pd();
1098             fjy2             = _mm_setzero_pd();
1099             fjz2             = _mm_setzero_pd();
1100             fjx3             = _mm_setzero_pd();
1101             fjy3             = _mm_setzero_pd();
1102             fjz3             = _mm_setzero_pd();
1103
1104             /**************************
1105              * CALCULATE INTERACTIONS *
1106              **************************/
1107
1108             /* REACTION-FIELD ELECTROSTATICS */
1109             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1110
1111             fscal            = felec;
1112
1113             /* Calculate temporary vectorial force */
1114             tx               = _mm_mul_pd(fscal,dx11);
1115             ty               = _mm_mul_pd(fscal,dy11);
1116             tz               = _mm_mul_pd(fscal,dz11);
1117
1118             /* Update vectorial force */
1119             fix1             = _mm_add_pd(fix1,tx);
1120             fiy1             = _mm_add_pd(fiy1,ty);
1121             fiz1             = _mm_add_pd(fiz1,tz);
1122
1123             fjx1             = _mm_add_pd(fjx1,tx);
1124             fjy1             = _mm_add_pd(fjy1,ty);
1125             fjz1             = _mm_add_pd(fjz1,tz);
1126
1127             /**************************
1128              * CALCULATE INTERACTIONS *
1129              **************************/
1130
1131             /* REACTION-FIELD ELECTROSTATICS */
1132             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1133
1134             fscal            = felec;
1135
1136             /* Calculate temporary vectorial force */
1137             tx               = _mm_mul_pd(fscal,dx12);
1138             ty               = _mm_mul_pd(fscal,dy12);
1139             tz               = _mm_mul_pd(fscal,dz12);
1140
1141             /* Update vectorial force */
1142             fix1             = _mm_add_pd(fix1,tx);
1143             fiy1             = _mm_add_pd(fiy1,ty);
1144             fiz1             = _mm_add_pd(fiz1,tz);
1145
1146             fjx2             = _mm_add_pd(fjx2,tx);
1147             fjy2             = _mm_add_pd(fjy2,ty);
1148             fjz2             = _mm_add_pd(fjz2,tz);
1149
1150             /**************************
1151              * CALCULATE INTERACTIONS *
1152              **************************/
1153
1154             /* REACTION-FIELD ELECTROSTATICS */
1155             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1156
1157             fscal            = felec;
1158
1159             /* Calculate temporary vectorial force */
1160             tx               = _mm_mul_pd(fscal,dx13);
1161             ty               = _mm_mul_pd(fscal,dy13);
1162             tz               = _mm_mul_pd(fscal,dz13);
1163
1164             /* Update vectorial force */
1165             fix1             = _mm_add_pd(fix1,tx);
1166             fiy1             = _mm_add_pd(fiy1,ty);
1167             fiz1             = _mm_add_pd(fiz1,tz);
1168
1169             fjx3             = _mm_add_pd(fjx3,tx);
1170             fjy3             = _mm_add_pd(fjy3,ty);
1171             fjz3             = _mm_add_pd(fjz3,tz);
1172
1173             /**************************
1174              * CALCULATE INTERACTIONS *
1175              **************************/
1176
1177             /* REACTION-FIELD ELECTROSTATICS */
1178             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1179
1180             fscal            = felec;
1181
1182             /* Calculate temporary vectorial force */
1183             tx               = _mm_mul_pd(fscal,dx21);
1184             ty               = _mm_mul_pd(fscal,dy21);
1185             tz               = _mm_mul_pd(fscal,dz21);
1186
1187             /* Update vectorial force */
1188             fix2             = _mm_add_pd(fix2,tx);
1189             fiy2             = _mm_add_pd(fiy2,ty);
1190             fiz2             = _mm_add_pd(fiz2,tz);
1191
1192             fjx1             = _mm_add_pd(fjx1,tx);
1193             fjy1             = _mm_add_pd(fjy1,ty);
1194             fjz1             = _mm_add_pd(fjz1,tz);
1195
1196             /**************************
1197              * CALCULATE INTERACTIONS *
1198              **************************/
1199
1200             /* REACTION-FIELD ELECTROSTATICS */
1201             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1202
1203             fscal            = felec;
1204
1205             /* Calculate temporary vectorial force */
1206             tx               = _mm_mul_pd(fscal,dx22);
1207             ty               = _mm_mul_pd(fscal,dy22);
1208             tz               = _mm_mul_pd(fscal,dz22);
1209
1210             /* Update vectorial force */
1211             fix2             = _mm_add_pd(fix2,tx);
1212             fiy2             = _mm_add_pd(fiy2,ty);
1213             fiz2             = _mm_add_pd(fiz2,tz);
1214
1215             fjx2             = _mm_add_pd(fjx2,tx);
1216             fjy2             = _mm_add_pd(fjy2,ty);
1217             fjz2             = _mm_add_pd(fjz2,tz);
1218
1219             /**************************
1220              * CALCULATE INTERACTIONS *
1221              **************************/
1222
1223             /* REACTION-FIELD ELECTROSTATICS */
1224             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1225
1226             fscal            = felec;
1227
1228             /* Calculate temporary vectorial force */
1229             tx               = _mm_mul_pd(fscal,dx23);
1230             ty               = _mm_mul_pd(fscal,dy23);
1231             tz               = _mm_mul_pd(fscal,dz23);
1232
1233             /* Update vectorial force */
1234             fix2             = _mm_add_pd(fix2,tx);
1235             fiy2             = _mm_add_pd(fiy2,ty);
1236             fiz2             = _mm_add_pd(fiz2,tz);
1237
1238             fjx3             = _mm_add_pd(fjx3,tx);
1239             fjy3             = _mm_add_pd(fjy3,ty);
1240             fjz3             = _mm_add_pd(fjz3,tz);
1241
1242             /**************************
1243              * CALCULATE INTERACTIONS *
1244              **************************/
1245
1246             /* REACTION-FIELD ELECTROSTATICS */
1247             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1248
1249             fscal            = felec;
1250
1251             /* Calculate temporary vectorial force */
1252             tx               = _mm_mul_pd(fscal,dx31);
1253             ty               = _mm_mul_pd(fscal,dy31);
1254             tz               = _mm_mul_pd(fscal,dz31);
1255
1256             /* Update vectorial force */
1257             fix3             = _mm_add_pd(fix3,tx);
1258             fiy3             = _mm_add_pd(fiy3,ty);
1259             fiz3             = _mm_add_pd(fiz3,tz);
1260
1261             fjx1             = _mm_add_pd(fjx1,tx);
1262             fjy1             = _mm_add_pd(fjy1,ty);
1263             fjz1             = _mm_add_pd(fjz1,tz);
1264
1265             /**************************
1266              * CALCULATE INTERACTIONS *
1267              **************************/
1268
1269             /* REACTION-FIELD ELECTROSTATICS */
1270             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1271
1272             fscal            = felec;
1273
1274             /* Calculate temporary vectorial force */
1275             tx               = _mm_mul_pd(fscal,dx32);
1276             ty               = _mm_mul_pd(fscal,dy32);
1277             tz               = _mm_mul_pd(fscal,dz32);
1278
1279             /* Update vectorial force */
1280             fix3             = _mm_add_pd(fix3,tx);
1281             fiy3             = _mm_add_pd(fiy3,ty);
1282             fiz3             = _mm_add_pd(fiz3,tz);
1283
1284             fjx2             = _mm_add_pd(fjx2,tx);
1285             fjy2             = _mm_add_pd(fjy2,ty);
1286             fjz2             = _mm_add_pd(fjz2,tz);
1287
1288             /**************************
1289              * CALCULATE INTERACTIONS *
1290              **************************/
1291
1292             /* REACTION-FIELD ELECTROSTATICS */
1293             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1294
1295             fscal            = felec;
1296
1297             /* Calculate temporary vectorial force */
1298             tx               = _mm_mul_pd(fscal,dx33);
1299             ty               = _mm_mul_pd(fscal,dy33);
1300             tz               = _mm_mul_pd(fscal,dz33);
1301
1302             /* Update vectorial force */
1303             fix3             = _mm_add_pd(fix3,tx);
1304             fiy3             = _mm_add_pd(fiy3,ty);
1305             fiz3             = _mm_add_pd(fiz3,tz);
1306
1307             fjx3             = _mm_add_pd(fjx3,tx);
1308             fjy3             = _mm_add_pd(fjy3,ty);
1309             fjz3             = _mm_add_pd(fjz3,tz);
1310
1311             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);
1312
1313             /* Inner loop uses 243 flops */
1314         }
1315
1316         if(jidx<j_index_end)
1317         {
1318
1319             jnrA             = jjnr[jidx];
1320             j_coord_offsetA  = DIM*jnrA;
1321
1322             /* load j atom coordinates */
1323             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1324                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1325
1326             /* Calculate displacement vector */
1327             dx11             = _mm_sub_pd(ix1,jx1);
1328             dy11             = _mm_sub_pd(iy1,jy1);
1329             dz11             = _mm_sub_pd(iz1,jz1);
1330             dx12             = _mm_sub_pd(ix1,jx2);
1331             dy12             = _mm_sub_pd(iy1,jy2);
1332             dz12             = _mm_sub_pd(iz1,jz2);
1333             dx13             = _mm_sub_pd(ix1,jx3);
1334             dy13             = _mm_sub_pd(iy1,jy3);
1335             dz13             = _mm_sub_pd(iz1,jz3);
1336             dx21             = _mm_sub_pd(ix2,jx1);
1337             dy21             = _mm_sub_pd(iy2,jy1);
1338             dz21             = _mm_sub_pd(iz2,jz1);
1339             dx22             = _mm_sub_pd(ix2,jx2);
1340             dy22             = _mm_sub_pd(iy2,jy2);
1341             dz22             = _mm_sub_pd(iz2,jz2);
1342             dx23             = _mm_sub_pd(ix2,jx3);
1343             dy23             = _mm_sub_pd(iy2,jy3);
1344             dz23             = _mm_sub_pd(iz2,jz3);
1345             dx31             = _mm_sub_pd(ix3,jx1);
1346             dy31             = _mm_sub_pd(iy3,jy1);
1347             dz31             = _mm_sub_pd(iz3,jz1);
1348             dx32             = _mm_sub_pd(ix3,jx2);
1349             dy32             = _mm_sub_pd(iy3,jy2);
1350             dz32             = _mm_sub_pd(iz3,jz2);
1351             dx33             = _mm_sub_pd(ix3,jx3);
1352             dy33             = _mm_sub_pd(iy3,jy3);
1353             dz33             = _mm_sub_pd(iz3,jz3);
1354
1355             /* Calculate squared distance and things based on it */
1356             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1357             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1358             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1359             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1360             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1361             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1362             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1363             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1364             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1365
1366             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1367             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1368             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1369             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1370             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1371             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1372             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1373             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1374             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1375
1376             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1377             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1378             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1379             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1380             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1381             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1382             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1383             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1384             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1385
1386             fjx1             = _mm_setzero_pd();
1387             fjy1             = _mm_setzero_pd();
1388             fjz1             = _mm_setzero_pd();
1389             fjx2             = _mm_setzero_pd();
1390             fjy2             = _mm_setzero_pd();
1391             fjz2             = _mm_setzero_pd();
1392             fjx3             = _mm_setzero_pd();
1393             fjy3             = _mm_setzero_pd();
1394             fjz3             = _mm_setzero_pd();
1395
1396             /**************************
1397              * CALCULATE INTERACTIONS *
1398              **************************/
1399
1400             /* REACTION-FIELD ELECTROSTATICS */
1401             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1402
1403             fscal            = felec;
1404
1405             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1406
1407             /* Calculate temporary vectorial force */
1408             tx               = _mm_mul_pd(fscal,dx11);
1409             ty               = _mm_mul_pd(fscal,dy11);
1410             tz               = _mm_mul_pd(fscal,dz11);
1411
1412             /* Update vectorial force */
1413             fix1             = _mm_add_pd(fix1,tx);
1414             fiy1             = _mm_add_pd(fiy1,ty);
1415             fiz1             = _mm_add_pd(fiz1,tz);
1416
1417             fjx1             = _mm_add_pd(fjx1,tx);
1418             fjy1             = _mm_add_pd(fjy1,ty);
1419             fjz1             = _mm_add_pd(fjz1,tz);
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             /* REACTION-FIELD ELECTROSTATICS */
1426             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1427
1428             fscal            = felec;
1429
1430             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1431
1432             /* Calculate temporary vectorial force */
1433             tx               = _mm_mul_pd(fscal,dx12);
1434             ty               = _mm_mul_pd(fscal,dy12);
1435             tz               = _mm_mul_pd(fscal,dz12);
1436
1437             /* Update vectorial force */
1438             fix1             = _mm_add_pd(fix1,tx);
1439             fiy1             = _mm_add_pd(fiy1,ty);
1440             fiz1             = _mm_add_pd(fiz1,tz);
1441
1442             fjx2             = _mm_add_pd(fjx2,tx);
1443             fjy2             = _mm_add_pd(fjy2,ty);
1444             fjz2             = _mm_add_pd(fjz2,tz);
1445
1446             /**************************
1447              * CALCULATE INTERACTIONS *
1448              **************************/
1449
1450             /* REACTION-FIELD ELECTROSTATICS */
1451             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1452
1453             fscal            = felec;
1454
1455             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1456
1457             /* Calculate temporary vectorial force */
1458             tx               = _mm_mul_pd(fscal,dx13);
1459             ty               = _mm_mul_pd(fscal,dy13);
1460             tz               = _mm_mul_pd(fscal,dz13);
1461
1462             /* Update vectorial force */
1463             fix1             = _mm_add_pd(fix1,tx);
1464             fiy1             = _mm_add_pd(fiy1,ty);
1465             fiz1             = _mm_add_pd(fiz1,tz);
1466
1467             fjx3             = _mm_add_pd(fjx3,tx);
1468             fjy3             = _mm_add_pd(fjy3,ty);
1469             fjz3             = _mm_add_pd(fjz3,tz);
1470
1471             /**************************
1472              * CALCULATE INTERACTIONS *
1473              **************************/
1474
1475             /* REACTION-FIELD ELECTROSTATICS */
1476             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1477
1478             fscal            = felec;
1479
1480             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1481
1482             /* Calculate temporary vectorial force */
1483             tx               = _mm_mul_pd(fscal,dx21);
1484             ty               = _mm_mul_pd(fscal,dy21);
1485             tz               = _mm_mul_pd(fscal,dz21);
1486
1487             /* Update vectorial force */
1488             fix2             = _mm_add_pd(fix2,tx);
1489             fiy2             = _mm_add_pd(fiy2,ty);
1490             fiz2             = _mm_add_pd(fiz2,tz);
1491
1492             fjx1             = _mm_add_pd(fjx1,tx);
1493             fjy1             = _mm_add_pd(fjy1,ty);
1494             fjz1             = _mm_add_pd(fjz1,tz);
1495
1496             /**************************
1497              * CALCULATE INTERACTIONS *
1498              **************************/
1499
1500             /* REACTION-FIELD ELECTROSTATICS */
1501             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1502
1503             fscal            = felec;
1504
1505             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1506
1507             /* Calculate temporary vectorial force */
1508             tx               = _mm_mul_pd(fscal,dx22);
1509             ty               = _mm_mul_pd(fscal,dy22);
1510             tz               = _mm_mul_pd(fscal,dz22);
1511
1512             /* Update vectorial force */
1513             fix2             = _mm_add_pd(fix2,tx);
1514             fiy2             = _mm_add_pd(fiy2,ty);
1515             fiz2             = _mm_add_pd(fiz2,tz);
1516
1517             fjx2             = _mm_add_pd(fjx2,tx);
1518             fjy2             = _mm_add_pd(fjy2,ty);
1519             fjz2             = _mm_add_pd(fjz2,tz);
1520
1521             /**************************
1522              * CALCULATE INTERACTIONS *
1523              **************************/
1524
1525             /* REACTION-FIELD ELECTROSTATICS */
1526             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1527
1528             fscal            = felec;
1529
1530             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1531
1532             /* Calculate temporary vectorial force */
1533             tx               = _mm_mul_pd(fscal,dx23);
1534             ty               = _mm_mul_pd(fscal,dy23);
1535             tz               = _mm_mul_pd(fscal,dz23);
1536
1537             /* Update vectorial force */
1538             fix2             = _mm_add_pd(fix2,tx);
1539             fiy2             = _mm_add_pd(fiy2,ty);
1540             fiz2             = _mm_add_pd(fiz2,tz);
1541
1542             fjx3             = _mm_add_pd(fjx3,tx);
1543             fjy3             = _mm_add_pd(fjy3,ty);
1544             fjz3             = _mm_add_pd(fjz3,tz);
1545
1546             /**************************
1547              * CALCULATE INTERACTIONS *
1548              **************************/
1549
1550             /* REACTION-FIELD ELECTROSTATICS */
1551             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1552
1553             fscal            = felec;
1554
1555             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1556
1557             /* Calculate temporary vectorial force */
1558             tx               = _mm_mul_pd(fscal,dx31);
1559             ty               = _mm_mul_pd(fscal,dy31);
1560             tz               = _mm_mul_pd(fscal,dz31);
1561
1562             /* Update vectorial force */
1563             fix3             = _mm_add_pd(fix3,tx);
1564             fiy3             = _mm_add_pd(fiy3,ty);
1565             fiz3             = _mm_add_pd(fiz3,tz);
1566
1567             fjx1             = _mm_add_pd(fjx1,tx);
1568             fjy1             = _mm_add_pd(fjy1,ty);
1569             fjz1             = _mm_add_pd(fjz1,tz);
1570
1571             /**************************
1572              * CALCULATE INTERACTIONS *
1573              **************************/
1574
1575             /* REACTION-FIELD ELECTROSTATICS */
1576             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1577
1578             fscal            = felec;
1579
1580             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1581
1582             /* Calculate temporary vectorial force */
1583             tx               = _mm_mul_pd(fscal,dx32);
1584             ty               = _mm_mul_pd(fscal,dy32);
1585             tz               = _mm_mul_pd(fscal,dz32);
1586
1587             /* Update vectorial force */
1588             fix3             = _mm_add_pd(fix3,tx);
1589             fiy3             = _mm_add_pd(fiy3,ty);
1590             fiz3             = _mm_add_pd(fiz3,tz);
1591
1592             fjx2             = _mm_add_pd(fjx2,tx);
1593             fjy2             = _mm_add_pd(fjy2,ty);
1594             fjz2             = _mm_add_pd(fjz2,tz);
1595
1596             /**************************
1597              * CALCULATE INTERACTIONS *
1598              **************************/
1599
1600             /* REACTION-FIELD ELECTROSTATICS */
1601             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1602
1603             fscal            = felec;
1604
1605             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1606
1607             /* Calculate temporary vectorial force */
1608             tx               = _mm_mul_pd(fscal,dx33);
1609             ty               = _mm_mul_pd(fscal,dy33);
1610             tz               = _mm_mul_pd(fscal,dz33);
1611
1612             /* Update vectorial force */
1613             fix3             = _mm_add_pd(fix3,tx);
1614             fiy3             = _mm_add_pd(fiy3,ty);
1615             fiz3             = _mm_add_pd(fiz3,tz);
1616
1617             fjx3             = _mm_add_pd(fjx3,tx);
1618             fjy3             = _mm_add_pd(fjy3,ty);
1619             fjz3             = _mm_add_pd(fjz3,tz);
1620
1621             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1622
1623             /* Inner loop uses 243 flops */
1624         }
1625
1626         /* End of innermost loop */
1627
1628         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1629                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1630
1631         /* Increment number of inner iterations */
1632         inneriter                  += j_index_end - j_index_start;
1633
1634         /* Outer loop uses 18 flops */
1635     }
1636
1637     /* Increment number of outer iterations */
1638     outeriter        += nri;
1639
1640     /* Update outer/inner flops */
1641
1642     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1643 }