src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRF_VdwCSTab_GeomW3W3_sse4_1_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse4_1_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_sse4_1_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_sse4_1_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     int              vdwjidx1A,vdwjidx1B;
  88     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 107     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 108     __m128i          vfitab;
 109     __m128i          ifour       = _mm_set1_epi32(4);
 110     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 111     real             *vftab;
 112     __m128d          dummy_mask,cutoff_mask;
 113     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 114     __m128d          one     = _mm_set1_pd(1.0);
 115     __m128d          two     = _mm_set1_pd(2.0);
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = _mm_set1_pd(fr->ic->epsfac);
 128     charge           = mdatoms->chargeA;
 129     krf              = _mm_set1_pd(fr->ic->k_rf);
 130     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 131     crf              = _mm_set1_pd(fr->ic->c_rf);
 132     nvdwtype         = fr->ntype;
 133     vdwparam         = fr->nbfp;
 134     vdwtype          = mdatoms->typeA;
 135
 136     vftab            = kernel_data->table_vdw->data;
 137     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 138
 139     /* Setup water-specific parameters */
 140     inr              = nlist->iinr[0];
 141     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 142     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 143     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 144     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 145
 146     jq0              = _mm_set1_pd(charge[inr+0]);
 147     jq1              = _mm_set1_pd(charge[inr+1]);
 148     jq2              = _mm_set1_pd(charge[inr+2]);
 149     vdwjidx0A        = 2*vdwtype[inr+0];
 150     qq00             = _mm_mul_pd(iq0,jq0);
 151     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq01             = _mm_mul_pd(iq0,jq1);
 154     qq02             = _mm_mul_pd(iq0,jq2);
 155     qq10             = _mm_mul_pd(iq1,jq0);
 156     qq11             = _mm_mul_pd(iq1,jq1);
 157     qq12             = _mm_mul_pd(iq1,jq2);
 158     qq20             = _mm_mul_pd(iq2,jq0);
 159     qq21             = _mm_mul_pd(iq2,jq1);
 160     qq22             = _mm_mul_pd(iq2,jq2);
 161
 162     /* Avoid stupid compiler warnings */
 163     jnrA = jnrB = 0;
 164     j_coord_offsetA = 0;
 165     j_coord_offsetB = 0;
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     /* Start outer loop over neighborlists */
 171     for(iidx=0; iidx<nri; iidx++)
 172     {
 173         /* Load shift vector for this list */
 174         i_shift_offset   = DIM*shiftidx[iidx];
 175
 176         /* Load limits for loop over neighbors */
 177         j_index_start    = jindex[iidx];
 178         j_index_end      = jindex[iidx+1];
 179
 180         /* Get outer coordinate index */
 181         inr              = iinr[iidx];
 182         i_coord_offset   = DIM*inr;
 183
 184         /* Load i particle coords and add shift vector */
 185         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 186                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 187
 188         fix0             = _mm_setzero_pd();
 189         fiy0             = _mm_setzero_pd();
 190         fiz0             = _mm_setzero_pd();
 191         fix1             = _mm_setzero_pd();
 192         fiy1             = _mm_setzero_pd();
 193         fiz1             = _mm_setzero_pd();
 194         fix2             = _mm_setzero_pd();
 195         fiy2             = _mm_setzero_pd();
 196         fiz2             = _mm_setzero_pd();
 197
 198         /* Reset potential sums */
 199         velecsum         = _mm_setzero_pd();
 200         vvdwsum          = _mm_setzero_pd();
 201
 202         /* Start inner kernel loop */
 203         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 204         {
 205
 206             /* Get j neighbor index, and coordinate index */
 207             jnrA             = jjnr[jidx];
 208             jnrB             = jjnr[jidx+1];
 209             j_coord_offsetA  = DIM*jnrA;
 210             j_coord_offsetB  = DIM*jnrB;
 211
 212             /* load j atom coordinates */
 213             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 214                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 215
 216             /* Calculate displacement vector */
 217             dx00             = _mm_sub_pd(ix0,jx0);
 218             dy00             = _mm_sub_pd(iy0,jy0);
 219             dz00             = _mm_sub_pd(iz0,jz0);
 220             dx01             = _mm_sub_pd(ix0,jx1);
 221             dy01             = _mm_sub_pd(iy0,jy1);
 222             dz01             = _mm_sub_pd(iz0,jz1);
 223             dx02             = _mm_sub_pd(ix0,jx2);
 224             dy02             = _mm_sub_pd(iy0,jy2);
 225             dz02             = _mm_sub_pd(iz0,jz2);
 226             dx10             = _mm_sub_pd(ix1,jx0);
 227             dy10             = _mm_sub_pd(iy1,jy0);
 228             dz10             = _mm_sub_pd(iz1,jz0);
 229             dx11             = _mm_sub_pd(ix1,jx1);
 230             dy11             = _mm_sub_pd(iy1,jy1);
 231             dz11             = _mm_sub_pd(iz1,jz1);
 232             dx12             = _mm_sub_pd(ix1,jx2);
 233             dy12             = _mm_sub_pd(iy1,jy2);
 234             dz12             = _mm_sub_pd(iz1,jz2);
 235             dx20             = _mm_sub_pd(ix2,jx0);
 236             dy20             = _mm_sub_pd(iy2,jy0);
 237             dz20             = _mm_sub_pd(iz2,jz0);
 238             dx21             = _mm_sub_pd(ix2,jx1);
 239             dy21             = _mm_sub_pd(iy2,jy1);
 240             dz21             = _mm_sub_pd(iz2,jz1);
 241             dx22             = _mm_sub_pd(ix2,jx2);
 242             dy22             = _mm_sub_pd(iy2,jy2);
 243             dz22             = _mm_sub_pd(iz2,jz2);
 244
 245             /* Calculate squared distance and things based on it */
 246             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 247             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 248             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 249             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 250             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 251             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 252             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 253             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 254             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 255
 256             rinv00           = sse41_invsqrt_d(rsq00);
 257             rinv01           = sse41_invsqrt_d(rsq01);
 258             rinv02           = sse41_invsqrt_d(rsq02);
 259             rinv10           = sse41_invsqrt_d(rsq10);
 260             rinv11           = sse41_invsqrt_d(rsq11);
 261             rinv12           = sse41_invsqrt_d(rsq12);
 262             rinv20           = sse41_invsqrt_d(rsq20);
 263             rinv21           = sse41_invsqrt_d(rsq21);
 264             rinv22           = sse41_invsqrt_d(rsq22);
 265
 266             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 267             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 268             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 269             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 270             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 271             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 272             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 273             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 274             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 275
 276             fjx0             = _mm_setzero_pd();
 277             fjy0             = _mm_setzero_pd();
 278             fjz0             = _mm_setzero_pd();
 279             fjx1             = _mm_setzero_pd();
 280             fjy1             = _mm_setzero_pd();
 281             fjz1             = _mm_setzero_pd();
 282             fjx2             = _mm_setzero_pd();
 283             fjy2             = _mm_setzero_pd();
 284             fjz2             = _mm_setzero_pd();
 285
 286             /**************************
 287              * CALCULATE INTERACTIONS *
 288              **************************/
 289
 290             r00              = _mm_mul_pd(rsq00,rinv00);
 291
 292             /* Calculate table index by multiplying r with table scale and truncate to integer */
 293             rt               = _mm_mul_pd(r00,vftabscale);
 294             vfitab           = _mm_cvttpd_epi32(rt);
 295             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 296             vfitab           = _mm_slli_epi32(vfitab,3);
 297
 298             /* REACTION-FIELD ELECTROSTATICS */
 299             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 300             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 301
 302             /* CUBIC SPLINE TABLE DISPERSION */
 303             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 304             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 305             GMX_MM_TRANSPOSE2_PD(Y,F);
 306             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 307             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 308             GMX_MM_TRANSPOSE2_PD(G,H);
 309             Heps             = _mm_mul_pd(vfeps,H);
 310             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 311             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 312             vvdw6            = _mm_mul_pd(c6_00,VV);
 313             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 314             fvdw6            = _mm_mul_pd(c6_00,FF);
 315
 316             /* CUBIC SPLINE TABLE REPULSION */
 317             vfitab           = _mm_add_epi32(vfitab,ifour);
 318             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 319             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 320             GMX_MM_TRANSPOSE2_PD(Y,F);
 321             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 322             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 323             GMX_MM_TRANSPOSE2_PD(G,H);
 324             Heps             = _mm_mul_pd(vfeps,H);
 325             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 326             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 327             vvdw12           = _mm_mul_pd(c12_00,VV);
 328             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 329             fvdw12           = _mm_mul_pd(c12_00,FF);
 330             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 331             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velecsum         = _mm_add_pd(velecsum,velec);
 335             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 336
 337             fscal            = _mm_add_pd(felec,fvdw);
 338
 339             /* Calculate temporary vectorial force */
 340             tx               = _mm_mul_pd(fscal,dx00);
 341             ty               = _mm_mul_pd(fscal,dy00);
 342             tz               = _mm_mul_pd(fscal,dz00);
 343
 344             /* Update vectorial force */
 345             fix0             = _mm_add_pd(fix0,tx);
 346             fiy0             = _mm_add_pd(fiy0,ty);
 347             fiz0             = _mm_add_pd(fiz0,tz);
 348
 349             fjx0             = _mm_add_pd(fjx0,tx);
 350             fjy0             = _mm_add_pd(fjy0,ty);
 351             fjz0             = _mm_add_pd(fjz0,tz);
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* REACTION-FIELD ELECTROSTATICS */
 358             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 359             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 360
 361             /* Update potential sum for this i atom from the interaction with this j atom. */
 362             velecsum         = _mm_add_pd(velecsum,velec);
 363
 364             fscal            = felec;
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = _mm_mul_pd(fscal,dx01);
 368             ty               = _mm_mul_pd(fscal,dy01);
 369             tz               = _mm_mul_pd(fscal,dz01);
 370
 371             /* Update vectorial force */
 372             fix0             = _mm_add_pd(fix0,tx);
 373             fiy0             = _mm_add_pd(fiy0,ty);
 374             fiz0             = _mm_add_pd(fiz0,tz);
 375
 376             fjx1             = _mm_add_pd(fjx1,tx);
 377             fjy1             = _mm_add_pd(fjy1,ty);
 378             fjz1             = _mm_add_pd(fjz1,tz);
 379
 380             /**************************
 381              * CALCULATE INTERACTIONS *
 382              **************************/
 383
 384             /* REACTION-FIELD ELECTROSTATICS */
 385             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 386             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 387
 388             /* Update potential sum for this i atom from the interaction with this j atom. */
 389             velecsum         = _mm_add_pd(velecsum,velec);
 390
 391             fscal            = felec;
 392
 393             /* Calculate temporary vectorial force */
 394             tx               = _mm_mul_pd(fscal,dx02);
 395             ty               = _mm_mul_pd(fscal,dy02);
 396             tz               = _mm_mul_pd(fscal,dz02);
 397
 398             /* Update vectorial force */
 399             fix0             = _mm_add_pd(fix0,tx);
 400             fiy0             = _mm_add_pd(fiy0,ty);
 401             fiz0             = _mm_add_pd(fiz0,tz);
 402
 403             fjx2             = _mm_add_pd(fjx2,tx);
 404             fjy2             = _mm_add_pd(fjy2,ty);
 405             fjz2             = _mm_add_pd(fjz2,tz);
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             /* REACTION-FIELD ELECTROSTATICS */
 412             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 413             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 414
 415             /* Update potential sum for this i atom from the interaction with this j atom. */
 416             velecsum         = _mm_add_pd(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             /* Calculate temporary vectorial force */
 421             tx               = _mm_mul_pd(fscal,dx10);
 422             ty               = _mm_mul_pd(fscal,dy10);
 423             tz               = _mm_mul_pd(fscal,dz10);
 424
 425             /* Update vectorial force */
 426             fix1             = _mm_add_pd(fix1,tx);
 427             fiy1             = _mm_add_pd(fiy1,ty);
 428             fiz1             = _mm_add_pd(fiz1,tz);
 429
 430             fjx0             = _mm_add_pd(fjx0,tx);
 431             fjy0             = _mm_add_pd(fjy0,ty);
 432             fjz0             = _mm_add_pd(fjz0,tz);
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 440             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velecsum         = _mm_add_pd(velecsum,velec);
 444
 445             fscal            = felec;
 446
 447             /* Calculate temporary vectorial force */
 448             tx               = _mm_mul_pd(fscal,dx11);
 449             ty               = _mm_mul_pd(fscal,dy11);
 450             tz               = _mm_mul_pd(fscal,dz11);
 451
 452             /* Update vectorial force */
 453             fix1             = _mm_add_pd(fix1,tx);
 454             fiy1             = _mm_add_pd(fiy1,ty);
 455             fiz1             = _mm_add_pd(fiz1,tz);
 456
 457             fjx1             = _mm_add_pd(fjx1,tx);
 458             fjy1             = _mm_add_pd(fjy1,ty);
 459             fjz1             = _mm_add_pd(fjz1,tz);
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             /* REACTION-FIELD ELECTROSTATICS */
 466             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 467             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 468
 469             /* Update potential sum for this i atom from the interaction with this j atom. */
 470             velecsum         = _mm_add_pd(velecsum,velec);
 471
 472             fscal            = felec;
 473
 474             /* Calculate temporary vectorial force */
 475             tx               = _mm_mul_pd(fscal,dx12);
 476             ty               = _mm_mul_pd(fscal,dy12);
 477             tz               = _mm_mul_pd(fscal,dz12);
 478
 479             /* Update vectorial force */
 480             fix1             = _mm_add_pd(fix1,tx);
 481             fiy1             = _mm_add_pd(fiy1,ty);
 482             fiz1             = _mm_add_pd(fiz1,tz);
 483
 484             fjx2             = _mm_add_pd(fjx2,tx);
 485             fjy2             = _mm_add_pd(fjy2,ty);
 486             fjz2             = _mm_add_pd(fjz2,tz);
 487
 488             /**************************
 489              * CALCULATE INTERACTIONS *
 490              **************************/
 491
 492             /* REACTION-FIELD ELECTROSTATICS */
 493             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 494             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velecsum         = _mm_add_pd(velecsum,velec);
 498
 499             fscal            = felec;
 500
 501             /* Calculate temporary vectorial force */
 502             tx               = _mm_mul_pd(fscal,dx20);
 503             ty               = _mm_mul_pd(fscal,dy20);
 504             tz               = _mm_mul_pd(fscal,dz20);
 505
 506             /* Update vectorial force */
 507             fix2             = _mm_add_pd(fix2,tx);
 508             fiy2             = _mm_add_pd(fiy2,ty);
 509             fiz2             = _mm_add_pd(fiz2,tz);
 510
 511             fjx0             = _mm_add_pd(fjx0,tx);
 512             fjy0             = _mm_add_pd(fjy0,ty);
 513             fjz0             = _mm_add_pd(fjz0,tz);
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             /* REACTION-FIELD ELECTROSTATICS */
 520             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 521             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 522
 523             /* Update potential sum for this i atom from the interaction with this j atom. */
 524             velecsum         = _mm_add_pd(velecsum,velec);
 525
 526             fscal            = felec;
 527
 528             /* Calculate temporary vectorial force */
 529             tx               = _mm_mul_pd(fscal,dx21);
 530             ty               = _mm_mul_pd(fscal,dy21);
 531             tz               = _mm_mul_pd(fscal,dz21);
 532
 533             /* Update vectorial force */
 534             fix2             = _mm_add_pd(fix2,tx);
 535             fiy2             = _mm_add_pd(fiy2,ty);
 536             fiz2             = _mm_add_pd(fiz2,tz);
 537
 538             fjx1             = _mm_add_pd(fjx1,tx);
 539             fjy1             = _mm_add_pd(fjy1,ty);
 540             fjz1             = _mm_add_pd(fjz1,tz);
 541
 542             /**************************
 543              * CALCULATE INTERACTIONS *
 544              **************************/
 545
 546             /* REACTION-FIELD ELECTROSTATICS */
 547             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 548             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 549
 550             /* Update potential sum for this i atom from the interaction with this j atom. */
 551             velecsum         = _mm_add_pd(velecsum,velec);
 552
 553             fscal            = felec;
 554
 555             /* Calculate temporary vectorial force */
 556             tx               = _mm_mul_pd(fscal,dx22);
 557             ty               = _mm_mul_pd(fscal,dy22);
 558             tz               = _mm_mul_pd(fscal,dz22);
 559
 560             /* Update vectorial force */
 561             fix2             = _mm_add_pd(fix2,tx);
 562             fiy2             = _mm_add_pd(fiy2,ty);
 563             fiz2             = _mm_add_pd(fiz2,tz);
 564
 565             fjx2             = _mm_add_pd(fjx2,tx);
 566             fjy2             = _mm_add_pd(fjy2,ty);
 567             fjz2             = _mm_add_pd(fjz2,tz);
 568
 569             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 570
 571             /* Inner loop uses 323 flops */
 572         }
 573
 574         if(jidx<j_index_end)
 575         {
 576
 577             jnrA             = jjnr[jidx];
 578             j_coord_offsetA  = DIM*jnrA;
 579
 580             /* load j atom coordinates */
 581             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 582                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 583
 584             /* Calculate displacement vector */
 585             dx00             = _mm_sub_pd(ix0,jx0);
 586             dy00             = _mm_sub_pd(iy0,jy0);
 587             dz00             = _mm_sub_pd(iz0,jz0);
 588             dx01             = _mm_sub_pd(ix0,jx1);
 589             dy01             = _mm_sub_pd(iy0,jy1);
 590             dz01             = _mm_sub_pd(iz0,jz1);
 591             dx02             = _mm_sub_pd(ix0,jx2);
 592             dy02             = _mm_sub_pd(iy0,jy2);
 593             dz02             = _mm_sub_pd(iz0,jz2);
 594             dx10             = _mm_sub_pd(ix1,jx0);
 595             dy10             = _mm_sub_pd(iy1,jy0);
 596             dz10             = _mm_sub_pd(iz1,jz0);
 597             dx11             = _mm_sub_pd(ix1,jx1);
 598             dy11             = _mm_sub_pd(iy1,jy1);
 599             dz11             = _mm_sub_pd(iz1,jz1);
 600             dx12             = _mm_sub_pd(ix1,jx2);
 601             dy12             = _mm_sub_pd(iy1,jy2);
 602             dz12             = _mm_sub_pd(iz1,jz2);
 603             dx20             = _mm_sub_pd(ix2,jx0);
 604             dy20             = _mm_sub_pd(iy2,jy0);
 605             dz20             = _mm_sub_pd(iz2,jz0);
 606             dx21             = _mm_sub_pd(ix2,jx1);
 607             dy21             = _mm_sub_pd(iy2,jy1);
 608             dz21             = _mm_sub_pd(iz2,jz1);
 609             dx22             = _mm_sub_pd(ix2,jx2);
 610             dy22             = _mm_sub_pd(iy2,jy2);
 611             dz22             = _mm_sub_pd(iz2,jz2);
 612
 613             /* Calculate squared distance and things based on it */
 614             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 615             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 616             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 617             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 618             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 619             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 620             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 621             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 622             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 623
 624             rinv00           = sse41_invsqrt_d(rsq00);
 625             rinv01           = sse41_invsqrt_d(rsq01);
 626             rinv02           = sse41_invsqrt_d(rsq02);
 627             rinv10           = sse41_invsqrt_d(rsq10);
 628             rinv11           = sse41_invsqrt_d(rsq11);
 629             rinv12           = sse41_invsqrt_d(rsq12);
 630             rinv20           = sse41_invsqrt_d(rsq20);
 631             rinv21           = sse41_invsqrt_d(rsq21);
 632             rinv22           = sse41_invsqrt_d(rsq22);
 633
 634             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 635             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 636             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 637             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 638             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 639             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 640             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 641             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 642             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 643
 644             fjx0             = _mm_setzero_pd();
 645             fjy0             = _mm_setzero_pd();
 646             fjz0             = _mm_setzero_pd();
 647             fjx1             = _mm_setzero_pd();
 648             fjy1             = _mm_setzero_pd();
 649             fjz1             = _mm_setzero_pd();
 650             fjx2             = _mm_setzero_pd();
 651             fjy2             = _mm_setzero_pd();
 652             fjz2             = _mm_setzero_pd();
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             r00              = _mm_mul_pd(rsq00,rinv00);
 659
 660             /* Calculate table index by multiplying r with table scale and truncate to integer */
 661             rt               = _mm_mul_pd(r00,vftabscale);
 662             vfitab           = _mm_cvttpd_epi32(rt);
 663             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 664             vfitab           = _mm_slli_epi32(vfitab,3);
 665
 666             /* REACTION-FIELD ELECTROSTATICS */
 667             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 668             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 669
 670             /* CUBIC SPLINE TABLE DISPERSION */
 671             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 672             F                = _mm_setzero_pd();
 673             GMX_MM_TRANSPOSE2_PD(Y,F);
 674             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 675             H                = _mm_setzero_pd();
 676             GMX_MM_TRANSPOSE2_PD(G,H);
 677             Heps             = _mm_mul_pd(vfeps,H);
 678             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 679             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 680             vvdw6            = _mm_mul_pd(c6_00,VV);
 681             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 682             fvdw6            = _mm_mul_pd(c6_00,FF);
 683
 684             /* CUBIC SPLINE TABLE REPULSION */
 685             vfitab           = _mm_add_epi32(vfitab,ifour);
 686             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 687             F                = _mm_setzero_pd();
 688             GMX_MM_TRANSPOSE2_PD(Y,F);
 689             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 690             H                = _mm_setzero_pd();
 691             GMX_MM_TRANSPOSE2_PD(G,H);
 692             Heps             = _mm_mul_pd(vfeps,H);
 693             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 694             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 695             vvdw12           = _mm_mul_pd(c12_00,VV);
 696             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 697             fvdw12           = _mm_mul_pd(c12_00,FF);
 698             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 699             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 700
 701             /* Update potential sum for this i atom from the interaction with this j atom. */
 702             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 703             velecsum         = _mm_add_pd(velecsum,velec);
 704             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 705             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 706
 707             fscal            = _mm_add_pd(felec,fvdw);
 708
 709             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 710
 711             /* Calculate temporary vectorial force */
 712             tx               = _mm_mul_pd(fscal,dx00);
 713             ty               = _mm_mul_pd(fscal,dy00);
 714             tz               = _mm_mul_pd(fscal,dz00);
 715
 716             /* Update vectorial force */
 717             fix0             = _mm_add_pd(fix0,tx);
 718             fiy0             = _mm_add_pd(fiy0,ty);
 719             fiz0             = _mm_add_pd(fiz0,tz);
 720
 721             fjx0             = _mm_add_pd(fjx0,tx);
 722             fjy0             = _mm_add_pd(fjy0,ty);
 723             fjz0             = _mm_add_pd(fjz0,tz);
 724
 725             /**************************
 726              * CALCULATE INTERACTIONS *
 727              **************************/
 728
 729             /* REACTION-FIELD ELECTROSTATICS */
 730             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 731             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 732
 733             /* Update potential sum for this i atom from the interaction with this j atom. */
 734             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 735             velecsum         = _mm_add_pd(velecsum,velec);
 736
 737             fscal            = felec;
 738
 739             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 740
 741             /* Calculate temporary vectorial force */
 742             tx               = _mm_mul_pd(fscal,dx01);
 743             ty               = _mm_mul_pd(fscal,dy01);
 744             tz               = _mm_mul_pd(fscal,dz01);
 745
 746             /* Update vectorial force */
 747             fix0             = _mm_add_pd(fix0,tx);
 748             fiy0             = _mm_add_pd(fiy0,ty);
 749             fiz0             = _mm_add_pd(fiz0,tz);
 750
 751             fjx1             = _mm_add_pd(fjx1,tx);
 752             fjy1             = _mm_add_pd(fjy1,ty);
 753             fjz1             = _mm_add_pd(fjz1,tz);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             /* REACTION-FIELD ELECTROSTATICS */
 760             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 761             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 765             velecsum         = _mm_add_pd(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 770
 771             /* Calculate temporary vectorial force */
 772             tx               = _mm_mul_pd(fscal,dx02);
 773             ty               = _mm_mul_pd(fscal,dy02);
 774             tz               = _mm_mul_pd(fscal,dz02);
 775
 776             /* Update vectorial force */
 777             fix0             = _mm_add_pd(fix0,tx);
 778             fiy0             = _mm_add_pd(fiy0,ty);
 779             fiz0             = _mm_add_pd(fiz0,tz);
 780
 781             fjx2             = _mm_add_pd(fjx2,tx);
 782             fjy2             = _mm_add_pd(fjy2,ty);
 783             fjz2             = _mm_add_pd(fjz2,tz);
 784
 785             /**************************
 786              * CALCULATE INTERACTIONS *
 787              **************************/
 788
 789             /* REACTION-FIELD ELECTROSTATICS */
 790             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 791             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 792
 793             /* Update potential sum for this i atom from the interaction with this j atom. */
 794             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 795             velecsum         = _mm_add_pd(velecsum,velec);
 796
 797             fscal            = felec;
 798
 799             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 800
 801             /* Calculate temporary vectorial force */
 802             tx               = _mm_mul_pd(fscal,dx10);
 803             ty               = _mm_mul_pd(fscal,dy10);
 804             tz               = _mm_mul_pd(fscal,dz10);
 805
 806             /* Update vectorial force */
 807             fix1             = _mm_add_pd(fix1,tx);
 808             fiy1             = _mm_add_pd(fiy1,ty);
 809             fiz1             = _mm_add_pd(fiz1,tz);
 810
 811             fjx0             = _mm_add_pd(fjx0,tx);
 812             fjy0             = _mm_add_pd(fjy0,ty);
 813             fjz0             = _mm_add_pd(fjz0,tz);
 814
 815             /**************************
 816              * CALCULATE INTERACTIONS *
 817              **************************/
 818
 819             /* REACTION-FIELD ELECTROSTATICS */
 820             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 821             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 822
 823             /* Update potential sum for this i atom from the interaction with this j atom. */
 824             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 825             velecsum         = _mm_add_pd(velecsum,velec);
 826
 827             fscal            = felec;
 828
 829             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 830
 831             /* Calculate temporary vectorial force */
 832             tx               = _mm_mul_pd(fscal,dx11);
 833             ty               = _mm_mul_pd(fscal,dy11);
 834             tz               = _mm_mul_pd(fscal,dz11);
 835
 836             /* Update vectorial force */
 837             fix1             = _mm_add_pd(fix1,tx);
 838             fiy1             = _mm_add_pd(fiy1,ty);
 839             fiz1             = _mm_add_pd(fiz1,tz);
 840
 841             fjx1             = _mm_add_pd(fjx1,tx);
 842             fjy1             = _mm_add_pd(fjy1,ty);
 843             fjz1             = _mm_add_pd(fjz1,tz);
 844
 845             /**************************
 846              * CALCULATE INTERACTIONS *
 847              **************************/
 848
 849             /* REACTION-FIELD ELECTROSTATICS */
 850             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 851             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 852
 853             /* Update potential sum for this i atom from the interaction with this j atom. */
 854             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 855             velecsum         = _mm_add_pd(velecsum,velec);
 856
 857             fscal            = felec;
 858
 859             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 860
 861             /* Calculate temporary vectorial force */
 862             tx               = _mm_mul_pd(fscal,dx12);
 863             ty               = _mm_mul_pd(fscal,dy12);
 864             tz               = _mm_mul_pd(fscal,dz12);
 865
 866             /* Update vectorial force */
 867             fix1             = _mm_add_pd(fix1,tx);
 868             fiy1             = _mm_add_pd(fiy1,ty);
 869             fiz1             = _mm_add_pd(fiz1,tz);
 870
 871             fjx2             = _mm_add_pd(fjx2,tx);
 872             fjy2             = _mm_add_pd(fjy2,ty);
 873             fjz2             = _mm_add_pd(fjz2,tz);
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             /* REACTION-FIELD ELECTROSTATICS */
 880             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 881             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 885             velecsum         = _mm_add_pd(velecsum,velec);
 886
 887             fscal            = felec;
 888
 889             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 890
 891             /* Calculate temporary vectorial force */
 892             tx               = _mm_mul_pd(fscal,dx20);
 893             ty               = _mm_mul_pd(fscal,dy20);
 894             tz               = _mm_mul_pd(fscal,dz20);
 895
 896             /* Update vectorial force */
 897             fix2             = _mm_add_pd(fix2,tx);
 898             fiy2             = _mm_add_pd(fiy2,ty);
 899             fiz2             = _mm_add_pd(fiz2,tz);
 900
 901             fjx0             = _mm_add_pd(fjx0,tx);
 902             fjy0             = _mm_add_pd(fjy0,ty);
 903             fjz0             = _mm_add_pd(fjz0,tz);
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             /* REACTION-FIELD ELECTROSTATICS */
 910             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 911             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 912
 913             /* Update potential sum for this i atom from the interaction with this j atom. */
 914             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 915             velecsum         = _mm_add_pd(velecsum,velec);
 916
 917             fscal            = felec;
 918
 919             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 920
 921             /* Calculate temporary vectorial force */
 922             tx               = _mm_mul_pd(fscal,dx21);
 923             ty               = _mm_mul_pd(fscal,dy21);
 924             tz               = _mm_mul_pd(fscal,dz21);
 925
 926             /* Update vectorial force */
 927             fix2             = _mm_add_pd(fix2,tx);
 928             fiy2             = _mm_add_pd(fiy2,ty);
 929             fiz2             = _mm_add_pd(fiz2,tz);
 930
 931             fjx1             = _mm_add_pd(fjx1,tx);
 932             fjy1             = _mm_add_pd(fjy1,ty);
 933             fjz1             = _mm_add_pd(fjz1,tz);
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             /* REACTION-FIELD ELECTROSTATICS */
 940             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 941             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 942
 943             /* Update potential sum for this i atom from the interaction with this j atom. */
 944             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 945             velecsum         = _mm_add_pd(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 950
 951             /* Calculate temporary vectorial force */
 952             tx               = _mm_mul_pd(fscal,dx22);
 953             ty               = _mm_mul_pd(fscal,dy22);
 954             tz               = _mm_mul_pd(fscal,dz22);
 955
 956             /* Update vectorial force */
 957             fix2             = _mm_add_pd(fix2,tx);
 958             fiy2             = _mm_add_pd(fiy2,ty);
 959             fiz2             = _mm_add_pd(fiz2,tz);
 960
 961             fjx2             = _mm_add_pd(fjx2,tx);
 962             fjy2             = _mm_add_pd(fjy2,ty);
 963             fjz2             = _mm_add_pd(fjz2,tz);
 964
 965             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 966
 967             /* Inner loop uses 323 flops */
 968         }
 969
 970         /* End of innermost loop */
 971
 972         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 973                                               f+i_coord_offset,fshift+i_shift_offset);
 974
 975         ggid                        = gid[iidx];
 976         /* Update potential energies */
 977         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 978         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 979
 980         /* Increment number of inner iterations */
 981         inneriter                  += j_index_end - j_index_start;
 982
 983         /* Outer loop uses 20 flops */
 984     }
 985
 986     /* Increment number of outer iterations */
 987     outeriter        += nri;
 988
 989     /* Update outer/inner flops */
 990
 991     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*323);
 992 }
 993 /*
 994  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_sse4_1_double
 995  * Electrostatics interaction: ReactionField
 996  * VdW interaction:            CubicSplineTable
 997  * Geometry:                   Water3-Water3
 998  * Calculate force/pot:        Force
 999  */
1000 void
1001 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_sse4_1_double
1002                     (t_nblist                    * gmx_restrict       nlist,
1003                      rvec                        * gmx_restrict          xx,
1004                      rvec                        * gmx_restrict          ff,
1005                      struct t_forcerec           * gmx_restrict          fr,
1006                      t_mdatoms                   * gmx_restrict     mdatoms,
1007                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1008                      t_nrnb                      * gmx_restrict        nrnb)
1009 {
1010     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1011      * just 0 for non-waters.
1012      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1013      * jnr indices corresponding to data put in the four positions in the SIMD register.
1014      */
1015     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1016     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1017     int              jnrA,jnrB;
1018     int              j_coord_offsetA,j_coord_offsetB;
1019     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1020     real             rcutoff_scalar;
1021     real             *shiftvec,*fshift,*x,*f;
1022     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1023     int              vdwioffset0;
1024     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1025     int              vdwioffset1;
1026     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1027     int              vdwioffset2;
1028     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1029     int              vdwjidx0A,vdwjidx0B;
1030     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1031     int              vdwjidx1A,vdwjidx1B;
1032     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1033     int              vdwjidx2A,vdwjidx2B;
1034     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1035     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1036     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1037     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1038     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1039     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1040     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1041     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1042     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1043     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1044     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1045     real             *charge;
1046     int              nvdwtype;
1047     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1048     int              *vdwtype;
1049     real             *vdwparam;
1050     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1051     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1052     __m128i          vfitab;
1053     __m128i          ifour       = _mm_set1_epi32(4);
1054     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1055     real             *vftab;
1056     __m128d          dummy_mask,cutoff_mask;
1057     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1058     __m128d          one     = _mm_set1_pd(1.0);
1059     __m128d          two     = _mm_set1_pd(2.0);
1060     x                = xx[0];
1061     f                = ff[0];
1062
1063     nri              = nlist->nri;
1064     iinr             = nlist->iinr;
1065     jindex           = nlist->jindex;
1066     jjnr             = nlist->jjnr;
1067     shiftidx         = nlist->shift;
1068     gid              = nlist->gid;
1069     shiftvec         = fr->shift_vec[0];
1070     fshift           = fr->fshift[0];
1071     facel            = _mm_set1_pd(fr->ic->epsfac);
1072     charge           = mdatoms->chargeA;
1073     krf              = _mm_set1_pd(fr->ic->k_rf);
1074     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1075     crf              = _mm_set1_pd(fr->ic->c_rf);
1076     nvdwtype         = fr->ntype;
1077     vdwparam         = fr->nbfp;
1078     vdwtype          = mdatoms->typeA;
1079
1080     vftab            = kernel_data->table_vdw->data;
1081     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1082
1083     /* Setup water-specific parameters */
1084     inr              = nlist->iinr[0];
1085     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1086     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1087     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1088     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1089
1090     jq0              = _mm_set1_pd(charge[inr+0]);
1091     jq1              = _mm_set1_pd(charge[inr+1]);
1092     jq2              = _mm_set1_pd(charge[inr+2]);
1093     vdwjidx0A        = 2*vdwtype[inr+0];
1094     qq00             = _mm_mul_pd(iq0,jq0);
1095     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1096     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1097     qq01             = _mm_mul_pd(iq0,jq1);
1098     qq02             = _mm_mul_pd(iq0,jq2);
1099     qq10             = _mm_mul_pd(iq1,jq0);
1100     qq11             = _mm_mul_pd(iq1,jq1);
1101     qq12             = _mm_mul_pd(iq1,jq2);
1102     qq20             = _mm_mul_pd(iq2,jq0);
1103     qq21             = _mm_mul_pd(iq2,jq1);
1104     qq22             = _mm_mul_pd(iq2,jq2);
1105
1106     /* Avoid stupid compiler warnings */
1107     jnrA = jnrB = 0;
1108     j_coord_offsetA = 0;
1109     j_coord_offsetB = 0;
1110
1111     outeriter        = 0;
1112     inneriter        = 0;
1113
1114     /* Start outer loop over neighborlists */
1115     for(iidx=0; iidx<nri; iidx++)
1116     {
1117         /* Load shift vector for this list */
1118         i_shift_offset   = DIM*shiftidx[iidx];
1119
1120         /* Load limits for loop over neighbors */
1121         j_index_start    = jindex[iidx];
1122         j_index_end      = jindex[iidx+1];
1123
1124         /* Get outer coordinate index */
1125         inr              = iinr[iidx];
1126         i_coord_offset   = DIM*inr;
1127
1128         /* Load i particle coords and add shift vector */
1129         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1130                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1131
1132         fix0             = _mm_setzero_pd();
1133         fiy0             = _mm_setzero_pd();
1134         fiz0             = _mm_setzero_pd();
1135         fix1             = _mm_setzero_pd();
1136         fiy1             = _mm_setzero_pd();
1137         fiz1             = _mm_setzero_pd();
1138         fix2             = _mm_setzero_pd();
1139         fiy2             = _mm_setzero_pd();
1140         fiz2             = _mm_setzero_pd();
1141
1142         /* Start inner kernel loop */
1143         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1144         {
1145
1146             /* Get j neighbor index, and coordinate index */
1147             jnrA             = jjnr[jidx];
1148             jnrB             = jjnr[jidx+1];
1149             j_coord_offsetA  = DIM*jnrA;
1150             j_coord_offsetB  = DIM*jnrB;
1151
1152             /* load j atom coordinates */
1153             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1154                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1155
1156             /* Calculate displacement vector */
1157             dx00             = _mm_sub_pd(ix0,jx0);
1158             dy00             = _mm_sub_pd(iy0,jy0);
1159             dz00             = _mm_sub_pd(iz0,jz0);
1160             dx01             = _mm_sub_pd(ix0,jx1);
1161             dy01             = _mm_sub_pd(iy0,jy1);
1162             dz01             = _mm_sub_pd(iz0,jz1);
1163             dx02             = _mm_sub_pd(ix0,jx2);
1164             dy02             = _mm_sub_pd(iy0,jy2);
1165             dz02             = _mm_sub_pd(iz0,jz2);
1166             dx10             = _mm_sub_pd(ix1,jx0);
1167             dy10             = _mm_sub_pd(iy1,jy0);
1168             dz10             = _mm_sub_pd(iz1,jz0);
1169             dx11             = _mm_sub_pd(ix1,jx1);
1170             dy11             = _mm_sub_pd(iy1,jy1);
1171             dz11             = _mm_sub_pd(iz1,jz1);
1172             dx12             = _mm_sub_pd(ix1,jx2);
1173             dy12             = _mm_sub_pd(iy1,jy2);
1174             dz12             = _mm_sub_pd(iz1,jz2);
1175             dx20             = _mm_sub_pd(ix2,jx0);
1176             dy20             = _mm_sub_pd(iy2,jy0);
1177             dz20             = _mm_sub_pd(iz2,jz0);
1178             dx21             = _mm_sub_pd(ix2,jx1);
1179             dy21             = _mm_sub_pd(iy2,jy1);
1180             dz21             = _mm_sub_pd(iz2,jz1);
1181             dx22             = _mm_sub_pd(ix2,jx2);
1182             dy22             = _mm_sub_pd(iy2,jy2);
1183             dz22             = _mm_sub_pd(iz2,jz2);
1184
1185             /* Calculate squared distance and things based on it */
1186             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1187             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1188             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1189             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1190             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1191             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1192             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1193             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1194             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1195
1196             rinv00           = sse41_invsqrt_d(rsq00);
1197             rinv01           = sse41_invsqrt_d(rsq01);
1198             rinv02           = sse41_invsqrt_d(rsq02);
1199             rinv10           = sse41_invsqrt_d(rsq10);
1200             rinv11           = sse41_invsqrt_d(rsq11);
1201             rinv12           = sse41_invsqrt_d(rsq12);
1202             rinv20           = sse41_invsqrt_d(rsq20);
1203             rinv21           = sse41_invsqrt_d(rsq21);
1204             rinv22           = sse41_invsqrt_d(rsq22);
1205
1206             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1207             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1208             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1209             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1210             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1211             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1212             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1213             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1214             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1215
1216             fjx0             = _mm_setzero_pd();
1217             fjy0             = _mm_setzero_pd();
1218             fjz0             = _mm_setzero_pd();
1219             fjx1             = _mm_setzero_pd();
1220             fjy1             = _mm_setzero_pd();
1221             fjz1             = _mm_setzero_pd();
1222             fjx2             = _mm_setzero_pd();
1223             fjy2             = _mm_setzero_pd();
1224             fjz2             = _mm_setzero_pd();
1225
1226             /**************************
1227              * CALCULATE INTERACTIONS *
1228              **************************/
1229
1230             r00              = _mm_mul_pd(rsq00,rinv00);
1231
1232             /* Calculate table index by multiplying r with table scale and truncate to integer */
1233             rt               = _mm_mul_pd(r00,vftabscale);
1234             vfitab           = _mm_cvttpd_epi32(rt);
1235             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1236             vfitab           = _mm_slli_epi32(vfitab,3);
1237
1238             /* REACTION-FIELD ELECTROSTATICS */
1239             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1240
1241             /* CUBIC SPLINE TABLE DISPERSION */
1242             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1243             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1244             GMX_MM_TRANSPOSE2_PD(Y,F);
1245             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1246             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1247             GMX_MM_TRANSPOSE2_PD(G,H);
1248             Heps             = _mm_mul_pd(vfeps,H);
1249             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1250             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1251             fvdw6            = _mm_mul_pd(c6_00,FF);
1252
1253             /* CUBIC SPLINE TABLE REPULSION */
1254             vfitab           = _mm_add_epi32(vfitab,ifour);
1255             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1256             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1257             GMX_MM_TRANSPOSE2_PD(Y,F);
1258             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1259             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1260             GMX_MM_TRANSPOSE2_PD(G,H);
1261             Heps             = _mm_mul_pd(vfeps,H);
1262             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1263             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1264             fvdw12           = _mm_mul_pd(c12_00,FF);
1265             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1266
1267             fscal            = _mm_add_pd(felec,fvdw);
1268
1269             /* Calculate temporary vectorial force */
1270             tx               = _mm_mul_pd(fscal,dx00);
1271             ty               = _mm_mul_pd(fscal,dy00);
1272             tz               = _mm_mul_pd(fscal,dz00);
1273
1274             /* Update vectorial force */
1275             fix0             = _mm_add_pd(fix0,tx);
1276             fiy0             = _mm_add_pd(fiy0,ty);
1277             fiz0             = _mm_add_pd(fiz0,tz);
1278
1279             fjx0             = _mm_add_pd(fjx0,tx);
1280             fjy0             = _mm_add_pd(fjy0,ty);
1281             fjz0             = _mm_add_pd(fjz0,tz);
1282
1283             /**************************
1284              * CALCULATE INTERACTIONS *
1285              **************************/
1286
1287             /* REACTION-FIELD ELECTROSTATICS */
1288             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1289
1290             fscal            = felec;
1291
1292             /* Calculate temporary vectorial force */
1293             tx               = _mm_mul_pd(fscal,dx01);
1294             ty               = _mm_mul_pd(fscal,dy01);
1295             tz               = _mm_mul_pd(fscal,dz01);
1296
1297             /* Update vectorial force */
1298             fix0             = _mm_add_pd(fix0,tx);
1299             fiy0             = _mm_add_pd(fiy0,ty);
1300             fiz0             = _mm_add_pd(fiz0,tz);
1301
1302             fjx1             = _mm_add_pd(fjx1,tx);
1303             fjy1             = _mm_add_pd(fjy1,ty);
1304             fjz1             = _mm_add_pd(fjz1,tz);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* REACTION-FIELD ELECTROSTATICS */
1311             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1312
1313             fscal            = felec;
1314
1315             /* Calculate temporary vectorial force */
1316             tx               = _mm_mul_pd(fscal,dx02);
1317             ty               = _mm_mul_pd(fscal,dy02);
1318             tz               = _mm_mul_pd(fscal,dz02);
1319
1320             /* Update vectorial force */
1321             fix0             = _mm_add_pd(fix0,tx);
1322             fiy0             = _mm_add_pd(fiy0,ty);
1323             fiz0             = _mm_add_pd(fiz0,tz);
1324
1325             fjx2             = _mm_add_pd(fjx2,tx);
1326             fjy2             = _mm_add_pd(fjy2,ty);
1327             fjz2             = _mm_add_pd(fjz2,tz);
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             /* REACTION-FIELD ELECTROSTATICS */
1334             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1335
1336             fscal            = felec;
1337
1338             /* Calculate temporary vectorial force */
1339             tx               = _mm_mul_pd(fscal,dx10);
1340             ty               = _mm_mul_pd(fscal,dy10);
1341             tz               = _mm_mul_pd(fscal,dz10);
1342
1343             /* Update vectorial force */
1344             fix1             = _mm_add_pd(fix1,tx);
1345             fiy1             = _mm_add_pd(fiy1,ty);
1346             fiz1             = _mm_add_pd(fiz1,tz);
1347
1348             fjx0             = _mm_add_pd(fjx0,tx);
1349             fjy0             = _mm_add_pd(fjy0,ty);
1350             fjz0             = _mm_add_pd(fjz0,tz);
1351
1352             /**************************
1353              * CALCULATE INTERACTIONS *
1354              **************************/
1355
1356             /* REACTION-FIELD ELECTROSTATICS */
1357             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1358
1359             fscal            = felec;
1360
1361             /* Calculate temporary vectorial force */
1362             tx               = _mm_mul_pd(fscal,dx11);
1363             ty               = _mm_mul_pd(fscal,dy11);
1364             tz               = _mm_mul_pd(fscal,dz11);
1365
1366             /* Update vectorial force */
1367             fix1             = _mm_add_pd(fix1,tx);
1368             fiy1             = _mm_add_pd(fiy1,ty);
1369             fiz1             = _mm_add_pd(fiz1,tz);
1370
1371             fjx1             = _mm_add_pd(fjx1,tx);
1372             fjy1             = _mm_add_pd(fjy1,ty);
1373             fjz1             = _mm_add_pd(fjz1,tz);
1374
1375             /**************************
1376              * CALCULATE INTERACTIONS *
1377              **************************/
1378
1379             /* REACTION-FIELD ELECTROSTATICS */
1380             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1381
1382             fscal            = felec;
1383
1384             /* Calculate temporary vectorial force */
1385             tx               = _mm_mul_pd(fscal,dx12);
1386             ty               = _mm_mul_pd(fscal,dy12);
1387             tz               = _mm_mul_pd(fscal,dz12);
1388
1389             /* Update vectorial force */
1390             fix1             = _mm_add_pd(fix1,tx);
1391             fiy1             = _mm_add_pd(fiy1,ty);
1392             fiz1             = _mm_add_pd(fiz1,tz);
1393
1394             fjx2             = _mm_add_pd(fjx2,tx);
1395             fjy2             = _mm_add_pd(fjy2,ty);
1396             fjz2             = _mm_add_pd(fjz2,tz);
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             /* REACTION-FIELD ELECTROSTATICS */
1403             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1404
1405             fscal            = felec;
1406
1407             /* Calculate temporary vectorial force */
1408             tx               = _mm_mul_pd(fscal,dx20);
1409             ty               = _mm_mul_pd(fscal,dy20);
1410             tz               = _mm_mul_pd(fscal,dz20);
1411
1412             /* Update vectorial force */
1413             fix2             = _mm_add_pd(fix2,tx);
1414             fiy2             = _mm_add_pd(fiy2,ty);
1415             fiz2             = _mm_add_pd(fiz2,tz);
1416
1417             fjx0             = _mm_add_pd(fjx0,tx);
1418             fjy0             = _mm_add_pd(fjy0,ty);
1419             fjz0             = _mm_add_pd(fjz0,tz);
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             /* REACTION-FIELD ELECTROSTATICS */
1426             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1427
1428             fscal            = felec;
1429
1430             /* Calculate temporary vectorial force */
1431             tx               = _mm_mul_pd(fscal,dx21);
1432             ty               = _mm_mul_pd(fscal,dy21);
1433             tz               = _mm_mul_pd(fscal,dz21);
1434
1435             /* Update vectorial force */
1436             fix2             = _mm_add_pd(fix2,tx);
1437             fiy2             = _mm_add_pd(fiy2,ty);
1438             fiz2             = _mm_add_pd(fiz2,tz);
1439
1440             fjx1             = _mm_add_pd(fjx1,tx);
1441             fjy1             = _mm_add_pd(fjy1,ty);
1442             fjz1             = _mm_add_pd(fjz1,tz);
1443
1444             /**************************
1445              * CALCULATE INTERACTIONS *
1446              **************************/
1447
1448             /* REACTION-FIELD ELECTROSTATICS */
1449             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1450
1451             fscal            = felec;
1452
1453             /* Calculate temporary vectorial force */
1454             tx               = _mm_mul_pd(fscal,dx22);
1455             ty               = _mm_mul_pd(fscal,dy22);
1456             tz               = _mm_mul_pd(fscal,dz22);
1457
1458             /* Update vectorial force */
1459             fix2             = _mm_add_pd(fix2,tx);
1460             fiy2             = _mm_add_pd(fiy2,ty);
1461             fiz2             = _mm_add_pd(fiz2,tz);
1462
1463             fjx2             = _mm_add_pd(fjx2,tx);
1464             fjy2             = _mm_add_pd(fjy2,ty);
1465             fjz2             = _mm_add_pd(fjz2,tz);
1466
1467             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1468
1469             /* Inner loop uses 270 flops */
1470         }
1471
1472         if(jidx<j_index_end)
1473         {
1474
1475             jnrA             = jjnr[jidx];
1476             j_coord_offsetA  = DIM*jnrA;
1477
1478             /* load j atom coordinates */
1479             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1480                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1481
1482             /* Calculate displacement vector */
1483             dx00             = _mm_sub_pd(ix0,jx0);
1484             dy00             = _mm_sub_pd(iy0,jy0);
1485             dz00             = _mm_sub_pd(iz0,jz0);
1486             dx01             = _mm_sub_pd(ix0,jx1);
1487             dy01             = _mm_sub_pd(iy0,jy1);
1488             dz01             = _mm_sub_pd(iz0,jz1);
1489             dx02             = _mm_sub_pd(ix0,jx2);
1490             dy02             = _mm_sub_pd(iy0,jy2);
1491             dz02             = _mm_sub_pd(iz0,jz2);
1492             dx10             = _mm_sub_pd(ix1,jx0);
1493             dy10             = _mm_sub_pd(iy1,jy0);
1494             dz10             = _mm_sub_pd(iz1,jz0);
1495             dx11             = _mm_sub_pd(ix1,jx1);
1496             dy11             = _mm_sub_pd(iy1,jy1);
1497             dz11             = _mm_sub_pd(iz1,jz1);
1498             dx12             = _mm_sub_pd(ix1,jx2);
1499             dy12             = _mm_sub_pd(iy1,jy2);
1500             dz12             = _mm_sub_pd(iz1,jz2);
1501             dx20             = _mm_sub_pd(ix2,jx0);
1502             dy20             = _mm_sub_pd(iy2,jy0);
1503             dz20             = _mm_sub_pd(iz2,jz0);
1504             dx21             = _mm_sub_pd(ix2,jx1);
1505             dy21             = _mm_sub_pd(iy2,jy1);
1506             dz21             = _mm_sub_pd(iz2,jz1);
1507             dx22             = _mm_sub_pd(ix2,jx2);
1508             dy22             = _mm_sub_pd(iy2,jy2);
1509             dz22             = _mm_sub_pd(iz2,jz2);
1510
1511             /* Calculate squared distance and things based on it */
1512             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1513             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1514             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1515             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1516             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1517             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1518             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1519             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1520             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1521
1522             rinv00           = sse41_invsqrt_d(rsq00);
1523             rinv01           = sse41_invsqrt_d(rsq01);
1524             rinv02           = sse41_invsqrt_d(rsq02);
1525             rinv10           = sse41_invsqrt_d(rsq10);
1526             rinv11           = sse41_invsqrt_d(rsq11);
1527             rinv12           = sse41_invsqrt_d(rsq12);
1528             rinv20           = sse41_invsqrt_d(rsq20);
1529             rinv21           = sse41_invsqrt_d(rsq21);
1530             rinv22           = sse41_invsqrt_d(rsq22);
1531
1532             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1533             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1534             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1535             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1536             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1537             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1538             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1539             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1540             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1541
1542             fjx0             = _mm_setzero_pd();
1543             fjy0             = _mm_setzero_pd();
1544             fjz0             = _mm_setzero_pd();
1545             fjx1             = _mm_setzero_pd();
1546             fjy1             = _mm_setzero_pd();
1547             fjz1             = _mm_setzero_pd();
1548             fjx2             = _mm_setzero_pd();
1549             fjy2             = _mm_setzero_pd();
1550             fjz2             = _mm_setzero_pd();
1551
1552             /**************************
1553              * CALCULATE INTERACTIONS *
1554              **************************/
1555
1556             r00              = _mm_mul_pd(rsq00,rinv00);
1557
1558             /* Calculate table index by multiplying r with table scale and truncate to integer */
1559             rt               = _mm_mul_pd(r00,vftabscale);
1560             vfitab           = _mm_cvttpd_epi32(rt);
1561             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1562             vfitab           = _mm_slli_epi32(vfitab,3);
1563
1564             /* REACTION-FIELD ELECTROSTATICS */
1565             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1566
1567             /* CUBIC SPLINE TABLE DISPERSION */
1568             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1569             F                = _mm_setzero_pd();
1570             GMX_MM_TRANSPOSE2_PD(Y,F);
1571             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1572             H                = _mm_setzero_pd();
1573             GMX_MM_TRANSPOSE2_PD(G,H);
1574             Heps             = _mm_mul_pd(vfeps,H);
1575             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1576             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1577             fvdw6            = _mm_mul_pd(c6_00,FF);
1578
1579             /* CUBIC SPLINE TABLE REPULSION */
1580             vfitab           = _mm_add_epi32(vfitab,ifour);
1581             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1582             F                = _mm_setzero_pd();
1583             GMX_MM_TRANSPOSE2_PD(Y,F);
1584             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1585             H                = _mm_setzero_pd();
1586             GMX_MM_TRANSPOSE2_PD(G,H);
1587             Heps             = _mm_mul_pd(vfeps,H);
1588             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1589             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1590             fvdw12           = _mm_mul_pd(c12_00,FF);
1591             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1592
1593             fscal            = _mm_add_pd(felec,fvdw);
1594
1595             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1596
1597             /* Calculate temporary vectorial force */
1598             tx               = _mm_mul_pd(fscal,dx00);
1599             ty               = _mm_mul_pd(fscal,dy00);
1600             tz               = _mm_mul_pd(fscal,dz00);
1601
1602             /* Update vectorial force */
1603             fix0             = _mm_add_pd(fix0,tx);
1604             fiy0             = _mm_add_pd(fiy0,ty);
1605             fiz0             = _mm_add_pd(fiz0,tz);
1606
1607             fjx0             = _mm_add_pd(fjx0,tx);
1608             fjy0             = _mm_add_pd(fjy0,ty);
1609             fjz0             = _mm_add_pd(fjz0,tz);
1610
1611             /**************************
1612              * CALCULATE INTERACTIONS *
1613              **************************/
1614
1615             /* REACTION-FIELD ELECTROSTATICS */
1616             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1617
1618             fscal            = felec;
1619
1620             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1621
1622             /* Calculate temporary vectorial force */
1623             tx               = _mm_mul_pd(fscal,dx01);
1624             ty               = _mm_mul_pd(fscal,dy01);
1625             tz               = _mm_mul_pd(fscal,dz01);
1626
1627             /* Update vectorial force */
1628             fix0             = _mm_add_pd(fix0,tx);
1629             fiy0             = _mm_add_pd(fiy0,ty);
1630             fiz0             = _mm_add_pd(fiz0,tz);
1631
1632             fjx1             = _mm_add_pd(fjx1,tx);
1633             fjy1             = _mm_add_pd(fjy1,ty);
1634             fjz1             = _mm_add_pd(fjz1,tz);
1635
1636             /**************************
1637              * CALCULATE INTERACTIONS *
1638              **************************/
1639
1640             /* REACTION-FIELD ELECTROSTATICS */
1641             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1642
1643             fscal            = felec;
1644
1645             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1646
1647             /* Calculate temporary vectorial force */
1648             tx               = _mm_mul_pd(fscal,dx02);
1649             ty               = _mm_mul_pd(fscal,dy02);
1650             tz               = _mm_mul_pd(fscal,dz02);
1651
1652             /* Update vectorial force */
1653             fix0             = _mm_add_pd(fix0,tx);
1654             fiy0             = _mm_add_pd(fiy0,ty);
1655             fiz0             = _mm_add_pd(fiz0,tz);
1656
1657             fjx2             = _mm_add_pd(fjx2,tx);
1658             fjy2             = _mm_add_pd(fjy2,ty);
1659             fjz2             = _mm_add_pd(fjz2,tz);
1660
1661             /**************************
1662              * CALCULATE INTERACTIONS *
1663              **************************/
1664
1665             /* REACTION-FIELD ELECTROSTATICS */
1666             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1667
1668             fscal            = felec;
1669
1670             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1671
1672             /* Calculate temporary vectorial force */
1673             tx               = _mm_mul_pd(fscal,dx10);
1674             ty               = _mm_mul_pd(fscal,dy10);
1675             tz               = _mm_mul_pd(fscal,dz10);
1676
1677             /* Update vectorial force */
1678             fix1             = _mm_add_pd(fix1,tx);
1679             fiy1             = _mm_add_pd(fiy1,ty);
1680             fiz1             = _mm_add_pd(fiz1,tz);
1681
1682             fjx0             = _mm_add_pd(fjx0,tx);
1683             fjy0             = _mm_add_pd(fjy0,ty);
1684             fjz0             = _mm_add_pd(fjz0,tz);
1685
1686             /**************************
1687              * CALCULATE INTERACTIONS *
1688              **************************/
1689
1690             /* REACTION-FIELD ELECTROSTATICS */
1691             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1692
1693             fscal            = felec;
1694
1695             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1696
1697             /* Calculate temporary vectorial force */
1698             tx               = _mm_mul_pd(fscal,dx11);
1699             ty               = _mm_mul_pd(fscal,dy11);
1700             tz               = _mm_mul_pd(fscal,dz11);
1701
1702             /* Update vectorial force */
1703             fix1             = _mm_add_pd(fix1,tx);
1704             fiy1             = _mm_add_pd(fiy1,ty);
1705             fiz1             = _mm_add_pd(fiz1,tz);
1706
1707             fjx1             = _mm_add_pd(fjx1,tx);
1708             fjy1             = _mm_add_pd(fjy1,ty);
1709             fjz1             = _mm_add_pd(fjz1,tz);
1710
1711             /**************************
1712              * CALCULATE INTERACTIONS *
1713              **************************/
1714
1715             /* REACTION-FIELD ELECTROSTATICS */
1716             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1717
1718             fscal            = felec;
1719
1720             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1721
1722             /* Calculate temporary vectorial force */
1723             tx               = _mm_mul_pd(fscal,dx12);
1724             ty               = _mm_mul_pd(fscal,dy12);
1725             tz               = _mm_mul_pd(fscal,dz12);
1726
1727             /* Update vectorial force */
1728             fix1             = _mm_add_pd(fix1,tx);
1729             fiy1             = _mm_add_pd(fiy1,ty);
1730             fiz1             = _mm_add_pd(fiz1,tz);
1731
1732             fjx2             = _mm_add_pd(fjx2,tx);
1733             fjy2             = _mm_add_pd(fjy2,ty);
1734             fjz2             = _mm_add_pd(fjz2,tz);
1735
1736             /**************************
1737              * CALCULATE INTERACTIONS *
1738              **************************/
1739
1740             /* REACTION-FIELD ELECTROSTATICS */
1741             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1742
1743             fscal            = felec;
1744
1745             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1746
1747             /* Calculate temporary vectorial force */
1748             tx               = _mm_mul_pd(fscal,dx20);
1749             ty               = _mm_mul_pd(fscal,dy20);
1750             tz               = _mm_mul_pd(fscal,dz20);
1751
1752             /* Update vectorial force */
1753             fix2             = _mm_add_pd(fix2,tx);
1754             fiy2             = _mm_add_pd(fiy2,ty);
1755             fiz2             = _mm_add_pd(fiz2,tz);
1756
1757             fjx0             = _mm_add_pd(fjx0,tx);
1758             fjy0             = _mm_add_pd(fjy0,ty);
1759             fjz0             = _mm_add_pd(fjz0,tz);
1760
1761             /**************************
1762              * CALCULATE INTERACTIONS *
1763              **************************/
1764
1765             /* REACTION-FIELD ELECTROSTATICS */
1766             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1767
1768             fscal            = felec;
1769
1770             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1771
1772             /* Calculate temporary vectorial force */
1773             tx               = _mm_mul_pd(fscal,dx21);
1774             ty               = _mm_mul_pd(fscal,dy21);
1775             tz               = _mm_mul_pd(fscal,dz21);
1776
1777             /* Update vectorial force */
1778             fix2             = _mm_add_pd(fix2,tx);
1779             fiy2             = _mm_add_pd(fiy2,ty);
1780             fiz2             = _mm_add_pd(fiz2,tz);
1781
1782             fjx1             = _mm_add_pd(fjx1,tx);
1783             fjy1             = _mm_add_pd(fjy1,ty);
1784             fjz1             = _mm_add_pd(fjz1,tz);
1785
1786             /**************************
1787              * CALCULATE INTERACTIONS *
1788              **************************/
1789
1790             /* REACTION-FIELD ELECTROSTATICS */
1791             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1792
1793             fscal            = felec;
1794
1795             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1796
1797             /* Calculate temporary vectorial force */
1798             tx               = _mm_mul_pd(fscal,dx22);
1799             ty               = _mm_mul_pd(fscal,dy22);
1800             tz               = _mm_mul_pd(fscal,dz22);
1801
1802             /* Update vectorial force */
1803             fix2             = _mm_add_pd(fix2,tx);
1804             fiy2             = _mm_add_pd(fiy2,ty);
1805             fiz2             = _mm_add_pd(fiz2,tz);
1806
1807             fjx2             = _mm_add_pd(fjx2,tx);
1808             fjy2             = _mm_add_pd(fjy2,ty);
1809             fjz2             = _mm_add_pd(fjz2,tz);
1810
1811             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1812
1813             /* Inner loop uses 270 flops */
1814         }
1815
1816         /* End of innermost loop */
1817
1818         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1819                                               f+i_coord_offset,fshift+i_shift_offset);
1820
1821         /* Increment number of inner iterations */
1822         inneriter                  += j_index_end - j_index_start;
1823
1824         /* Outer loop uses 18 flops */
1825     }
1826
1827     /* Increment number of outer iterations */
1828     outeriter        += nri;
1829
1830     /* Update outer/inner flops */
1831
1832     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*270);
1833 }