src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwLJ_GeomW3W3_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_VdwLJ_GeomW3W3_VF_sse4_1_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwLJ_GeomW3W3_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              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 110     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 111     __m128d          dummy_mask,cutoff_mask;
 112     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 113     __m128d          one     = _mm_set1_pd(1.0);
 114     __m128d          two     = _mm_set1_pd(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_pd(fr->epsfac);
 127     charge           = mdatoms->chargeA;
 128     krf              = _mm_set1_pd(fr->ic->k_rf);
 129     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 130     crf              = _mm_set1_pd(fr->ic->c_rf);
 131     nvdwtype         = fr->ntype;
 132     vdwparam         = fr->nbfp;
 133     vdwtype          = mdatoms->typeA;
 134
 135     /* Setup water-specific parameters */
 136     inr              = nlist->iinr[0];
 137     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 138     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 139     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 140     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 141
 142     jq0              = _mm_set1_pd(charge[inr+0]);
 143     jq1              = _mm_set1_pd(charge[inr+1]);
 144     jq2              = _mm_set1_pd(charge[inr+2]);
 145     vdwjidx0A        = 2*vdwtype[inr+0];
 146     qq00             = _mm_mul_pd(iq0,jq0);
 147     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 148     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 149     qq01             = _mm_mul_pd(iq0,jq1);
 150     qq02             = _mm_mul_pd(iq0,jq2);
 151     qq10             = _mm_mul_pd(iq1,jq0);
 152     qq11             = _mm_mul_pd(iq1,jq1);
 153     qq12             = _mm_mul_pd(iq1,jq2);
 154     qq20             = _mm_mul_pd(iq2,jq0);
 155     qq21             = _mm_mul_pd(iq2,jq1);
 156     qq22             = _mm_mul_pd(iq2,jq2);
 157
 158     /* Avoid stupid compiler warnings */
 159     jnrA = jnrB = 0;
 160     j_coord_offsetA = 0;
 161     j_coord_offsetB = 0;
 162
 163     outeriter        = 0;
 164     inneriter        = 0;
 165
 166     /* Start outer loop over neighborlists */
 167     for(iidx=0; iidx<nri; iidx++)
 168     {
 169         /* Load shift vector for this list */
 170         i_shift_offset   = DIM*shiftidx[iidx];
 171
 172         /* Load limits for loop over neighbors */
 173         j_index_start    = jindex[iidx];
 174         j_index_end      = jindex[iidx+1];
 175
 176         /* Get outer coordinate index */
 177         inr              = iinr[iidx];
 178         i_coord_offset   = DIM*inr;
 179
 180         /* Load i particle coords and add shift vector */
 181         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 182                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 183
 184         fix0             = _mm_setzero_pd();
 185         fiy0             = _mm_setzero_pd();
 186         fiz0             = _mm_setzero_pd();
 187         fix1             = _mm_setzero_pd();
 188         fiy1             = _mm_setzero_pd();
 189         fiz1             = _mm_setzero_pd();
 190         fix2             = _mm_setzero_pd();
 191         fiy2             = _mm_setzero_pd();
 192         fiz2             = _mm_setzero_pd();
 193
 194         /* Reset potential sums */
 195         velecsum         = _mm_setzero_pd();
 196         vvdwsum          = _mm_setzero_pd();
 197
 198         /* Start inner kernel loop */
 199         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 200         {
 201
 202             /* Get j neighbor index, and coordinate index */
 203             jnrA             = jjnr[jidx];
 204             jnrB             = jjnr[jidx+1];
 205             j_coord_offsetA  = DIM*jnrA;
 206             j_coord_offsetB  = DIM*jnrB;
 207
 208             /* load j atom coordinates */
 209             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 210                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 211
 212             /* Calculate displacement vector */
 213             dx00             = _mm_sub_pd(ix0,jx0);
 214             dy00             = _mm_sub_pd(iy0,jy0);
 215             dz00             = _mm_sub_pd(iz0,jz0);
 216             dx01             = _mm_sub_pd(ix0,jx1);
 217             dy01             = _mm_sub_pd(iy0,jy1);
 218             dz01             = _mm_sub_pd(iz0,jz1);
 219             dx02             = _mm_sub_pd(ix0,jx2);
 220             dy02             = _mm_sub_pd(iy0,jy2);
 221             dz02             = _mm_sub_pd(iz0,jz2);
 222             dx10             = _mm_sub_pd(ix1,jx0);
 223             dy10             = _mm_sub_pd(iy1,jy0);
 224             dz10             = _mm_sub_pd(iz1,jz0);
 225             dx11             = _mm_sub_pd(ix1,jx1);
 226             dy11             = _mm_sub_pd(iy1,jy1);
 227             dz11             = _mm_sub_pd(iz1,jz1);
 228             dx12             = _mm_sub_pd(ix1,jx2);
 229             dy12             = _mm_sub_pd(iy1,jy2);
 230             dz12             = _mm_sub_pd(iz1,jz2);
 231             dx20             = _mm_sub_pd(ix2,jx0);
 232             dy20             = _mm_sub_pd(iy2,jy0);
 233             dz20             = _mm_sub_pd(iz2,jz0);
 234             dx21             = _mm_sub_pd(ix2,jx1);
 235             dy21             = _mm_sub_pd(iy2,jy1);
 236             dz21             = _mm_sub_pd(iz2,jz1);
 237             dx22             = _mm_sub_pd(ix2,jx2);
 238             dy22             = _mm_sub_pd(iy2,jy2);
 239             dz22             = _mm_sub_pd(iz2,jz2);
 240
 241             /* Calculate squared distance and things based on it */
 242             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 243             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 244             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 245             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 246             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 247             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 248             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 249             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 250             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 251
 252             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 253             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 254             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 255             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 256             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 257             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 258             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 259             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 260             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 261
 262             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 263             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 264             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 265             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 266             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 267             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 268             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 269             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 270             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 271
 272             fjx0             = _mm_setzero_pd();
 273             fjy0             = _mm_setzero_pd();
 274             fjz0             = _mm_setzero_pd();
 275             fjx1             = _mm_setzero_pd();
 276             fjy1             = _mm_setzero_pd();
 277             fjz1             = _mm_setzero_pd();
 278             fjx2             = _mm_setzero_pd();
 279             fjy2             = _mm_setzero_pd();
 280             fjz2             = _mm_setzero_pd();
 281
 282             /**************************
 283              * CALCULATE INTERACTIONS *
 284              **************************/
 285
 286             /* REACTION-FIELD ELECTROSTATICS */
 287             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 288             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 289
 290             /* LENNARD-JONES DISPERSION/REPULSION */
 291
 292             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 293             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 294             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 295             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 296             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 297
 298             /* Update potential sum for this i atom from the interaction with this j atom. */
 299             velecsum         = _mm_add_pd(velecsum,velec);
 300             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 301
 302             fscal            = _mm_add_pd(felec,fvdw);
 303
 304             /* Calculate temporary vectorial force */
 305             tx               = _mm_mul_pd(fscal,dx00);
 306             ty               = _mm_mul_pd(fscal,dy00);
 307             tz               = _mm_mul_pd(fscal,dz00);
 308
 309             /* Update vectorial force */
 310             fix0             = _mm_add_pd(fix0,tx);
 311             fiy0             = _mm_add_pd(fiy0,ty);
 312             fiz0             = _mm_add_pd(fiz0,tz);
 313
 314             fjx0             = _mm_add_pd(fjx0,tx);
 315             fjy0             = _mm_add_pd(fjy0,ty);
 316             fjz0             = _mm_add_pd(fjz0,tz);
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             /* REACTION-FIELD ELECTROSTATICS */
 323             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 324             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 325
 326             /* Update potential sum for this i atom from the interaction with this j atom. */
 327             velecsum         = _mm_add_pd(velecsum,velec);
 328
 329             fscal            = felec;
 330
 331             /* Calculate temporary vectorial force */
 332             tx               = _mm_mul_pd(fscal,dx01);
 333             ty               = _mm_mul_pd(fscal,dy01);
 334             tz               = _mm_mul_pd(fscal,dz01);
 335
 336             /* Update vectorial force */
 337             fix0             = _mm_add_pd(fix0,tx);
 338             fiy0             = _mm_add_pd(fiy0,ty);
 339             fiz0             = _mm_add_pd(fiz0,tz);
 340
 341             fjx1             = _mm_add_pd(fjx1,tx);
 342             fjy1             = _mm_add_pd(fjy1,ty);
 343             fjz1             = _mm_add_pd(fjz1,tz);
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             /* REACTION-FIELD ELECTROSTATICS */
 350             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 351             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velecsum         = _mm_add_pd(velecsum,velec);
 355
 356             fscal            = felec;
 357
 358             /* Calculate temporary vectorial force */
 359             tx               = _mm_mul_pd(fscal,dx02);
 360             ty               = _mm_mul_pd(fscal,dy02);
 361             tz               = _mm_mul_pd(fscal,dz02);
 362
 363             /* Update vectorial force */
 364             fix0             = _mm_add_pd(fix0,tx);
 365             fiy0             = _mm_add_pd(fiy0,ty);
 366             fiz0             = _mm_add_pd(fiz0,tz);
 367
 368             fjx2             = _mm_add_pd(fjx2,tx);
 369             fjy2             = _mm_add_pd(fjy2,ty);
 370             fjz2             = _mm_add_pd(fjz2,tz);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             /* REACTION-FIELD ELECTROSTATICS */
 377             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 378             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             /* Calculate temporary vectorial force */
 386             tx               = _mm_mul_pd(fscal,dx10);
 387             ty               = _mm_mul_pd(fscal,dy10);
 388             tz               = _mm_mul_pd(fscal,dz10);
 389
 390             /* Update vectorial force */
 391             fix1             = _mm_add_pd(fix1,tx);
 392             fiy1             = _mm_add_pd(fiy1,ty);
 393             fiz1             = _mm_add_pd(fiz1,tz);
 394
 395             fjx0             = _mm_add_pd(fjx0,tx);
 396             fjy0             = _mm_add_pd(fjy0,ty);
 397             fjz0             = _mm_add_pd(fjz0,tz);
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 405             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velecsum         = _mm_add_pd(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412             /* Calculate temporary vectorial force */
 413             tx               = _mm_mul_pd(fscal,dx11);
 414             ty               = _mm_mul_pd(fscal,dy11);
 415             tz               = _mm_mul_pd(fscal,dz11);
 416
 417             /* Update vectorial force */
 418             fix1             = _mm_add_pd(fix1,tx);
 419             fiy1             = _mm_add_pd(fiy1,ty);
 420             fiz1             = _mm_add_pd(fiz1,tz);
 421
 422             fjx1             = _mm_add_pd(fjx1,tx);
 423             fjy1             = _mm_add_pd(fjy1,ty);
 424             fjz1             = _mm_add_pd(fjz1,tz);
 425
 426             /**************************
 427              * CALCULATE INTERACTIONS *
 428              **************************/
 429
 430             /* REACTION-FIELD ELECTROSTATICS */
 431             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 432             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 433
 434             /* Update potential sum for this i atom from the interaction with this j atom. */
 435             velecsum         = _mm_add_pd(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             /* Calculate temporary vectorial force */
 440             tx               = _mm_mul_pd(fscal,dx12);
 441             ty               = _mm_mul_pd(fscal,dy12);
 442             tz               = _mm_mul_pd(fscal,dz12);
 443
 444             /* Update vectorial force */
 445             fix1             = _mm_add_pd(fix1,tx);
 446             fiy1             = _mm_add_pd(fiy1,ty);
 447             fiz1             = _mm_add_pd(fiz1,tz);
 448
 449             fjx2             = _mm_add_pd(fjx2,tx);
 450             fjy2             = _mm_add_pd(fjy2,ty);
 451             fjz2             = _mm_add_pd(fjz2,tz);
 452
 453             /**************************
 454              * CALCULATE INTERACTIONS *
 455              **************************/
 456
 457             /* REACTION-FIELD ELECTROSTATICS */
 458             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 459             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velecsum         = _mm_add_pd(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466             /* Calculate temporary vectorial force */
 467             tx               = _mm_mul_pd(fscal,dx20);
 468             ty               = _mm_mul_pd(fscal,dy20);
 469             tz               = _mm_mul_pd(fscal,dz20);
 470
 471             /* Update vectorial force */
 472             fix2             = _mm_add_pd(fix2,tx);
 473             fiy2             = _mm_add_pd(fiy2,ty);
 474             fiz2             = _mm_add_pd(fiz2,tz);
 475
 476             fjx0             = _mm_add_pd(fjx0,tx);
 477             fjy0             = _mm_add_pd(fjy0,ty);
 478             fjz0             = _mm_add_pd(fjz0,tz);
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             /* REACTION-FIELD ELECTROSTATICS */
 485             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 486             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 487
 488             /* Update potential sum for this i atom from the interaction with this j atom. */
 489             velecsum         = _mm_add_pd(velecsum,velec);
 490
 491             fscal            = felec;
 492
 493             /* Calculate temporary vectorial force */
 494             tx               = _mm_mul_pd(fscal,dx21);
 495             ty               = _mm_mul_pd(fscal,dy21);
 496             tz               = _mm_mul_pd(fscal,dz21);
 497
 498             /* Update vectorial force */
 499             fix2             = _mm_add_pd(fix2,tx);
 500             fiy2             = _mm_add_pd(fiy2,ty);
 501             fiz2             = _mm_add_pd(fiz2,tz);
 502
 503             fjx1             = _mm_add_pd(fjx1,tx);
 504             fjy1             = _mm_add_pd(fjy1,ty);
 505             fjz1             = _mm_add_pd(fjz1,tz);
 506
 507             /**************************
 508              * CALCULATE INTERACTIONS *
 509              **************************/
 510
 511             /* REACTION-FIELD ELECTROSTATICS */
 512             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 513             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 514
 515             /* Update potential sum for this i atom from the interaction with this j atom. */
 516             velecsum         = _mm_add_pd(velecsum,velec);
 517
 518             fscal            = felec;
 519
 520             /* Calculate temporary vectorial force */
 521             tx               = _mm_mul_pd(fscal,dx22);
 522             ty               = _mm_mul_pd(fscal,dy22);
 523             tz               = _mm_mul_pd(fscal,dz22);
 524
 525             /* Update vectorial force */
 526             fix2             = _mm_add_pd(fix2,tx);
 527             fiy2             = _mm_add_pd(fiy2,ty);
 528             fiz2             = _mm_add_pd(fiz2,tz);
 529
 530             fjx2             = _mm_add_pd(fjx2,tx);
 531             fjy2             = _mm_add_pd(fjy2,ty);
 532             fjz2             = _mm_add_pd(fjz2,tz);
 533
 534             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 535
 536             /* Inner loop uses 300 flops */
 537         }
 538
 539         if(jidx<j_index_end)
 540         {
 541
 542             jnrA             = jjnr[jidx];
 543             j_coord_offsetA  = DIM*jnrA;
 544
 545             /* load j atom coordinates */
 546             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 547                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 548
 549             /* Calculate displacement vector */
 550             dx00             = _mm_sub_pd(ix0,jx0);
 551             dy00             = _mm_sub_pd(iy0,jy0);
 552             dz00             = _mm_sub_pd(iz0,jz0);
 553             dx01             = _mm_sub_pd(ix0,jx1);
 554             dy01             = _mm_sub_pd(iy0,jy1);
 555             dz01             = _mm_sub_pd(iz0,jz1);
 556             dx02             = _mm_sub_pd(ix0,jx2);
 557             dy02             = _mm_sub_pd(iy0,jy2);
 558             dz02             = _mm_sub_pd(iz0,jz2);
 559             dx10             = _mm_sub_pd(ix1,jx0);
 560             dy10             = _mm_sub_pd(iy1,jy0);
 561             dz10             = _mm_sub_pd(iz1,jz0);
 562             dx11             = _mm_sub_pd(ix1,jx1);
 563             dy11             = _mm_sub_pd(iy1,jy1);
 564             dz11             = _mm_sub_pd(iz1,jz1);
 565             dx12             = _mm_sub_pd(ix1,jx2);
 566             dy12             = _mm_sub_pd(iy1,jy2);
 567             dz12             = _mm_sub_pd(iz1,jz2);
 568             dx20             = _mm_sub_pd(ix2,jx0);
 569             dy20             = _mm_sub_pd(iy2,jy0);
 570             dz20             = _mm_sub_pd(iz2,jz0);
 571             dx21             = _mm_sub_pd(ix2,jx1);
 572             dy21             = _mm_sub_pd(iy2,jy1);
 573             dz21             = _mm_sub_pd(iz2,jz1);
 574             dx22             = _mm_sub_pd(ix2,jx2);
 575             dy22             = _mm_sub_pd(iy2,jy2);
 576             dz22             = _mm_sub_pd(iz2,jz2);
 577
 578             /* Calculate squared distance and things based on it */
 579             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 580             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 581             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 582             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 583             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 584             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 585             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 586             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 587             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 588
 589             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 590             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 591             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 592             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 593             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 594             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 595             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 596             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 597             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 598
 599             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 600             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 601             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 602             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 603             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 604             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 605             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 606             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 607             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 608
 609             fjx0             = _mm_setzero_pd();
 610             fjy0             = _mm_setzero_pd();
 611             fjz0             = _mm_setzero_pd();
 612             fjx1             = _mm_setzero_pd();
 613             fjy1             = _mm_setzero_pd();
 614             fjz1             = _mm_setzero_pd();
 615             fjx2             = _mm_setzero_pd();
 616             fjy2             = _mm_setzero_pd();
 617             fjz2             = _mm_setzero_pd();
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             /* REACTION-FIELD ELECTROSTATICS */
 624             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 625             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 626
 627             /* LENNARD-JONES DISPERSION/REPULSION */
 628
 629             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 630             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 631             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 632             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 633             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 634
 635             /* Update potential sum for this i atom from the interaction with this j atom. */
 636             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 637             velecsum         = _mm_add_pd(velecsum,velec);
 638             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 639             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 640
 641             fscal            = _mm_add_pd(felec,fvdw);
 642
 643             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 644
 645             /* Calculate temporary vectorial force */
 646             tx               = _mm_mul_pd(fscal,dx00);
 647             ty               = _mm_mul_pd(fscal,dy00);
 648             tz               = _mm_mul_pd(fscal,dz00);
 649
 650             /* Update vectorial force */
 651             fix0             = _mm_add_pd(fix0,tx);
 652             fiy0             = _mm_add_pd(fiy0,ty);
 653             fiz0             = _mm_add_pd(fiz0,tz);
 654
 655             fjx0             = _mm_add_pd(fjx0,tx);
 656             fjy0             = _mm_add_pd(fjy0,ty);
 657             fjz0             = _mm_add_pd(fjz0,tz);
 658
 659             /**************************
 660              * CALCULATE INTERACTIONS *
 661              **************************/
 662
 663             /* REACTION-FIELD ELECTROSTATICS */
 664             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 665             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 666
 667             /* Update potential sum for this i atom from the interaction with this j atom. */
 668             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 669             velecsum         = _mm_add_pd(velecsum,velec);
 670
 671             fscal            = felec;
 672
 673             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 674
 675             /* Calculate temporary vectorial force */
 676             tx               = _mm_mul_pd(fscal,dx01);
 677             ty               = _mm_mul_pd(fscal,dy01);
 678             tz               = _mm_mul_pd(fscal,dz01);
 679
 680             /* Update vectorial force */
 681             fix0             = _mm_add_pd(fix0,tx);
 682             fiy0             = _mm_add_pd(fiy0,ty);
 683             fiz0             = _mm_add_pd(fiz0,tz);
 684
 685             fjx1             = _mm_add_pd(fjx1,tx);
 686             fjy1             = _mm_add_pd(fjy1,ty);
 687             fjz1             = _mm_add_pd(fjz1,tz);
 688
 689             /**************************
 690              * CALCULATE INTERACTIONS *
 691              **************************/
 692
 693             /* REACTION-FIELD ELECTROSTATICS */
 694             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 695             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 696
 697             /* Update potential sum for this i atom from the interaction with this j atom. */
 698             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 699             velecsum         = _mm_add_pd(velecsum,velec);
 700
 701             fscal            = felec;
 702
 703             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 704
 705             /* Calculate temporary vectorial force */
 706             tx               = _mm_mul_pd(fscal,dx02);
 707             ty               = _mm_mul_pd(fscal,dy02);
 708             tz               = _mm_mul_pd(fscal,dz02);
 709
 710             /* Update vectorial force */
 711             fix0             = _mm_add_pd(fix0,tx);
 712             fiy0             = _mm_add_pd(fiy0,ty);
 713             fiz0             = _mm_add_pd(fiz0,tz);
 714
 715             fjx2             = _mm_add_pd(fjx2,tx);
 716             fjy2             = _mm_add_pd(fjy2,ty);
 717             fjz2             = _mm_add_pd(fjz2,tz);
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 725             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 729             velecsum         = _mm_add_pd(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 734
 735             /* Calculate temporary vectorial force */
 736             tx               = _mm_mul_pd(fscal,dx10);
 737             ty               = _mm_mul_pd(fscal,dy10);
 738             tz               = _mm_mul_pd(fscal,dz10);
 739
 740             /* Update vectorial force */
 741             fix1             = _mm_add_pd(fix1,tx);
 742             fiy1             = _mm_add_pd(fiy1,ty);
 743             fiz1             = _mm_add_pd(fiz1,tz);
 744
 745             fjx0             = _mm_add_pd(fjx0,tx);
 746             fjy0             = _mm_add_pd(fjy0,ty);
 747             fjz0             = _mm_add_pd(fjz0,tz);
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             /* REACTION-FIELD ELECTROSTATICS */
 754             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 755             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 756
 757             /* Update potential sum for this i atom from the interaction with this j atom. */
 758             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 759             velecsum         = _mm_add_pd(velecsum,velec);
 760
 761             fscal            = felec;
 762
 763             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 764
 765             /* Calculate temporary vectorial force */
 766             tx               = _mm_mul_pd(fscal,dx11);
 767             ty               = _mm_mul_pd(fscal,dy11);
 768             tz               = _mm_mul_pd(fscal,dz11);
 769
 770             /* Update vectorial force */
 771             fix1             = _mm_add_pd(fix1,tx);
 772             fiy1             = _mm_add_pd(fiy1,ty);
 773             fiz1             = _mm_add_pd(fiz1,tz);
 774
 775             fjx1             = _mm_add_pd(fjx1,tx);
 776             fjy1             = _mm_add_pd(fjy1,ty);
 777             fjz1             = _mm_add_pd(fjz1,tz);
 778
 779             /**************************
 780              * CALCULATE INTERACTIONS *
 781              **************************/
 782
 783             /* REACTION-FIELD ELECTROSTATICS */
 784             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 785             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 786
 787             /* Update potential sum for this i atom from the interaction with this j atom. */
 788             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 789             velecsum         = _mm_add_pd(velecsum,velec);
 790
 791             fscal            = felec;
 792
 793             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 794
 795             /* Calculate temporary vectorial force */
 796             tx               = _mm_mul_pd(fscal,dx12);
 797             ty               = _mm_mul_pd(fscal,dy12);
 798             tz               = _mm_mul_pd(fscal,dz12);
 799
 800             /* Update vectorial force */
 801             fix1             = _mm_add_pd(fix1,tx);
 802             fiy1             = _mm_add_pd(fiy1,ty);
 803             fiz1             = _mm_add_pd(fiz1,tz);
 804
 805             fjx2             = _mm_add_pd(fjx2,tx);
 806             fjy2             = _mm_add_pd(fjy2,ty);
 807             fjz2             = _mm_add_pd(fjz2,tz);
 808
 809             /**************************
 810              * CALCULATE INTERACTIONS *
 811              **************************/
 812
 813             /* REACTION-FIELD ELECTROSTATICS */
 814             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 815             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 816
 817             /* Update potential sum for this i atom from the interaction with this j atom. */
 818             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 819             velecsum         = _mm_add_pd(velecsum,velec);
 820
 821             fscal            = felec;
 822
 823             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 824
 825             /* Calculate temporary vectorial force */
 826             tx               = _mm_mul_pd(fscal,dx20);
 827             ty               = _mm_mul_pd(fscal,dy20);
 828             tz               = _mm_mul_pd(fscal,dz20);
 829
 830             /* Update vectorial force */
 831             fix2             = _mm_add_pd(fix2,tx);
 832             fiy2             = _mm_add_pd(fiy2,ty);
 833             fiz2             = _mm_add_pd(fiz2,tz);
 834
 835             fjx0             = _mm_add_pd(fjx0,tx);
 836             fjy0             = _mm_add_pd(fjy0,ty);
 837             fjz0             = _mm_add_pd(fjz0,tz);
 838
 839             /**************************
 840              * CALCULATE INTERACTIONS *
 841              **************************/
 842
 843             /* REACTION-FIELD ELECTROSTATICS */
 844             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 845             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 846
 847             /* Update potential sum for this i atom from the interaction with this j atom. */
 848             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 849             velecsum         = _mm_add_pd(velecsum,velec);
 850
 851             fscal            = felec;
 852
 853             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 854
 855             /* Calculate temporary vectorial force */
 856             tx               = _mm_mul_pd(fscal,dx21);
 857             ty               = _mm_mul_pd(fscal,dy21);
 858             tz               = _mm_mul_pd(fscal,dz21);
 859
 860             /* Update vectorial force */
 861             fix2             = _mm_add_pd(fix2,tx);
 862             fiy2             = _mm_add_pd(fiy2,ty);
 863             fiz2             = _mm_add_pd(fiz2,tz);
 864
 865             fjx1             = _mm_add_pd(fjx1,tx);
 866             fjy1             = _mm_add_pd(fjy1,ty);
 867             fjz1             = _mm_add_pd(fjz1,tz);
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             /* REACTION-FIELD ELECTROSTATICS */
 874             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 875             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 876
 877             /* Update potential sum for this i atom from the interaction with this j atom. */
 878             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 879             velecsum         = _mm_add_pd(velecsum,velec);
 880
 881             fscal            = felec;
 882
 883             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 884
 885             /* Calculate temporary vectorial force */
 886             tx               = _mm_mul_pd(fscal,dx22);
 887             ty               = _mm_mul_pd(fscal,dy22);
 888             tz               = _mm_mul_pd(fscal,dz22);
 889
 890             /* Update vectorial force */
 891             fix2             = _mm_add_pd(fix2,tx);
 892             fiy2             = _mm_add_pd(fiy2,ty);
 893             fiz2             = _mm_add_pd(fiz2,tz);
 894
 895             fjx2             = _mm_add_pd(fjx2,tx);
 896             fjy2             = _mm_add_pd(fjy2,ty);
 897             fjz2             = _mm_add_pd(fjz2,tz);
 898
 899             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 900
 901             /* Inner loop uses 300 flops */
 902         }
 903
 904         /* End of innermost loop */
 905
 906         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 907                                               f+i_coord_offset,fshift+i_shift_offset);
 908
 909         ggid                        = gid[iidx];
 910         /* Update potential energies */
 911         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 912         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 913
 914         /* Increment number of inner iterations */
 915         inneriter                  += j_index_end - j_index_start;
 916
 917         /* Outer loop uses 20 flops */
 918     }
 919
 920     /* Increment number of outer iterations */
 921     outeriter        += nri;
 922
 923     /* Update outer/inner flops */
 924
 925     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*300);
 926 }
 927 /*
 928  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse4_1_double
 929  * Electrostatics interaction: ReactionField
 930  * VdW interaction:            LennardJones
 931  * Geometry:                   Water3-Water3
 932  * Calculate force/pot:        Force
 933  */
 934 void
 935 nb_kernel_ElecRF_VdwLJ_GeomW3W3_F_sse4_1_double
 936                     (t_nblist                    * gmx_restrict       nlist,
 937                      rvec                        * gmx_restrict          xx,
 938                      rvec                        * gmx_restrict          ff,
 939                      t_forcerec                  * gmx_restrict          fr,
 940                      t_mdatoms                   * gmx_restrict     mdatoms,
 941                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 942                      t_nrnb                      * gmx_restrict        nrnb)
 943 {
 944     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 945      * just 0 for non-waters.
 946      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 947      * jnr indices corresponding to data put in the four positions in the SIMD register.
 948      */
 949     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 950     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 951     int              jnrA,jnrB;
 952     int              j_coord_offsetA,j_coord_offsetB;
 953     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 954     real             rcutoff_scalar;
 955     real             *shiftvec,*fshift,*x,*f;
 956     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 957     int              vdwioffset0;
 958     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 959     int              vdwioffset1;
 960     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 961     int              vdwioffset2;
 962     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 963     int              vdwjidx0A,vdwjidx0B;
 964     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 965     int              vdwjidx1A,vdwjidx1B;
 966     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 967     int              vdwjidx2A,vdwjidx2B;
 968     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 969     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 970     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 971     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 972     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 973     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 974     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 975     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 976     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 977     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 978     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 979     real             *charge;
 980     int              nvdwtype;
 981     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 982     int              *vdwtype;
 983     real             *vdwparam;
 984     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 985     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 986     __m128d          dummy_mask,cutoff_mask;
 987     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 988     __m128d          one     = _mm_set1_pd(1.0);
 989     __m128d          two     = _mm_set1_pd(2.0);
 990     x                = xx[0];
 991     f                = ff[0];
 992
 993     nri              = nlist->nri;
 994     iinr             = nlist->iinr;
 995     jindex           = nlist->jindex;
 996     jjnr             = nlist->jjnr;
 997     shiftidx         = nlist->shift;
 998     gid              = nlist->gid;
 999     shiftvec         = fr->shift_vec[0];
1000     fshift           = fr->fshift[0];
1001     facel            = _mm_set1_pd(fr->epsfac);
1002     charge           = mdatoms->chargeA;
1003     krf              = _mm_set1_pd(fr->ic->k_rf);
1004     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1005     crf              = _mm_set1_pd(fr->ic->c_rf);
1006     nvdwtype         = fr->ntype;
1007     vdwparam         = fr->nbfp;
1008     vdwtype          = mdatoms->typeA;
1009
1010     /* Setup water-specific parameters */
1011     inr              = nlist->iinr[0];
1012     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1013     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1014     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1015     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1016
1017     jq0              = _mm_set1_pd(charge[inr+0]);
1018     jq1              = _mm_set1_pd(charge[inr+1]);
1019     jq2              = _mm_set1_pd(charge[inr+2]);
1020     vdwjidx0A        = 2*vdwtype[inr+0];
1021     qq00             = _mm_mul_pd(iq0,jq0);
1022     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1023     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1024     qq01             = _mm_mul_pd(iq0,jq1);
1025     qq02             = _mm_mul_pd(iq0,jq2);
1026     qq10             = _mm_mul_pd(iq1,jq0);
1027     qq11             = _mm_mul_pd(iq1,jq1);
1028     qq12             = _mm_mul_pd(iq1,jq2);
1029     qq20             = _mm_mul_pd(iq2,jq0);
1030     qq21             = _mm_mul_pd(iq2,jq1);
1031     qq22             = _mm_mul_pd(iq2,jq2);
1032
1033     /* Avoid stupid compiler warnings */
1034     jnrA = jnrB = 0;
1035     j_coord_offsetA = 0;
1036     j_coord_offsetB = 0;
1037
1038     outeriter        = 0;
1039     inneriter        = 0;
1040
1041     /* Start outer loop over neighborlists */
1042     for(iidx=0; iidx<nri; iidx++)
1043     {
1044         /* Load shift vector for this list */
1045         i_shift_offset   = DIM*shiftidx[iidx];
1046
1047         /* Load limits for loop over neighbors */
1048         j_index_start    = jindex[iidx];
1049         j_index_end      = jindex[iidx+1];
1050
1051         /* Get outer coordinate index */
1052         inr              = iinr[iidx];
1053         i_coord_offset   = DIM*inr;
1054
1055         /* Load i particle coords and add shift vector */
1056         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1057                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1058
1059         fix0             = _mm_setzero_pd();
1060         fiy0             = _mm_setzero_pd();
1061         fiz0             = _mm_setzero_pd();
1062         fix1             = _mm_setzero_pd();
1063         fiy1             = _mm_setzero_pd();
1064         fiz1             = _mm_setzero_pd();
1065         fix2             = _mm_setzero_pd();
1066         fiy2             = _mm_setzero_pd();
1067         fiz2             = _mm_setzero_pd();
1068
1069         /* Start inner kernel loop */
1070         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1071         {
1072
1073             /* Get j neighbor index, and coordinate index */
1074             jnrA             = jjnr[jidx];
1075             jnrB             = jjnr[jidx+1];
1076             j_coord_offsetA  = DIM*jnrA;
1077             j_coord_offsetB  = DIM*jnrB;
1078
1079             /* load j atom coordinates */
1080             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1081                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1082
1083             /* Calculate displacement vector */
1084             dx00             = _mm_sub_pd(ix0,jx0);
1085             dy00             = _mm_sub_pd(iy0,jy0);
1086             dz00             = _mm_sub_pd(iz0,jz0);
1087             dx01             = _mm_sub_pd(ix0,jx1);
1088             dy01             = _mm_sub_pd(iy0,jy1);
1089             dz01             = _mm_sub_pd(iz0,jz1);
1090             dx02             = _mm_sub_pd(ix0,jx2);
1091             dy02             = _mm_sub_pd(iy0,jy2);
1092             dz02             = _mm_sub_pd(iz0,jz2);
1093             dx10             = _mm_sub_pd(ix1,jx0);
1094             dy10             = _mm_sub_pd(iy1,jy0);
1095             dz10             = _mm_sub_pd(iz1,jz0);
1096             dx11             = _mm_sub_pd(ix1,jx1);
1097             dy11             = _mm_sub_pd(iy1,jy1);
1098             dz11             = _mm_sub_pd(iz1,jz1);
1099             dx12             = _mm_sub_pd(ix1,jx2);
1100             dy12             = _mm_sub_pd(iy1,jy2);
1101             dz12             = _mm_sub_pd(iz1,jz2);
1102             dx20             = _mm_sub_pd(ix2,jx0);
1103             dy20             = _mm_sub_pd(iy2,jy0);
1104             dz20             = _mm_sub_pd(iz2,jz0);
1105             dx21             = _mm_sub_pd(ix2,jx1);
1106             dy21             = _mm_sub_pd(iy2,jy1);
1107             dz21             = _mm_sub_pd(iz2,jz1);
1108             dx22             = _mm_sub_pd(ix2,jx2);
1109             dy22             = _mm_sub_pd(iy2,jy2);
1110             dz22             = _mm_sub_pd(iz2,jz2);
1111
1112             /* Calculate squared distance and things based on it */
1113             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1114             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1115             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1116             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1117             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1118             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1119             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1120             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1121             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1122
1123             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1124             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1125             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1126             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1127             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1128             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1129             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1130             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1131             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1132
1133             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1134             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1135             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1136             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1137             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1138             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1139             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1140             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1141             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1142
1143             fjx0             = _mm_setzero_pd();
1144             fjy0             = _mm_setzero_pd();
1145             fjz0             = _mm_setzero_pd();
1146             fjx1             = _mm_setzero_pd();
1147             fjy1             = _mm_setzero_pd();
1148             fjz1             = _mm_setzero_pd();
1149             fjx2             = _mm_setzero_pd();
1150             fjy2             = _mm_setzero_pd();
1151             fjz2             = _mm_setzero_pd();
1152
1153             /**************************
1154              * CALCULATE INTERACTIONS *
1155              **************************/
1156
1157             /* REACTION-FIELD ELECTROSTATICS */
1158             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1159
1160             /* LENNARD-JONES DISPERSION/REPULSION */
1161
1162             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1163             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1164
1165             fscal            = _mm_add_pd(felec,fvdw);
1166
1167             /* Calculate temporary vectorial force */
1168             tx               = _mm_mul_pd(fscal,dx00);
1169             ty               = _mm_mul_pd(fscal,dy00);
1170             tz               = _mm_mul_pd(fscal,dz00);
1171
1172             /* Update vectorial force */
1173             fix0             = _mm_add_pd(fix0,tx);
1174             fiy0             = _mm_add_pd(fiy0,ty);
1175             fiz0             = _mm_add_pd(fiz0,tz);
1176
1177             fjx0             = _mm_add_pd(fjx0,tx);
1178             fjy0             = _mm_add_pd(fjy0,ty);
1179             fjz0             = _mm_add_pd(fjz0,tz);
1180
1181             /**************************
1182              * CALCULATE INTERACTIONS *
1183              **************************/
1184
1185             /* REACTION-FIELD ELECTROSTATICS */
1186             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1187
1188             fscal            = felec;
1189
1190             /* Calculate temporary vectorial force */
1191             tx               = _mm_mul_pd(fscal,dx01);
1192             ty               = _mm_mul_pd(fscal,dy01);
1193             tz               = _mm_mul_pd(fscal,dz01);
1194
1195             /* Update vectorial force */
1196             fix0             = _mm_add_pd(fix0,tx);
1197             fiy0             = _mm_add_pd(fiy0,ty);
1198             fiz0             = _mm_add_pd(fiz0,tz);
1199
1200             fjx1             = _mm_add_pd(fjx1,tx);
1201             fjy1             = _mm_add_pd(fjy1,ty);
1202             fjz1             = _mm_add_pd(fjz1,tz);
1203
1204             /**************************
1205              * CALCULATE INTERACTIONS *
1206              **************************/
1207
1208             /* REACTION-FIELD ELECTROSTATICS */
1209             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1210
1211             fscal            = felec;
1212
1213             /* Calculate temporary vectorial force */
1214             tx               = _mm_mul_pd(fscal,dx02);
1215             ty               = _mm_mul_pd(fscal,dy02);
1216             tz               = _mm_mul_pd(fscal,dz02);
1217
1218             /* Update vectorial force */
1219             fix0             = _mm_add_pd(fix0,tx);
1220             fiy0             = _mm_add_pd(fiy0,ty);
1221             fiz0             = _mm_add_pd(fiz0,tz);
1222
1223             fjx2             = _mm_add_pd(fjx2,tx);
1224             fjy2             = _mm_add_pd(fjy2,ty);
1225             fjz2             = _mm_add_pd(fjz2,tz);
1226
1227             /**************************
1228              * CALCULATE INTERACTIONS *
1229              **************************/
1230
1231             /* REACTION-FIELD ELECTROSTATICS */
1232             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1233
1234             fscal            = felec;
1235
1236             /* Calculate temporary vectorial force */
1237             tx               = _mm_mul_pd(fscal,dx10);
1238             ty               = _mm_mul_pd(fscal,dy10);
1239             tz               = _mm_mul_pd(fscal,dz10);
1240
1241             /* Update vectorial force */
1242             fix1             = _mm_add_pd(fix1,tx);
1243             fiy1             = _mm_add_pd(fiy1,ty);
1244             fiz1             = _mm_add_pd(fiz1,tz);
1245
1246             fjx0             = _mm_add_pd(fjx0,tx);
1247             fjy0             = _mm_add_pd(fjy0,ty);
1248             fjz0             = _mm_add_pd(fjz0,tz);
1249
1250             /**************************
1251              * CALCULATE INTERACTIONS *
1252              **************************/
1253
1254             /* REACTION-FIELD ELECTROSTATICS */
1255             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1256
1257             fscal            = felec;
1258
1259             /* Calculate temporary vectorial force */
1260             tx               = _mm_mul_pd(fscal,dx11);
1261             ty               = _mm_mul_pd(fscal,dy11);
1262             tz               = _mm_mul_pd(fscal,dz11);
1263
1264             /* Update vectorial force */
1265             fix1             = _mm_add_pd(fix1,tx);
1266             fiy1             = _mm_add_pd(fiy1,ty);
1267             fiz1             = _mm_add_pd(fiz1,tz);
1268
1269             fjx1             = _mm_add_pd(fjx1,tx);
1270             fjy1             = _mm_add_pd(fjy1,ty);
1271             fjz1             = _mm_add_pd(fjz1,tz);
1272
1273             /**************************
1274              * CALCULATE INTERACTIONS *
1275              **************************/
1276
1277             /* REACTION-FIELD ELECTROSTATICS */
1278             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1279
1280             fscal            = felec;
1281
1282             /* Calculate temporary vectorial force */
1283             tx               = _mm_mul_pd(fscal,dx12);
1284             ty               = _mm_mul_pd(fscal,dy12);
1285             tz               = _mm_mul_pd(fscal,dz12);
1286
1287             /* Update vectorial force */
1288             fix1             = _mm_add_pd(fix1,tx);
1289             fiy1             = _mm_add_pd(fiy1,ty);
1290             fiz1             = _mm_add_pd(fiz1,tz);
1291
1292             fjx2             = _mm_add_pd(fjx2,tx);
1293             fjy2             = _mm_add_pd(fjy2,ty);
1294             fjz2             = _mm_add_pd(fjz2,tz);
1295
1296             /**************************
1297              * CALCULATE INTERACTIONS *
1298              **************************/
1299
1300             /* REACTION-FIELD ELECTROSTATICS */
1301             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1302
1303             fscal            = felec;
1304
1305             /* Calculate temporary vectorial force */
1306             tx               = _mm_mul_pd(fscal,dx20);
1307             ty               = _mm_mul_pd(fscal,dy20);
1308             tz               = _mm_mul_pd(fscal,dz20);
1309
1310             /* Update vectorial force */
1311             fix2             = _mm_add_pd(fix2,tx);
1312             fiy2             = _mm_add_pd(fiy2,ty);
1313             fiz2             = _mm_add_pd(fiz2,tz);
1314
1315             fjx0             = _mm_add_pd(fjx0,tx);
1316             fjy0             = _mm_add_pd(fjy0,ty);
1317             fjz0             = _mm_add_pd(fjz0,tz);
1318
1319             /**************************
1320              * CALCULATE INTERACTIONS *
1321              **************************/
1322
1323             /* REACTION-FIELD ELECTROSTATICS */
1324             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1325
1326             fscal            = felec;
1327
1328             /* Calculate temporary vectorial force */
1329             tx               = _mm_mul_pd(fscal,dx21);
1330             ty               = _mm_mul_pd(fscal,dy21);
1331             tz               = _mm_mul_pd(fscal,dz21);
1332
1333             /* Update vectorial force */
1334             fix2             = _mm_add_pd(fix2,tx);
1335             fiy2             = _mm_add_pd(fiy2,ty);
1336             fiz2             = _mm_add_pd(fiz2,tz);
1337
1338             fjx1             = _mm_add_pd(fjx1,tx);
1339             fjy1             = _mm_add_pd(fjy1,ty);
1340             fjz1             = _mm_add_pd(fjz1,tz);
1341
1342             /**************************
1343              * CALCULATE INTERACTIONS *
1344              **************************/
1345
1346             /* REACTION-FIELD ELECTROSTATICS */
1347             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1348
1349             fscal            = felec;
1350
1351             /* Calculate temporary vectorial force */
1352             tx               = _mm_mul_pd(fscal,dx22);
1353             ty               = _mm_mul_pd(fscal,dy22);
1354             tz               = _mm_mul_pd(fscal,dz22);
1355
1356             /* Update vectorial force */
1357             fix2             = _mm_add_pd(fix2,tx);
1358             fiy2             = _mm_add_pd(fiy2,ty);
1359             fiz2             = _mm_add_pd(fiz2,tz);
1360
1361             fjx2             = _mm_add_pd(fjx2,tx);
1362             fjy2             = _mm_add_pd(fjy2,ty);
1363             fjz2             = _mm_add_pd(fjz2,tz);
1364
1365             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1366
1367             /* Inner loop uses 250 flops */
1368         }
1369
1370         if(jidx<j_index_end)
1371         {
1372
1373             jnrA             = jjnr[jidx];
1374             j_coord_offsetA  = DIM*jnrA;
1375
1376             /* load j atom coordinates */
1377             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1378                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1379
1380             /* Calculate displacement vector */
1381             dx00             = _mm_sub_pd(ix0,jx0);
1382             dy00             = _mm_sub_pd(iy0,jy0);
1383             dz00             = _mm_sub_pd(iz0,jz0);
1384             dx01             = _mm_sub_pd(ix0,jx1);
1385             dy01             = _mm_sub_pd(iy0,jy1);
1386             dz01             = _mm_sub_pd(iz0,jz1);
1387             dx02             = _mm_sub_pd(ix0,jx2);
1388             dy02             = _mm_sub_pd(iy0,jy2);
1389             dz02             = _mm_sub_pd(iz0,jz2);
1390             dx10             = _mm_sub_pd(ix1,jx0);
1391             dy10             = _mm_sub_pd(iy1,jy0);
1392             dz10             = _mm_sub_pd(iz1,jz0);
1393             dx11             = _mm_sub_pd(ix1,jx1);
1394             dy11             = _mm_sub_pd(iy1,jy1);
1395             dz11             = _mm_sub_pd(iz1,jz1);
1396             dx12             = _mm_sub_pd(ix1,jx2);
1397             dy12             = _mm_sub_pd(iy1,jy2);
1398             dz12             = _mm_sub_pd(iz1,jz2);
1399             dx20             = _mm_sub_pd(ix2,jx0);
1400             dy20             = _mm_sub_pd(iy2,jy0);
1401             dz20             = _mm_sub_pd(iz2,jz0);
1402             dx21             = _mm_sub_pd(ix2,jx1);
1403             dy21             = _mm_sub_pd(iy2,jy1);
1404             dz21             = _mm_sub_pd(iz2,jz1);
1405             dx22             = _mm_sub_pd(ix2,jx2);
1406             dy22             = _mm_sub_pd(iy2,jy2);
1407             dz22             = _mm_sub_pd(iz2,jz2);
1408
1409             /* Calculate squared distance and things based on it */
1410             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1411             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1412             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1413             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1414             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1415             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1416             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1417             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1418             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1419
1420             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1421             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1422             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1423             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1424             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1425             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1426             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1427             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1428             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1429
1430             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1431             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1432             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1433             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1434             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1435             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1436             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1437             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1438             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1439
1440             fjx0             = _mm_setzero_pd();
1441             fjy0             = _mm_setzero_pd();
1442             fjz0             = _mm_setzero_pd();
1443             fjx1             = _mm_setzero_pd();
1444             fjy1             = _mm_setzero_pd();
1445             fjz1             = _mm_setzero_pd();
1446             fjx2             = _mm_setzero_pd();
1447             fjy2             = _mm_setzero_pd();
1448             fjz2             = _mm_setzero_pd();
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             /* REACTION-FIELD ELECTROSTATICS */
1455             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1456
1457             /* LENNARD-JONES DISPERSION/REPULSION */
1458
1459             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1460             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1461
1462             fscal            = _mm_add_pd(felec,fvdw);
1463
1464             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1465
1466             /* Calculate temporary vectorial force */
1467             tx               = _mm_mul_pd(fscal,dx00);
1468             ty               = _mm_mul_pd(fscal,dy00);
1469             tz               = _mm_mul_pd(fscal,dz00);
1470
1471             /* Update vectorial force */
1472             fix0             = _mm_add_pd(fix0,tx);
1473             fiy0             = _mm_add_pd(fiy0,ty);
1474             fiz0             = _mm_add_pd(fiz0,tz);
1475
1476             fjx0             = _mm_add_pd(fjx0,tx);
1477             fjy0             = _mm_add_pd(fjy0,ty);
1478             fjz0             = _mm_add_pd(fjz0,tz);
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             /* REACTION-FIELD ELECTROSTATICS */
1485             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1486
1487             fscal            = felec;
1488
1489             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1490
1491             /* Calculate temporary vectorial force */
1492             tx               = _mm_mul_pd(fscal,dx01);
1493             ty               = _mm_mul_pd(fscal,dy01);
1494             tz               = _mm_mul_pd(fscal,dz01);
1495
1496             /* Update vectorial force */
1497             fix0             = _mm_add_pd(fix0,tx);
1498             fiy0             = _mm_add_pd(fiy0,ty);
1499             fiz0             = _mm_add_pd(fiz0,tz);
1500
1501             fjx1             = _mm_add_pd(fjx1,tx);
1502             fjy1             = _mm_add_pd(fjy1,ty);
1503             fjz1             = _mm_add_pd(fjz1,tz);
1504
1505             /**************************
1506              * CALCULATE INTERACTIONS *
1507              **************************/
1508
1509             /* REACTION-FIELD ELECTROSTATICS */
1510             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1511
1512             fscal            = felec;
1513
1514             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1515
1516             /* Calculate temporary vectorial force */
1517             tx               = _mm_mul_pd(fscal,dx02);
1518             ty               = _mm_mul_pd(fscal,dy02);
1519             tz               = _mm_mul_pd(fscal,dz02);
1520
1521             /* Update vectorial force */
1522             fix0             = _mm_add_pd(fix0,tx);
1523             fiy0             = _mm_add_pd(fiy0,ty);
1524             fiz0             = _mm_add_pd(fiz0,tz);
1525
1526             fjx2             = _mm_add_pd(fjx2,tx);
1527             fjy2             = _mm_add_pd(fjy2,ty);
1528             fjz2             = _mm_add_pd(fjz2,tz);
1529
1530             /**************************
1531              * CALCULATE INTERACTIONS *
1532              **************************/
1533
1534             /* REACTION-FIELD ELECTROSTATICS */
1535             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1536
1537             fscal            = felec;
1538
1539             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1540
1541             /* Calculate temporary vectorial force */
1542             tx               = _mm_mul_pd(fscal,dx10);
1543             ty               = _mm_mul_pd(fscal,dy10);
1544             tz               = _mm_mul_pd(fscal,dz10);
1545
1546             /* Update vectorial force */
1547             fix1             = _mm_add_pd(fix1,tx);
1548             fiy1             = _mm_add_pd(fiy1,ty);
1549             fiz1             = _mm_add_pd(fiz1,tz);
1550
1551             fjx0             = _mm_add_pd(fjx0,tx);
1552             fjy0             = _mm_add_pd(fjy0,ty);
1553             fjz0             = _mm_add_pd(fjz0,tz);
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             /* REACTION-FIELD ELECTROSTATICS */
1560             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1561
1562             fscal            = felec;
1563
1564             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1565
1566             /* Calculate temporary vectorial force */
1567             tx               = _mm_mul_pd(fscal,dx11);
1568             ty               = _mm_mul_pd(fscal,dy11);
1569             tz               = _mm_mul_pd(fscal,dz11);
1570
1571             /* Update vectorial force */
1572             fix1             = _mm_add_pd(fix1,tx);
1573             fiy1             = _mm_add_pd(fiy1,ty);
1574             fiz1             = _mm_add_pd(fiz1,tz);
1575
1576             fjx1             = _mm_add_pd(fjx1,tx);
1577             fjy1             = _mm_add_pd(fjy1,ty);
1578             fjz1             = _mm_add_pd(fjz1,tz);
1579
1580             /**************************
1581              * CALCULATE INTERACTIONS *
1582              **************************/
1583
1584             /* REACTION-FIELD ELECTROSTATICS */
1585             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1586
1587             fscal            = felec;
1588
1589             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1590
1591             /* Calculate temporary vectorial force */
1592             tx               = _mm_mul_pd(fscal,dx12);
1593             ty               = _mm_mul_pd(fscal,dy12);
1594             tz               = _mm_mul_pd(fscal,dz12);
1595
1596             /* Update vectorial force */
1597             fix1             = _mm_add_pd(fix1,tx);
1598             fiy1             = _mm_add_pd(fiy1,ty);
1599             fiz1             = _mm_add_pd(fiz1,tz);
1600
1601             fjx2             = _mm_add_pd(fjx2,tx);
1602             fjy2             = _mm_add_pd(fjy2,ty);
1603             fjz2             = _mm_add_pd(fjz2,tz);
1604
1605             /**************************
1606              * CALCULATE INTERACTIONS *
1607              **************************/
1608
1609             /* REACTION-FIELD ELECTROSTATICS */
1610             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1611
1612             fscal            = felec;
1613
1614             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1615
1616             /* Calculate temporary vectorial force */
1617             tx               = _mm_mul_pd(fscal,dx20);
1618             ty               = _mm_mul_pd(fscal,dy20);
1619             tz               = _mm_mul_pd(fscal,dz20);
1620
1621             /* Update vectorial force */
1622             fix2             = _mm_add_pd(fix2,tx);
1623             fiy2             = _mm_add_pd(fiy2,ty);
1624             fiz2             = _mm_add_pd(fiz2,tz);
1625
1626             fjx0             = _mm_add_pd(fjx0,tx);
1627             fjy0             = _mm_add_pd(fjy0,ty);
1628             fjz0             = _mm_add_pd(fjz0,tz);
1629
1630             /**************************
1631              * CALCULATE INTERACTIONS *
1632              **************************/
1633
1634             /* REACTION-FIELD ELECTROSTATICS */
1635             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1636
1637             fscal            = felec;
1638
1639             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1640
1641             /* Calculate temporary vectorial force */
1642             tx               = _mm_mul_pd(fscal,dx21);
1643             ty               = _mm_mul_pd(fscal,dy21);
1644             tz               = _mm_mul_pd(fscal,dz21);
1645
1646             /* Update vectorial force */
1647             fix2             = _mm_add_pd(fix2,tx);
1648             fiy2             = _mm_add_pd(fiy2,ty);
1649             fiz2             = _mm_add_pd(fiz2,tz);
1650
1651             fjx1             = _mm_add_pd(fjx1,tx);
1652             fjy1             = _mm_add_pd(fjy1,ty);
1653             fjz1             = _mm_add_pd(fjz1,tz);
1654
1655             /**************************
1656              * CALCULATE INTERACTIONS *
1657              **************************/
1658
1659             /* REACTION-FIELD ELECTROSTATICS */
1660             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1661
1662             fscal            = felec;
1663
1664             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1665
1666             /* Calculate temporary vectorial force */
1667             tx               = _mm_mul_pd(fscal,dx22);
1668             ty               = _mm_mul_pd(fscal,dy22);
1669             tz               = _mm_mul_pd(fscal,dz22);
1670
1671             /* Update vectorial force */
1672             fix2             = _mm_add_pd(fix2,tx);
1673             fiy2             = _mm_add_pd(fiy2,ty);
1674             fiz2             = _mm_add_pd(fiz2,tz);
1675
1676             fjx2             = _mm_add_pd(fjx2,tx);
1677             fjy2             = _mm_add_pd(fjy2,ty);
1678             fjz2             = _mm_add_pd(fjz2,tz);
1679
1680             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1681
1682             /* Inner loop uses 250 flops */
1683         }
1684
1685         /* End of innermost loop */
1686
1687         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1688                                               f+i_coord_offset,fshift+i_shift_offset);
1689
1690         /* Increment number of inner iterations */
1691         inneriter                  += j_index_end - j_index_start;
1692
1693         /* Outer loop uses 18 flops */
1694     }
1695
1696     /* Increment number of outer iterations */
1697     outeriter        += nri;
1698
1699     /* Update outer/inner flops */
1700
1701     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1702 }