src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_sse2_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 sse2_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse2_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_sse2_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_sse2_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     __m128d          dummy_mask,cutoff_mask;
 109     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 110     __m128d          one     = _mm_set1_pd(1.0);
 111     __m128d          two     = _mm_set1_pd(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_pd(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     krf              = _mm_set1_pd(fr->ic->k_rf);
 126     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 127     crf              = _mm_set1_pd(fr->ic->c_rf);
 128     nvdwtype         = fr->ntype;
 129     vdwparam         = fr->nbfp;
 130     vdwtype          = mdatoms->typeA;
 131
 132     /* Setup water-specific parameters */
 133     inr              = nlist->iinr[0];
 134     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 135     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 136     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 137     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 138
 139     jq0              = _mm_set1_pd(charge[inr+0]);
 140     jq1              = _mm_set1_pd(charge[inr+1]);
 141     jq2              = _mm_set1_pd(charge[inr+2]);
 142     vdwjidx0A        = 2*vdwtype[inr+0];
 143     qq00             = _mm_mul_pd(iq0,jq0);
 144     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 145     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 146     qq01             = _mm_mul_pd(iq0,jq1);
 147     qq02             = _mm_mul_pd(iq0,jq2);
 148     qq10             = _mm_mul_pd(iq1,jq0);
 149     qq11             = _mm_mul_pd(iq1,jq1);
 150     qq12             = _mm_mul_pd(iq1,jq2);
 151     qq20             = _mm_mul_pd(iq2,jq0);
 152     qq21             = _mm_mul_pd(iq2,jq1);
 153     qq22             = _mm_mul_pd(iq2,jq2);
 154
 155     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 156     rcutoff_scalar   = fr->ic->rcoulomb;
 157     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 158     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 159
 160     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 161     rvdw             = _mm_set1_pd(fr->ic->rvdw);
 162
 163     /* Avoid stupid compiler warnings */
 164     jnrA = jnrB = 0;
 165     j_coord_offsetA = 0;
 166     j_coord_offsetB = 0;
 167
 168     outeriter        = 0;
 169     inneriter        = 0;
 170
 171     /* Start outer loop over neighborlists */
 172     for(iidx=0; iidx<nri; iidx++)
 173     {
 174         /* Load shift vector for this list */
 175         i_shift_offset   = DIM*shiftidx[iidx];
 176
 177         /* Load limits for loop over neighbors */
 178         j_index_start    = jindex[iidx];
 179         j_index_end      = jindex[iidx+1];
 180
 181         /* Get outer coordinate index */
 182         inr              = iinr[iidx];
 183         i_coord_offset   = DIM*inr;
 184
 185         /* Load i particle coords and add shift vector */
 186         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 187                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 188
 189         fix0             = _mm_setzero_pd();
 190         fiy0             = _mm_setzero_pd();
 191         fiz0             = _mm_setzero_pd();
 192         fix1             = _mm_setzero_pd();
 193         fiy1             = _mm_setzero_pd();
 194         fiz1             = _mm_setzero_pd();
 195         fix2             = _mm_setzero_pd();
 196         fiy2             = _mm_setzero_pd();
 197         fiz2             = _mm_setzero_pd();
 198
 199         /* Reset potential sums */
 200         velecsum         = _mm_setzero_pd();
 201         vvdwsum          = _mm_setzero_pd();
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 205         {
 206
 207             /* Get j neighbor index, and coordinate index */
 208             jnrA             = jjnr[jidx];
 209             jnrB             = jjnr[jidx+1];
 210             j_coord_offsetA  = DIM*jnrA;
 211             j_coord_offsetB  = DIM*jnrB;
 212
 213             /* load j atom coordinates */
 214             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 215                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 216
 217             /* Calculate displacement vector */
 218             dx00             = _mm_sub_pd(ix0,jx0);
 219             dy00             = _mm_sub_pd(iy0,jy0);
 220             dz00             = _mm_sub_pd(iz0,jz0);
 221             dx01             = _mm_sub_pd(ix0,jx1);
 222             dy01             = _mm_sub_pd(iy0,jy1);
 223             dz01             = _mm_sub_pd(iz0,jz1);
 224             dx02             = _mm_sub_pd(ix0,jx2);
 225             dy02             = _mm_sub_pd(iy0,jy2);
 226             dz02             = _mm_sub_pd(iz0,jz2);
 227             dx10             = _mm_sub_pd(ix1,jx0);
 228             dy10             = _mm_sub_pd(iy1,jy0);
 229             dz10             = _mm_sub_pd(iz1,jz0);
 230             dx11             = _mm_sub_pd(ix1,jx1);
 231             dy11             = _mm_sub_pd(iy1,jy1);
 232             dz11             = _mm_sub_pd(iz1,jz1);
 233             dx12             = _mm_sub_pd(ix1,jx2);
 234             dy12             = _mm_sub_pd(iy1,jy2);
 235             dz12             = _mm_sub_pd(iz1,jz2);
 236             dx20             = _mm_sub_pd(ix2,jx0);
 237             dy20             = _mm_sub_pd(iy2,jy0);
 238             dz20             = _mm_sub_pd(iz2,jz0);
 239             dx21             = _mm_sub_pd(ix2,jx1);
 240             dy21             = _mm_sub_pd(iy2,jy1);
 241             dz21             = _mm_sub_pd(iz2,jz1);
 242             dx22             = _mm_sub_pd(ix2,jx2);
 243             dy22             = _mm_sub_pd(iy2,jy2);
 244             dz22             = _mm_sub_pd(iz2,jz2);
 245
 246             /* Calculate squared distance and things based on it */
 247             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 248             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 249             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 250             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 251             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 252             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 253             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 254             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 255             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 256
 257             rinv00           = sse2_invsqrt_d(rsq00);
 258             rinv01           = sse2_invsqrt_d(rsq01);
 259             rinv02           = sse2_invsqrt_d(rsq02);
 260             rinv10           = sse2_invsqrt_d(rsq10);
 261             rinv11           = sse2_invsqrt_d(rsq11);
 262             rinv12           = sse2_invsqrt_d(rsq12);
 263             rinv20           = sse2_invsqrt_d(rsq20);
 264             rinv21           = sse2_invsqrt_d(rsq21);
 265             rinv22           = sse2_invsqrt_d(rsq22);
 266
 267             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 268             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 269             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 270             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 271             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 272             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 273             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 274             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 275             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 276
 277             fjx0             = _mm_setzero_pd();
 278             fjy0             = _mm_setzero_pd();
 279             fjz0             = _mm_setzero_pd();
 280             fjx1             = _mm_setzero_pd();
 281             fjy1             = _mm_setzero_pd();
 282             fjz1             = _mm_setzero_pd();
 283             fjx2             = _mm_setzero_pd();
 284             fjy2             = _mm_setzero_pd();
 285             fjz2             = _mm_setzero_pd();
 286
 287             /**************************
 288              * CALCULATE INTERACTIONS *
 289              **************************/
 290
 291             if (gmx_mm_any_lt(rsq00,rcutoff2))
 292             {
 293
 294             /* REACTION-FIELD ELECTROSTATICS */
 295             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 296             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 297
 298             /* LENNARD-JONES DISPERSION/REPULSION */
 299
 300             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 301             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 302             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 303             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 304                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 305             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 306
 307             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 308
 309             /* Update potential sum for this i atom from the interaction with this j atom. */
 310             velec            = _mm_and_pd(velec,cutoff_mask);
 311             velecsum         = _mm_add_pd(velecsum,velec);
 312             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 313             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 314
 315             fscal            = _mm_add_pd(felec,fvdw);
 316
 317             fscal            = _mm_and_pd(fscal,cutoff_mask);
 318
 319             /* Calculate temporary vectorial force */
 320             tx               = _mm_mul_pd(fscal,dx00);
 321             ty               = _mm_mul_pd(fscal,dy00);
 322             tz               = _mm_mul_pd(fscal,dz00);
 323
 324             /* Update vectorial force */
 325             fix0             = _mm_add_pd(fix0,tx);
 326             fiy0             = _mm_add_pd(fiy0,ty);
 327             fiz0             = _mm_add_pd(fiz0,tz);
 328
 329             fjx0             = _mm_add_pd(fjx0,tx);
 330             fjy0             = _mm_add_pd(fjy0,ty);
 331             fjz0             = _mm_add_pd(fjz0,tz);
 332
 333             }
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             if (gmx_mm_any_lt(rsq01,rcutoff2))
 340             {
 341
 342             /* REACTION-FIELD ELECTROSTATICS */
 343             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 344             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 345
 346             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             velec            = _mm_and_pd(velec,cutoff_mask);
 350             velecsum         = _mm_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             fscal            = _mm_and_pd(fscal,cutoff_mask);
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm_mul_pd(fscal,dx01);
 358             ty               = _mm_mul_pd(fscal,dy01);
 359             tz               = _mm_mul_pd(fscal,dz01);
 360
 361             /* Update vectorial force */
 362             fix0             = _mm_add_pd(fix0,tx);
 363             fiy0             = _mm_add_pd(fiy0,ty);
 364             fiz0             = _mm_add_pd(fiz0,tz);
 365
 366             fjx1             = _mm_add_pd(fjx1,tx);
 367             fjy1             = _mm_add_pd(fjy1,ty);
 368             fjz1             = _mm_add_pd(fjz1,tz);
 369
 370             }
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             if (gmx_mm_any_lt(rsq02,rcutoff2))
 377             {
 378
 379             /* REACTION-FIELD ELECTROSTATICS */
 380             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 381             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 382
 383             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velec            = _mm_and_pd(velec,cutoff_mask);
 387             velecsum         = _mm_add_pd(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391             fscal            = _mm_and_pd(fscal,cutoff_mask);
 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
 409             /**************************
 410              * CALCULATE INTERACTIONS *
 411              **************************/
 412
 413             if (gmx_mm_any_lt(rsq10,rcutoff2))
 414             {
 415
 416             /* REACTION-FIELD ELECTROSTATICS */
 417             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 418             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 419
 420             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 421
 422             /* Update potential sum for this i atom from the interaction with this j atom. */
 423             velec            = _mm_and_pd(velec,cutoff_mask);
 424             velecsum         = _mm_add_pd(velecsum,velec);
 425
 426             fscal            = felec;
 427
 428             fscal            = _mm_and_pd(fscal,cutoff_mask);
 429
 430             /* Calculate temporary vectorial force */
 431             tx               = _mm_mul_pd(fscal,dx10);
 432             ty               = _mm_mul_pd(fscal,dy10);
 433             tz               = _mm_mul_pd(fscal,dz10);
 434
 435             /* Update vectorial force */
 436             fix1             = _mm_add_pd(fix1,tx);
 437             fiy1             = _mm_add_pd(fiy1,ty);
 438             fiz1             = _mm_add_pd(fiz1,tz);
 439
 440             fjx0             = _mm_add_pd(fjx0,tx);
 441             fjy0             = _mm_add_pd(fjy0,ty);
 442             fjz0             = _mm_add_pd(fjz0,tz);
 443
 444             }
 445
 446             /**************************
 447              * CALCULATE INTERACTIONS *
 448              **************************/
 449
 450             if (gmx_mm_any_lt(rsq11,rcutoff2))
 451             {
 452
 453             /* REACTION-FIELD ELECTROSTATICS */
 454             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 455             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 456
 457             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             velec            = _mm_and_pd(velec,cutoff_mask);
 461             velecsum         = _mm_add_pd(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             fscal            = _mm_and_pd(fscal,cutoff_mask);
 466
 467             /* Calculate temporary vectorial force */
 468             tx               = _mm_mul_pd(fscal,dx11);
 469             ty               = _mm_mul_pd(fscal,dy11);
 470             tz               = _mm_mul_pd(fscal,dz11);
 471
 472             /* Update vectorial force */
 473             fix1             = _mm_add_pd(fix1,tx);
 474             fiy1             = _mm_add_pd(fiy1,ty);
 475             fiz1             = _mm_add_pd(fiz1,tz);
 476
 477             fjx1             = _mm_add_pd(fjx1,tx);
 478             fjy1             = _mm_add_pd(fjy1,ty);
 479             fjz1             = _mm_add_pd(fjz1,tz);
 480
 481             }
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             if (gmx_mm_any_lt(rsq12,rcutoff2))
 488             {
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 492             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 493
 494             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velec            = _mm_and_pd(velec,cutoff_mask);
 498             velecsum         = _mm_add_pd(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502             fscal            = _mm_and_pd(fscal,cutoff_mask);
 503
 504             /* Calculate temporary vectorial force */
 505             tx               = _mm_mul_pd(fscal,dx12);
 506             ty               = _mm_mul_pd(fscal,dy12);
 507             tz               = _mm_mul_pd(fscal,dz12);
 508
 509             /* Update vectorial force */
 510             fix1             = _mm_add_pd(fix1,tx);
 511             fiy1             = _mm_add_pd(fiy1,ty);
 512             fiz1             = _mm_add_pd(fiz1,tz);
 513
 514             fjx2             = _mm_add_pd(fjx2,tx);
 515             fjy2             = _mm_add_pd(fjy2,ty);
 516             fjz2             = _mm_add_pd(fjz2,tz);
 517
 518             }
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             if (gmx_mm_any_lt(rsq20,rcutoff2))
 525             {
 526
 527             /* REACTION-FIELD ELECTROSTATICS */
 528             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 529             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 530
 531             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 532
 533             /* Update potential sum for this i atom from the interaction with this j atom. */
 534             velec            = _mm_and_pd(velec,cutoff_mask);
 535             velecsum         = _mm_add_pd(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539             fscal            = _mm_and_pd(fscal,cutoff_mask);
 540
 541             /* Calculate temporary vectorial force */
 542             tx               = _mm_mul_pd(fscal,dx20);
 543             ty               = _mm_mul_pd(fscal,dy20);
 544             tz               = _mm_mul_pd(fscal,dz20);
 545
 546             /* Update vectorial force */
 547             fix2             = _mm_add_pd(fix2,tx);
 548             fiy2             = _mm_add_pd(fiy2,ty);
 549             fiz2             = _mm_add_pd(fiz2,tz);
 550
 551             fjx0             = _mm_add_pd(fjx0,tx);
 552             fjy0             = _mm_add_pd(fjy0,ty);
 553             fjz0             = _mm_add_pd(fjz0,tz);
 554
 555             }
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             if (gmx_mm_any_lt(rsq21,rcutoff2))
 562             {
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 566             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 567
 568             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _mm_and_pd(velec,cutoff_mask);
 572             velecsum         = _mm_add_pd(velecsum,velec);
 573
 574             fscal            = felec;
 575
 576             fscal            = _mm_and_pd(fscal,cutoff_mask);
 577
 578             /* Calculate temporary vectorial force */
 579             tx               = _mm_mul_pd(fscal,dx21);
 580             ty               = _mm_mul_pd(fscal,dy21);
 581             tz               = _mm_mul_pd(fscal,dz21);
 582
 583             /* Update vectorial force */
 584             fix2             = _mm_add_pd(fix2,tx);
 585             fiy2             = _mm_add_pd(fiy2,ty);
 586             fiz2             = _mm_add_pd(fiz2,tz);
 587
 588             fjx1             = _mm_add_pd(fjx1,tx);
 589             fjy1             = _mm_add_pd(fjy1,ty);
 590             fjz1             = _mm_add_pd(fjz1,tz);
 591
 592             }
 593
 594             /**************************
 595              * CALCULATE INTERACTIONS *
 596              **************************/
 597
 598             if (gmx_mm_any_lt(rsq22,rcutoff2))
 599             {
 600
 601             /* REACTION-FIELD ELECTROSTATICS */
 602             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 603             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 604
 605             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 606
 607             /* Update potential sum for this i atom from the interaction with this j atom. */
 608             velec            = _mm_and_pd(velec,cutoff_mask);
 609             velecsum         = _mm_add_pd(velecsum,velec);
 610
 611             fscal            = felec;
 612
 613             fscal            = _mm_and_pd(fscal,cutoff_mask);
 614
 615             /* Calculate temporary vectorial force */
 616             tx               = _mm_mul_pd(fscal,dx22);
 617             ty               = _mm_mul_pd(fscal,dy22);
 618             tz               = _mm_mul_pd(fscal,dz22);
 619
 620             /* Update vectorial force */
 621             fix2             = _mm_add_pd(fix2,tx);
 622             fiy2             = _mm_add_pd(fiy2,ty);
 623             fiz2             = _mm_add_pd(fiz2,tz);
 624
 625             fjx2             = _mm_add_pd(fjx2,tx);
 626             fjy2             = _mm_add_pd(fjy2,ty);
 627             fjz2             = _mm_add_pd(fjz2,tz);
 628
 629             }
 630
 631             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 632
 633             /* Inner loop uses 342 flops */
 634         }
 635
 636         if(jidx<j_index_end)
 637         {
 638
 639             jnrA             = jjnr[jidx];
 640             j_coord_offsetA  = DIM*jnrA;
 641
 642             /* load j atom coordinates */
 643             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 644                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 645
 646             /* Calculate displacement vector */
 647             dx00             = _mm_sub_pd(ix0,jx0);
 648             dy00             = _mm_sub_pd(iy0,jy0);
 649             dz00             = _mm_sub_pd(iz0,jz0);
 650             dx01             = _mm_sub_pd(ix0,jx1);
 651             dy01             = _mm_sub_pd(iy0,jy1);
 652             dz01             = _mm_sub_pd(iz0,jz1);
 653             dx02             = _mm_sub_pd(ix0,jx2);
 654             dy02             = _mm_sub_pd(iy0,jy2);
 655             dz02             = _mm_sub_pd(iz0,jz2);
 656             dx10             = _mm_sub_pd(ix1,jx0);
 657             dy10             = _mm_sub_pd(iy1,jy0);
 658             dz10             = _mm_sub_pd(iz1,jz0);
 659             dx11             = _mm_sub_pd(ix1,jx1);
 660             dy11             = _mm_sub_pd(iy1,jy1);
 661             dz11             = _mm_sub_pd(iz1,jz1);
 662             dx12             = _mm_sub_pd(ix1,jx2);
 663             dy12             = _mm_sub_pd(iy1,jy2);
 664             dz12             = _mm_sub_pd(iz1,jz2);
 665             dx20             = _mm_sub_pd(ix2,jx0);
 666             dy20             = _mm_sub_pd(iy2,jy0);
 667             dz20             = _mm_sub_pd(iz2,jz0);
 668             dx21             = _mm_sub_pd(ix2,jx1);
 669             dy21             = _mm_sub_pd(iy2,jy1);
 670             dz21             = _mm_sub_pd(iz2,jz1);
 671             dx22             = _mm_sub_pd(ix2,jx2);
 672             dy22             = _mm_sub_pd(iy2,jy2);
 673             dz22             = _mm_sub_pd(iz2,jz2);
 674
 675             /* Calculate squared distance and things based on it */
 676             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 677             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 678             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 679             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 680             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 681             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 682             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 683             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 684             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 685
 686             rinv00           = sse2_invsqrt_d(rsq00);
 687             rinv01           = sse2_invsqrt_d(rsq01);
 688             rinv02           = sse2_invsqrt_d(rsq02);
 689             rinv10           = sse2_invsqrt_d(rsq10);
 690             rinv11           = sse2_invsqrt_d(rsq11);
 691             rinv12           = sse2_invsqrt_d(rsq12);
 692             rinv20           = sse2_invsqrt_d(rsq20);
 693             rinv21           = sse2_invsqrt_d(rsq21);
 694             rinv22           = sse2_invsqrt_d(rsq22);
 695
 696             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 697             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 698             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 699             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 700             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 701             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 702             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 703             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 704             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 705
 706             fjx0             = _mm_setzero_pd();
 707             fjy0             = _mm_setzero_pd();
 708             fjz0             = _mm_setzero_pd();
 709             fjx1             = _mm_setzero_pd();
 710             fjy1             = _mm_setzero_pd();
 711             fjz1             = _mm_setzero_pd();
 712             fjx2             = _mm_setzero_pd();
 713             fjy2             = _mm_setzero_pd();
 714             fjz2             = _mm_setzero_pd();
 715
 716             /**************************
 717              * CALCULATE INTERACTIONS *
 718              **************************/
 719
 720             if (gmx_mm_any_lt(rsq00,rcutoff2))
 721             {
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 725             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 726
 727             /* LENNARD-JONES DISPERSION/REPULSION */
 728
 729             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 730             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 731             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 732             vvdw             = _mm_sub_pd(_mm_mul_pd( _mm_sub_pd(vvdw12 , _mm_mul_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 733                                           _mm_mul_pd( _mm_sub_pd(vvdw6,_mm_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 734             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 735
 736             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             velec            = _mm_and_pd(velec,cutoff_mask);
 740             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 741             velecsum         = _mm_add_pd(velecsum,velec);
 742             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 743             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 744             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 745
 746             fscal            = _mm_add_pd(felec,fvdw);
 747
 748             fscal            = _mm_and_pd(fscal,cutoff_mask);
 749
 750             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 751
 752             /* Calculate temporary vectorial force */
 753             tx               = _mm_mul_pd(fscal,dx00);
 754             ty               = _mm_mul_pd(fscal,dy00);
 755             tz               = _mm_mul_pd(fscal,dz00);
 756
 757             /* Update vectorial force */
 758             fix0             = _mm_add_pd(fix0,tx);
 759             fiy0             = _mm_add_pd(fiy0,ty);
 760             fiz0             = _mm_add_pd(fiz0,tz);
 761
 762             fjx0             = _mm_add_pd(fjx0,tx);
 763             fjy0             = _mm_add_pd(fjy0,ty);
 764             fjz0             = _mm_add_pd(fjz0,tz);
 765
 766             }
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             if (gmx_mm_any_lt(rsq01,rcutoff2))
 773             {
 774
 775             /* REACTION-FIELD ELECTROSTATICS */
 776             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 777             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 778
 779             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 780
 781             /* Update potential sum for this i atom from the interaction with this j atom. */
 782             velec            = _mm_and_pd(velec,cutoff_mask);
 783             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 784             velecsum         = _mm_add_pd(velecsum,velec);
 785
 786             fscal            = felec;
 787
 788             fscal            = _mm_and_pd(fscal,cutoff_mask);
 789
 790             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 791
 792             /* Calculate temporary vectorial force */
 793             tx               = _mm_mul_pd(fscal,dx01);
 794             ty               = _mm_mul_pd(fscal,dy01);
 795             tz               = _mm_mul_pd(fscal,dz01);
 796
 797             /* Update vectorial force */
 798             fix0             = _mm_add_pd(fix0,tx);
 799             fiy0             = _mm_add_pd(fiy0,ty);
 800             fiz0             = _mm_add_pd(fiz0,tz);
 801
 802             fjx1             = _mm_add_pd(fjx1,tx);
 803             fjy1             = _mm_add_pd(fjy1,ty);
 804             fjz1             = _mm_add_pd(fjz1,tz);
 805
 806             }
 807
 808             /**************************
 809              * CALCULATE INTERACTIONS *
 810              **************************/
 811
 812             if (gmx_mm_any_lt(rsq02,rcutoff2))
 813             {
 814
 815             /* REACTION-FIELD ELECTROSTATICS */
 816             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 817             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 818
 819             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 820
 821             /* Update potential sum for this i atom from the interaction with this j atom. */
 822             velec            = _mm_and_pd(velec,cutoff_mask);
 823             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 824             velecsum         = _mm_add_pd(velecsum,velec);
 825
 826             fscal            = felec;
 827
 828             fscal            = _mm_and_pd(fscal,cutoff_mask);
 829
 830             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 831
 832             /* Calculate temporary vectorial force */
 833             tx               = _mm_mul_pd(fscal,dx02);
 834             ty               = _mm_mul_pd(fscal,dy02);
 835             tz               = _mm_mul_pd(fscal,dz02);
 836
 837             /* Update vectorial force */
 838             fix0             = _mm_add_pd(fix0,tx);
 839             fiy0             = _mm_add_pd(fiy0,ty);
 840             fiz0             = _mm_add_pd(fiz0,tz);
 841
 842             fjx2             = _mm_add_pd(fjx2,tx);
 843             fjy2             = _mm_add_pd(fjy2,ty);
 844             fjz2             = _mm_add_pd(fjz2,tz);
 845
 846             }
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             if (gmx_mm_any_lt(rsq10,rcutoff2))
 853             {
 854
 855             /* REACTION-FIELD ELECTROSTATICS */
 856             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 857             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 858
 859             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 860
 861             /* Update potential sum for this i atom from the interaction with this j atom. */
 862             velec            = _mm_and_pd(velec,cutoff_mask);
 863             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 864             velecsum         = _mm_add_pd(velecsum,velec);
 865
 866             fscal            = felec;
 867
 868             fscal            = _mm_and_pd(fscal,cutoff_mask);
 869
 870             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 871
 872             /* Calculate temporary vectorial force */
 873             tx               = _mm_mul_pd(fscal,dx10);
 874             ty               = _mm_mul_pd(fscal,dy10);
 875             tz               = _mm_mul_pd(fscal,dz10);
 876
 877             /* Update vectorial force */
 878             fix1             = _mm_add_pd(fix1,tx);
 879             fiy1             = _mm_add_pd(fiy1,ty);
 880             fiz1             = _mm_add_pd(fiz1,tz);
 881
 882             fjx0             = _mm_add_pd(fjx0,tx);
 883             fjy0             = _mm_add_pd(fjy0,ty);
 884             fjz0             = _mm_add_pd(fjz0,tz);
 885
 886             }
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             if (gmx_mm_any_lt(rsq11,rcutoff2))
 893             {
 894
 895             /* REACTION-FIELD ELECTROSTATICS */
 896             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 897             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 898
 899             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 900
 901             /* Update potential sum for this i atom from the interaction with this j atom. */
 902             velec            = _mm_and_pd(velec,cutoff_mask);
 903             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 904             velecsum         = _mm_add_pd(velecsum,velec);
 905
 906             fscal            = felec;
 907
 908             fscal            = _mm_and_pd(fscal,cutoff_mask);
 909
 910             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 911
 912             /* Calculate temporary vectorial force */
 913             tx               = _mm_mul_pd(fscal,dx11);
 914             ty               = _mm_mul_pd(fscal,dy11);
 915             tz               = _mm_mul_pd(fscal,dz11);
 916
 917             /* Update vectorial force */
 918             fix1             = _mm_add_pd(fix1,tx);
 919             fiy1             = _mm_add_pd(fiy1,ty);
 920             fiz1             = _mm_add_pd(fiz1,tz);
 921
 922             fjx1             = _mm_add_pd(fjx1,tx);
 923             fjy1             = _mm_add_pd(fjy1,ty);
 924             fjz1             = _mm_add_pd(fjz1,tz);
 925
 926             }
 927
 928             /**************************
 929              * CALCULATE INTERACTIONS *
 930              **************************/
 931
 932             if (gmx_mm_any_lt(rsq12,rcutoff2))
 933             {
 934
 935             /* REACTION-FIELD ELECTROSTATICS */
 936             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 937             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 938
 939             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 940
 941             /* Update potential sum for this i atom from the interaction with this j atom. */
 942             velec            = _mm_and_pd(velec,cutoff_mask);
 943             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 944             velecsum         = _mm_add_pd(velecsum,velec);
 945
 946             fscal            = felec;
 947
 948             fscal            = _mm_and_pd(fscal,cutoff_mask);
 949
 950             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 951
 952             /* Calculate temporary vectorial force */
 953             tx               = _mm_mul_pd(fscal,dx12);
 954             ty               = _mm_mul_pd(fscal,dy12);
 955             tz               = _mm_mul_pd(fscal,dz12);
 956
 957             /* Update vectorial force */
 958             fix1             = _mm_add_pd(fix1,tx);
 959             fiy1             = _mm_add_pd(fiy1,ty);
 960             fiz1             = _mm_add_pd(fiz1,tz);
 961
 962             fjx2             = _mm_add_pd(fjx2,tx);
 963             fjy2             = _mm_add_pd(fjy2,ty);
 964             fjz2             = _mm_add_pd(fjz2,tz);
 965
 966             }
 967
 968             /**************************
 969              * CALCULATE INTERACTIONS *
 970              **************************/
 971
 972             if (gmx_mm_any_lt(rsq20,rcutoff2))
 973             {
 974
 975             /* REACTION-FIELD ELECTROSTATICS */
 976             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 977             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 978
 979             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 980
 981             /* Update potential sum for this i atom from the interaction with this j atom. */
 982             velec            = _mm_and_pd(velec,cutoff_mask);
 983             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 984             velecsum         = _mm_add_pd(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _mm_and_pd(fscal,cutoff_mask);
 989
 990             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 991
 992             /* Calculate temporary vectorial force */
 993             tx               = _mm_mul_pd(fscal,dx20);
 994             ty               = _mm_mul_pd(fscal,dy20);
 995             tz               = _mm_mul_pd(fscal,dz20);
 996
 997             /* Update vectorial force */
 998             fix2             = _mm_add_pd(fix2,tx);
 999             fiy2             = _mm_add_pd(fiy2,ty);
1000             fiz2             = _mm_add_pd(fiz2,tz);
1001
1002             fjx0             = _mm_add_pd(fjx0,tx);
1003             fjy0             = _mm_add_pd(fjy0,ty);
1004             fjz0             = _mm_add_pd(fjz0,tz);
1005
1006             }
1007
1008             /**************************
1009              * CALCULATE INTERACTIONS *
1010              **************************/
1011
1012             if (gmx_mm_any_lt(rsq21,rcutoff2))
1013             {
1014
1015             /* REACTION-FIELD ELECTROSTATICS */
1016             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
1017             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1018
1019             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1020
1021             /* Update potential sum for this i atom from the interaction with this j atom. */
1022             velec            = _mm_and_pd(velec,cutoff_mask);
1023             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1024             velecsum         = _mm_add_pd(velecsum,velec);
1025
1026             fscal            = felec;
1027
1028             fscal            = _mm_and_pd(fscal,cutoff_mask);
1029
1030             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1031
1032             /* Calculate temporary vectorial force */
1033             tx               = _mm_mul_pd(fscal,dx21);
1034             ty               = _mm_mul_pd(fscal,dy21);
1035             tz               = _mm_mul_pd(fscal,dz21);
1036
1037             /* Update vectorial force */
1038             fix2             = _mm_add_pd(fix2,tx);
1039             fiy2             = _mm_add_pd(fiy2,ty);
1040             fiz2             = _mm_add_pd(fiz2,tz);
1041
1042             fjx1             = _mm_add_pd(fjx1,tx);
1043             fjy1             = _mm_add_pd(fjy1,ty);
1044             fjz1             = _mm_add_pd(fjz1,tz);
1045
1046             }
1047
1048             /**************************
1049              * CALCULATE INTERACTIONS *
1050              **************************/
1051
1052             if (gmx_mm_any_lt(rsq22,rcutoff2))
1053             {
1054
1055             /* REACTION-FIELD ELECTROSTATICS */
1056             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1057             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1058
1059             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1060
1061             /* Update potential sum for this i atom from the interaction with this j atom. */
1062             velec            = _mm_and_pd(velec,cutoff_mask);
1063             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1064             velecsum         = _mm_add_pd(velecsum,velec);
1065
1066             fscal            = felec;
1067
1068             fscal            = _mm_and_pd(fscal,cutoff_mask);
1069
1070             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1071
1072             /* Calculate temporary vectorial force */
1073             tx               = _mm_mul_pd(fscal,dx22);
1074             ty               = _mm_mul_pd(fscal,dy22);
1075             tz               = _mm_mul_pd(fscal,dz22);
1076
1077             /* Update vectorial force */
1078             fix2             = _mm_add_pd(fix2,tx);
1079             fiy2             = _mm_add_pd(fiy2,ty);
1080             fiz2             = _mm_add_pd(fiz2,tz);
1081
1082             fjx2             = _mm_add_pd(fjx2,tx);
1083             fjy2             = _mm_add_pd(fjy2,ty);
1084             fjz2             = _mm_add_pd(fjz2,tz);
1085
1086             }
1087
1088             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1089
1090             /* Inner loop uses 342 flops */
1091         }
1092
1093         /* End of innermost loop */
1094
1095         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1096                                               f+i_coord_offset,fshift+i_shift_offset);
1097
1098         ggid                        = gid[iidx];
1099         /* Update potential energies */
1100         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1101         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1102
1103         /* Increment number of inner iterations */
1104         inneriter                  += j_index_end - j_index_start;
1105
1106         /* Outer loop uses 20 flops */
1107     }
1108
1109     /* Increment number of outer iterations */
1110     outeriter        += nri;
1111
1112     /* Update outer/inner flops */
1113
1114     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*342);
1115 }
1116 /*
1117  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sse2_double
1118  * Electrostatics interaction: ReactionField
1119  * VdW interaction:            LennardJones
1120  * Geometry:                   Water3-Water3
1121  * Calculate force/pot:        Force
1122  */
1123 void
1124 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_sse2_double
1125                     (t_nblist                    * gmx_restrict       nlist,
1126                      rvec                        * gmx_restrict          xx,
1127                      rvec                        * gmx_restrict          ff,
1128                      struct t_forcerec           * gmx_restrict          fr,
1129                      t_mdatoms                   * gmx_restrict     mdatoms,
1130                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1131                      t_nrnb                      * gmx_restrict        nrnb)
1132 {
1133     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1134      * just 0 for non-waters.
1135      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1136      * jnr indices corresponding to data put in the four positions in the SIMD register.
1137      */
1138     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1139     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1140     int              jnrA,jnrB;
1141     int              j_coord_offsetA,j_coord_offsetB;
1142     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1143     real             rcutoff_scalar;
1144     real             *shiftvec,*fshift,*x,*f;
1145     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1146     int              vdwioffset0;
1147     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1148     int              vdwioffset1;
1149     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1150     int              vdwioffset2;
1151     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1152     int              vdwjidx0A,vdwjidx0B;
1153     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1154     int              vdwjidx1A,vdwjidx1B;
1155     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1156     int              vdwjidx2A,vdwjidx2B;
1157     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1158     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1159     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1160     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1161     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1162     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1163     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1164     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1165     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1166     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1167     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1168     real             *charge;
1169     int              nvdwtype;
1170     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1171     int              *vdwtype;
1172     real             *vdwparam;
1173     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1174     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1175     __m128d          dummy_mask,cutoff_mask;
1176     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1177     __m128d          one     = _mm_set1_pd(1.0);
1178     __m128d          two     = _mm_set1_pd(2.0);
1179     x                = xx[0];
1180     f                = ff[0];
1181
1182     nri              = nlist->nri;
1183     iinr             = nlist->iinr;
1184     jindex           = nlist->jindex;
1185     jjnr             = nlist->jjnr;
1186     shiftidx         = nlist->shift;
1187     gid              = nlist->gid;
1188     shiftvec         = fr->shift_vec[0];
1189     fshift           = fr->fshift[0];
1190     facel            = _mm_set1_pd(fr->ic->epsfac);
1191     charge           = mdatoms->chargeA;
1192     krf              = _mm_set1_pd(fr->ic->k_rf);
1193     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1194     crf              = _mm_set1_pd(fr->ic->c_rf);
1195     nvdwtype         = fr->ntype;
1196     vdwparam         = fr->nbfp;
1197     vdwtype          = mdatoms->typeA;
1198
1199     /* Setup water-specific parameters */
1200     inr              = nlist->iinr[0];
1201     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1202     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1203     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1204     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1205
1206     jq0              = _mm_set1_pd(charge[inr+0]);
1207     jq1              = _mm_set1_pd(charge[inr+1]);
1208     jq2              = _mm_set1_pd(charge[inr+2]);
1209     vdwjidx0A        = 2*vdwtype[inr+0];
1210     qq00             = _mm_mul_pd(iq0,jq0);
1211     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1212     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1213     qq01             = _mm_mul_pd(iq0,jq1);
1214     qq02             = _mm_mul_pd(iq0,jq2);
1215     qq10             = _mm_mul_pd(iq1,jq0);
1216     qq11             = _mm_mul_pd(iq1,jq1);
1217     qq12             = _mm_mul_pd(iq1,jq2);
1218     qq20             = _mm_mul_pd(iq2,jq0);
1219     qq21             = _mm_mul_pd(iq2,jq1);
1220     qq22             = _mm_mul_pd(iq2,jq2);
1221
1222     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1223     rcutoff_scalar   = fr->ic->rcoulomb;
1224     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1225     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1226
1227     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1228     rvdw             = _mm_set1_pd(fr->ic->rvdw);
1229
1230     /* Avoid stupid compiler warnings */
1231     jnrA = jnrB = 0;
1232     j_coord_offsetA = 0;
1233     j_coord_offsetB = 0;
1234
1235     outeriter        = 0;
1236     inneriter        = 0;
1237
1238     /* Start outer loop over neighborlists */
1239     for(iidx=0; iidx<nri; iidx++)
1240     {
1241         /* Load shift vector for this list */
1242         i_shift_offset   = DIM*shiftidx[iidx];
1243
1244         /* Load limits for loop over neighbors */
1245         j_index_start    = jindex[iidx];
1246         j_index_end      = jindex[iidx+1];
1247
1248         /* Get outer coordinate index */
1249         inr              = iinr[iidx];
1250         i_coord_offset   = DIM*inr;
1251
1252         /* Load i particle coords and add shift vector */
1253         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1254                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1255
1256         fix0             = _mm_setzero_pd();
1257         fiy0             = _mm_setzero_pd();
1258         fiz0             = _mm_setzero_pd();
1259         fix1             = _mm_setzero_pd();
1260         fiy1             = _mm_setzero_pd();
1261         fiz1             = _mm_setzero_pd();
1262         fix2             = _mm_setzero_pd();
1263         fiy2             = _mm_setzero_pd();
1264         fiz2             = _mm_setzero_pd();
1265
1266         /* Start inner kernel loop */
1267         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1268         {
1269
1270             /* Get j neighbor index, and coordinate index */
1271             jnrA             = jjnr[jidx];
1272             jnrB             = jjnr[jidx+1];
1273             j_coord_offsetA  = DIM*jnrA;
1274             j_coord_offsetB  = DIM*jnrB;
1275
1276             /* load j atom coordinates */
1277             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1278                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1279
1280             /* Calculate displacement vector */
1281             dx00             = _mm_sub_pd(ix0,jx0);
1282             dy00             = _mm_sub_pd(iy0,jy0);
1283             dz00             = _mm_sub_pd(iz0,jz0);
1284             dx01             = _mm_sub_pd(ix0,jx1);
1285             dy01             = _mm_sub_pd(iy0,jy1);
1286             dz01             = _mm_sub_pd(iz0,jz1);
1287             dx02             = _mm_sub_pd(ix0,jx2);
1288             dy02             = _mm_sub_pd(iy0,jy2);
1289             dz02             = _mm_sub_pd(iz0,jz2);
1290             dx10             = _mm_sub_pd(ix1,jx0);
1291             dy10             = _mm_sub_pd(iy1,jy0);
1292             dz10             = _mm_sub_pd(iz1,jz0);
1293             dx11             = _mm_sub_pd(ix1,jx1);
1294             dy11             = _mm_sub_pd(iy1,jy1);
1295             dz11             = _mm_sub_pd(iz1,jz1);
1296             dx12             = _mm_sub_pd(ix1,jx2);
1297             dy12             = _mm_sub_pd(iy1,jy2);
1298             dz12             = _mm_sub_pd(iz1,jz2);
1299             dx20             = _mm_sub_pd(ix2,jx0);
1300             dy20             = _mm_sub_pd(iy2,jy0);
1301             dz20             = _mm_sub_pd(iz2,jz0);
1302             dx21             = _mm_sub_pd(ix2,jx1);
1303             dy21             = _mm_sub_pd(iy2,jy1);
1304             dz21             = _mm_sub_pd(iz2,jz1);
1305             dx22             = _mm_sub_pd(ix2,jx2);
1306             dy22             = _mm_sub_pd(iy2,jy2);
1307             dz22             = _mm_sub_pd(iz2,jz2);
1308
1309             /* Calculate squared distance and things based on it */
1310             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1311             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1312             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1313             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1314             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1315             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1316             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1317             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1318             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1319
1320             rinv00           = sse2_invsqrt_d(rsq00);
1321             rinv01           = sse2_invsqrt_d(rsq01);
1322             rinv02           = sse2_invsqrt_d(rsq02);
1323             rinv10           = sse2_invsqrt_d(rsq10);
1324             rinv11           = sse2_invsqrt_d(rsq11);
1325             rinv12           = sse2_invsqrt_d(rsq12);
1326             rinv20           = sse2_invsqrt_d(rsq20);
1327             rinv21           = sse2_invsqrt_d(rsq21);
1328             rinv22           = sse2_invsqrt_d(rsq22);
1329
1330             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1331             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1332             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1333             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1334             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1335             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1336             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1337             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1338             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1339
1340             fjx0             = _mm_setzero_pd();
1341             fjy0             = _mm_setzero_pd();
1342             fjz0             = _mm_setzero_pd();
1343             fjx1             = _mm_setzero_pd();
1344             fjy1             = _mm_setzero_pd();
1345             fjz1             = _mm_setzero_pd();
1346             fjx2             = _mm_setzero_pd();
1347             fjy2             = _mm_setzero_pd();
1348             fjz2             = _mm_setzero_pd();
1349
1350             /**************************
1351              * CALCULATE INTERACTIONS *
1352              **************************/
1353
1354             if (gmx_mm_any_lt(rsq00,rcutoff2))
1355             {
1356
1357             /* REACTION-FIELD ELECTROSTATICS */
1358             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1359
1360             /* LENNARD-JONES DISPERSION/REPULSION */
1361
1362             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1363             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1364
1365             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1366
1367             fscal            = _mm_add_pd(felec,fvdw);
1368
1369             fscal            = _mm_and_pd(fscal,cutoff_mask);
1370
1371             /* Calculate temporary vectorial force */
1372             tx               = _mm_mul_pd(fscal,dx00);
1373             ty               = _mm_mul_pd(fscal,dy00);
1374             tz               = _mm_mul_pd(fscal,dz00);
1375
1376             /* Update vectorial force */
1377             fix0             = _mm_add_pd(fix0,tx);
1378             fiy0             = _mm_add_pd(fiy0,ty);
1379             fiz0             = _mm_add_pd(fiz0,tz);
1380
1381             fjx0             = _mm_add_pd(fjx0,tx);
1382             fjy0             = _mm_add_pd(fjy0,ty);
1383             fjz0             = _mm_add_pd(fjz0,tz);
1384
1385             }
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             if (gmx_mm_any_lt(rsq01,rcutoff2))
1392             {
1393
1394             /* REACTION-FIELD ELECTROSTATICS */
1395             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1396
1397             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1398
1399             fscal            = felec;
1400
1401             fscal            = _mm_and_pd(fscal,cutoff_mask);
1402
1403             /* Calculate temporary vectorial force */
1404             tx               = _mm_mul_pd(fscal,dx01);
1405             ty               = _mm_mul_pd(fscal,dy01);
1406             tz               = _mm_mul_pd(fscal,dz01);
1407
1408             /* Update vectorial force */
1409             fix0             = _mm_add_pd(fix0,tx);
1410             fiy0             = _mm_add_pd(fiy0,ty);
1411             fiz0             = _mm_add_pd(fiz0,tz);
1412
1413             fjx1             = _mm_add_pd(fjx1,tx);
1414             fjy1             = _mm_add_pd(fjy1,ty);
1415             fjz1             = _mm_add_pd(fjz1,tz);
1416
1417             }
1418
1419             /**************************
1420              * CALCULATE INTERACTIONS *
1421              **************************/
1422
1423             if (gmx_mm_any_lt(rsq02,rcutoff2))
1424             {
1425
1426             /* REACTION-FIELD ELECTROSTATICS */
1427             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1428
1429             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1430
1431             fscal            = felec;
1432
1433             fscal            = _mm_and_pd(fscal,cutoff_mask);
1434
1435             /* Calculate temporary vectorial force */
1436             tx               = _mm_mul_pd(fscal,dx02);
1437             ty               = _mm_mul_pd(fscal,dy02);
1438             tz               = _mm_mul_pd(fscal,dz02);
1439
1440             /* Update vectorial force */
1441             fix0             = _mm_add_pd(fix0,tx);
1442             fiy0             = _mm_add_pd(fiy0,ty);
1443             fiz0             = _mm_add_pd(fiz0,tz);
1444
1445             fjx2             = _mm_add_pd(fjx2,tx);
1446             fjy2             = _mm_add_pd(fjy2,ty);
1447             fjz2             = _mm_add_pd(fjz2,tz);
1448
1449             }
1450
1451             /**************************
1452              * CALCULATE INTERACTIONS *
1453              **************************/
1454
1455             if (gmx_mm_any_lt(rsq10,rcutoff2))
1456             {
1457
1458             /* REACTION-FIELD ELECTROSTATICS */
1459             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1460
1461             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1462
1463             fscal            = felec;
1464
1465             fscal            = _mm_and_pd(fscal,cutoff_mask);
1466
1467             /* Calculate temporary vectorial force */
1468             tx               = _mm_mul_pd(fscal,dx10);
1469             ty               = _mm_mul_pd(fscal,dy10);
1470             tz               = _mm_mul_pd(fscal,dz10);
1471
1472             /* Update vectorial force */
1473             fix1             = _mm_add_pd(fix1,tx);
1474             fiy1             = _mm_add_pd(fiy1,ty);
1475             fiz1             = _mm_add_pd(fiz1,tz);
1476
1477             fjx0             = _mm_add_pd(fjx0,tx);
1478             fjy0             = _mm_add_pd(fjy0,ty);
1479             fjz0             = _mm_add_pd(fjz0,tz);
1480
1481             }
1482
1483             /**************************
1484              * CALCULATE INTERACTIONS *
1485              **************************/
1486
1487             if (gmx_mm_any_lt(rsq11,rcutoff2))
1488             {
1489
1490             /* REACTION-FIELD ELECTROSTATICS */
1491             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1492
1493             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1494
1495             fscal            = felec;
1496
1497             fscal            = _mm_and_pd(fscal,cutoff_mask);
1498
1499             /* Calculate temporary vectorial force */
1500             tx               = _mm_mul_pd(fscal,dx11);
1501             ty               = _mm_mul_pd(fscal,dy11);
1502             tz               = _mm_mul_pd(fscal,dz11);
1503
1504             /* Update vectorial force */
1505             fix1             = _mm_add_pd(fix1,tx);
1506             fiy1             = _mm_add_pd(fiy1,ty);
1507             fiz1             = _mm_add_pd(fiz1,tz);
1508
1509             fjx1             = _mm_add_pd(fjx1,tx);
1510             fjy1             = _mm_add_pd(fjy1,ty);
1511             fjz1             = _mm_add_pd(fjz1,tz);
1512
1513             }
1514
1515             /**************************
1516              * CALCULATE INTERACTIONS *
1517              **************************/
1518
1519             if (gmx_mm_any_lt(rsq12,rcutoff2))
1520             {
1521
1522             /* REACTION-FIELD ELECTROSTATICS */
1523             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1524
1525             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1526
1527             fscal            = felec;
1528
1529             fscal            = _mm_and_pd(fscal,cutoff_mask);
1530
1531             /* Calculate temporary vectorial force */
1532             tx               = _mm_mul_pd(fscal,dx12);
1533             ty               = _mm_mul_pd(fscal,dy12);
1534             tz               = _mm_mul_pd(fscal,dz12);
1535
1536             /* Update vectorial force */
1537             fix1             = _mm_add_pd(fix1,tx);
1538             fiy1             = _mm_add_pd(fiy1,ty);
1539             fiz1             = _mm_add_pd(fiz1,tz);
1540
1541             fjx2             = _mm_add_pd(fjx2,tx);
1542             fjy2             = _mm_add_pd(fjy2,ty);
1543             fjz2             = _mm_add_pd(fjz2,tz);
1544
1545             }
1546
1547             /**************************
1548              * CALCULATE INTERACTIONS *
1549              **************************/
1550
1551             if (gmx_mm_any_lt(rsq20,rcutoff2))
1552             {
1553
1554             /* REACTION-FIELD ELECTROSTATICS */
1555             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1556
1557             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1558
1559             fscal            = felec;
1560
1561             fscal            = _mm_and_pd(fscal,cutoff_mask);
1562
1563             /* Calculate temporary vectorial force */
1564             tx               = _mm_mul_pd(fscal,dx20);
1565             ty               = _mm_mul_pd(fscal,dy20);
1566             tz               = _mm_mul_pd(fscal,dz20);
1567
1568             /* Update vectorial force */
1569             fix2             = _mm_add_pd(fix2,tx);
1570             fiy2             = _mm_add_pd(fiy2,ty);
1571             fiz2             = _mm_add_pd(fiz2,tz);
1572
1573             fjx0             = _mm_add_pd(fjx0,tx);
1574             fjy0             = _mm_add_pd(fjy0,ty);
1575             fjz0             = _mm_add_pd(fjz0,tz);
1576
1577             }
1578
1579             /**************************
1580              * CALCULATE INTERACTIONS *
1581              **************************/
1582
1583             if (gmx_mm_any_lt(rsq21,rcutoff2))
1584             {
1585
1586             /* REACTION-FIELD ELECTROSTATICS */
1587             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1588
1589             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1590
1591             fscal            = felec;
1592
1593             fscal            = _mm_and_pd(fscal,cutoff_mask);
1594
1595             /* Calculate temporary vectorial force */
1596             tx               = _mm_mul_pd(fscal,dx21);
1597             ty               = _mm_mul_pd(fscal,dy21);
1598             tz               = _mm_mul_pd(fscal,dz21);
1599
1600             /* Update vectorial force */
1601             fix2             = _mm_add_pd(fix2,tx);
1602             fiy2             = _mm_add_pd(fiy2,ty);
1603             fiz2             = _mm_add_pd(fiz2,tz);
1604
1605             fjx1             = _mm_add_pd(fjx1,tx);
1606             fjy1             = _mm_add_pd(fjy1,ty);
1607             fjz1             = _mm_add_pd(fjz1,tz);
1608
1609             }
1610
1611             /**************************
1612              * CALCULATE INTERACTIONS *
1613              **************************/
1614
1615             if (gmx_mm_any_lt(rsq22,rcutoff2))
1616             {
1617
1618             /* REACTION-FIELD ELECTROSTATICS */
1619             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1620
1621             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1622
1623             fscal            = felec;
1624
1625             fscal            = _mm_and_pd(fscal,cutoff_mask);
1626
1627             /* Calculate temporary vectorial force */
1628             tx               = _mm_mul_pd(fscal,dx22);
1629             ty               = _mm_mul_pd(fscal,dy22);
1630             tz               = _mm_mul_pd(fscal,dz22);
1631
1632             /* Update vectorial force */
1633             fix2             = _mm_add_pd(fix2,tx);
1634             fiy2             = _mm_add_pd(fiy2,ty);
1635             fiz2             = _mm_add_pd(fiz2,tz);
1636
1637             fjx2             = _mm_add_pd(fjx2,tx);
1638             fjy2             = _mm_add_pd(fjy2,ty);
1639             fjz2             = _mm_add_pd(fjz2,tz);
1640
1641             }
1642
1643             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1644
1645             /* Inner loop uses 277 flops */
1646         }
1647
1648         if(jidx<j_index_end)
1649         {
1650
1651             jnrA             = jjnr[jidx];
1652             j_coord_offsetA  = DIM*jnrA;
1653
1654             /* load j atom coordinates */
1655             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1656                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1657
1658             /* Calculate displacement vector */
1659             dx00             = _mm_sub_pd(ix0,jx0);
1660             dy00             = _mm_sub_pd(iy0,jy0);
1661             dz00             = _mm_sub_pd(iz0,jz0);
1662             dx01             = _mm_sub_pd(ix0,jx1);
1663             dy01             = _mm_sub_pd(iy0,jy1);
1664             dz01             = _mm_sub_pd(iz0,jz1);
1665             dx02             = _mm_sub_pd(ix0,jx2);
1666             dy02             = _mm_sub_pd(iy0,jy2);
1667             dz02             = _mm_sub_pd(iz0,jz2);
1668             dx10             = _mm_sub_pd(ix1,jx0);
1669             dy10             = _mm_sub_pd(iy1,jy0);
1670             dz10             = _mm_sub_pd(iz1,jz0);
1671             dx11             = _mm_sub_pd(ix1,jx1);
1672             dy11             = _mm_sub_pd(iy1,jy1);
1673             dz11             = _mm_sub_pd(iz1,jz1);
1674             dx12             = _mm_sub_pd(ix1,jx2);
1675             dy12             = _mm_sub_pd(iy1,jy2);
1676             dz12             = _mm_sub_pd(iz1,jz2);
1677             dx20             = _mm_sub_pd(ix2,jx0);
1678             dy20             = _mm_sub_pd(iy2,jy0);
1679             dz20             = _mm_sub_pd(iz2,jz0);
1680             dx21             = _mm_sub_pd(ix2,jx1);
1681             dy21             = _mm_sub_pd(iy2,jy1);
1682             dz21             = _mm_sub_pd(iz2,jz1);
1683             dx22             = _mm_sub_pd(ix2,jx2);
1684             dy22             = _mm_sub_pd(iy2,jy2);
1685             dz22             = _mm_sub_pd(iz2,jz2);
1686
1687             /* Calculate squared distance and things based on it */
1688             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1689             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1690             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1691             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1692             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1693             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1694             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1695             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1696             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1697
1698             rinv00           = sse2_invsqrt_d(rsq00);
1699             rinv01           = sse2_invsqrt_d(rsq01);
1700             rinv02           = sse2_invsqrt_d(rsq02);
1701             rinv10           = sse2_invsqrt_d(rsq10);
1702             rinv11           = sse2_invsqrt_d(rsq11);
1703             rinv12           = sse2_invsqrt_d(rsq12);
1704             rinv20           = sse2_invsqrt_d(rsq20);
1705             rinv21           = sse2_invsqrt_d(rsq21);
1706             rinv22           = sse2_invsqrt_d(rsq22);
1707
1708             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1709             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1710             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1711             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1712             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1713             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1714             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1715             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1716             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1717
1718             fjx0             = _mm_setzero_pd();
1719             fjy0             = _mm_setzero_pd();
1720             fjz0             = _mm_setzero_pd();
1721             fjx1             = _mm_setzero_pd();
1722             fjy1             = _mm_setzero_pd();
1723             fjz1             = _mm_setzero_pd();
1724             fjx2             = _mm_setzero_pd();
1725             fjy2             = _mm_setzero_pd();
1726             fjz2             = _mm_setzero_pd();
1727
1728             /**************************
1729              * CALCULATE INTERACTIONS *
1730              **************************/
1731
1732             if (gmx_mm_any_lt(rsq00,rcutoff2))
1733             {
1734
1735             /* REACTION-FIELD ELECTROSTATICS */
1736             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1737
1738             /* LENNARD-JONES DISPERSION/REPULSION */
1739
1740             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1741             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1742
1743             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1744
1745             fscal            = _mm_add_pd(felec,fvdw);
1746
1747             fscal            = _mm_and_pd(fscal,cutoff_mask);
1748
1749             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1750
1751             /* Calculate temporary vectorial force */
1752             tx               = _mm_mul_pd(fscal,dx00);
1753             ty               = _mm_mul_pd(fscal,dy00);
1754             tz               = _mm_mul_pd(fscal,dz00);
1755
1756             /* Update vectorial force */
1757             fix0             = _mm_add_pd(fix0,tx);
1758             fiy0             = _mm_add_pd(fiy0,ty);
1759             fiz0             = _mm_add_pd(fiz0,tz);
1760
1761             fjx0             = _mm_add_pd(fjx0,tx);
1762             fjy0             = _mm_add_pd(fjy0,ty);
1763             fjz0             = _mm_add_pd(fjz0,tz);
1764
1765             }
1766
1767             /**************************
1768              * CALCULATE INTERACTIONS *
1769              **************************/
1770
1771             if (gmx_mm_any_lt(rsq01,rcutoff2))
1772             {
1773
1774             /* REACTION-FIELD ELECTROSTATICS */
1775             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1776
1777             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1778
1779             fscal            = felec;
1780
1781             fscal            = _mm_and_pd(fscal,cutoff_mask);
1782
1783             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1784
1785             /* Calculate temporary vectorial force */
1786             tx               = _mm_mul_pd(fscal,dx01);
1787             ty               = _mm_mul_pd(fscal,dy01);
1788             tz               = _mm_mul_pd(fscal,dz01);
1789
1790             /* Update vectorial force */
1791             fix0             = _mm_add_pd(fix0,tx);
1792             fiy0             = _mm_add_pd(fiy0,ty);
1793             fiz0             = _mm_add_pd(fiz0,tz);
1794
1795             fjx1             = _mm_add_pd(fjx1,tx);
1796             fjy1             = _mm_add_pd(fjy1,ty);
1797             fjz1             = _mm_add_pd(fjz1,tz);
1798
1799             }
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             if (gmx_mm_any_lt(rsq02,rcutoff2))
1806             {
1807
1808             /* REACTION-FIELD ELECTROSTATICS */
1809             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1810
1811             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1812
1813             fscal            = felec;
1814
1815             fscal            = _mm_and_pd(fscal,cutoff_mask);
1816
1817             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1818
1819             /* Calculate temporary vectorial force */
1820             tx               = _mm_mul_pd(fscal,dx02);
1821             ty               = _mm_mul_pd(fscal,dy02);
1822             tz               = _mm_mul_pd(fscal,dz02);
1823
1824             /* Update vectorial force */
1825             fix0             = _mm_add_pd(fix0,tx);
1826             fiy0             = _mm_add_pd(fiy0,ty);
1827             fiz0             = _mm_add_pd(fiz0,tz);
1828
1829             fjx2             = _mm_add_pd(fjx2,tx);
1830             fjy2             = _mm_add_pd(fjy2,ty);
1831             fjz2             = _mm_add_pd(fjz2,tz);
1832
1833             }
1834
1835             /**************************
1836              * CALCULATE INTERACTIONS *
1837              **************************/
1838
1839             if (gmx_mm_any_lt(rsq10,rcutoff2))
1840             {
1841
1842             /* REACTION-FIELD ELECTROSTATICS */
1843             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1844
1845             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1846
1847             fscal            = felec;
1848
1849             fscal            = _mm_and_pd(fscal,cutoff_mask);
1850
1851             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1852
1853             /* Calculate temporary vectorial force */
1854             tx               = _mm_mul_pd(fscal,dx10);
1855             ty               = _mm_mul_pd(fscal,dy10);
1856             tz               = _mm_mul_pd(fscal,dz10);
1857
1858             /* Update vectorial force */
1859             fix1             = _mm_add_pd(fix1,tx);
1860             fiy1             = _mm_add_pd(fiy1,ty);
1861             fiz1             = _mm_add_pd(fiz1,tz);
1862
1863             fjx0             = _mm_add_pd(fjx0,tx);
1864             fjy0             = _mm_add_pd(fjy0,ty);
1865             fjz0             = _mm_add_pd(fjz0,tz);
1866
1867             }
1868
1869             /**************************
1870              * CALCULATE INTERACTIONS *
1871              **************************/
1872
1873             if (gmx_mm_any_lt(rsq11,rcutoff2))
1874             {
1875
1876             /* REACTION-FIELD ELECTROSTATICS */
1877             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1878
1879             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1880
1881             fscal            = felec;
1882
1883             fscal            = _mm_and_pd(fscal,cutoff_mask);
1884
1885             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1886
1887             /* Calculate temporary vectorial force */
1888             tx               = _mm_mul_pd(fscal,dx11);
1889             ty               = _mm_mul_pd(fscal,dy11);
1890             tz               = _mm_mul_pd(fscal,dz11);
1891
1892             /* Update vectorial force */
1893             fix1             = _mm_add_pd(fix1,tx);
1894             fiy1             = _mm_add_pd(fiy1,ty);
1895             fiz1             = _mm_add_pd(fiz1,tz);
1896
1897             fjx1             = _mm_add_pd(fjx1,tx);
1898             fjy1             = _mm_add_pd(fjy1,ty);
1899             fjz1             = _mm_add_pd(fjz1,tz);
1900
1901             }
1902
1903             /**************************
1904              * CALCULATE INTERACTIONS *
1905              **************************/
1906
1907             if (gmx_mm_any_lt(rsq12,rcutoff2))
1908             {
1909
1910             /* REACTION-FIELD ELECTROSTATICS */
1911             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1912
1913             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1914
1915             fscal            = felec;
1916
1917             fscal            = _mm_and_pd(fscal,cutoff_mask);
1918
1919             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1920
1921             /* Calculate temporary vectorial force */
1922             tx               = _mm_mul_pd(fscal,dx12);
1923             ty               = _mm_mul_pd(fscal,dy12);
1924             tz               = _mm_mul_pd(fscal,dz12);
1925
1926             /* Update vectorial force */
1927             fix1             = _mm_add_pd(fix1,tx);
1928             fiy1             = _mm_add_pd(fiy1,ty);
1929             fiz1             = _mm_add_pd(fiz1,tz);
1930
1931             fjx2             = _mm_add_pd(fjx2,tx);
1932             fjy2             = _mm_add_pd(fjy2,ty);
1933             fjz2             = _mm_add_pd(fjz2,tz);
1934
1935             }
1936
1937             /**************************
1938              * CALCULATE INTERACTIONS *
1939              **************************/
1940
1941             if (gmx_mm_any_lt(rsq20,rcutoff2))
1942             {
1943
1944             /* REACTION-FIELD ELECTROSTATICS */
1945             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1946
1947             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1948
1949             fscal            = felec;
1950
1951             fscal            = _mm_and_pd(fscal,cutoff_mask);
1952
1953             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1954
1955             /* Calculate temporary vectorial force */
1956             tx               = _mm_mul_pd(fscal,dx20);
1957             ty               = _mm_mul_pd(fscal,dy20);
1958             tz               = _mm_mul_pd(fscal,dz20);
1959
1960             /* Update vectorial force */
1961             fix2             = _mm_add_pd(fix2,tx);
1962             fiy2             = _mm_add_pd(fiy2,ty);
1963             fiz2             = _mm_add_pd(fiz2,tz);
1964
1965             fjx0             = _mm_add_pd(fjx0,tx);
1966             fjy0             = _mm_add_pd(fjy0,ty);
1967             fjz0             = _mm_add_pd(fjz0,tz);
1968
1969             }
1970
1971             /**************************
1972              * CALCULATE INTERACTIONS *
1973              **************************/
1974
1975             if (gmx_mm_any_lt(rsq21,rcutoff2))
1976             {
1977
1978             /* REACTION-FIELD ELECTROSTATICS */
1979             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1980
1981             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1982
1983             fscal            = felec;
1984
1985             fscal            = _mm_and_pd(fscal,cutoff_mask);
1986
1987             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1988
1989             /* Calculate temporary vectorial force */
1990             tx               = _mm_mul_pd(fscal,dx21);
1991             ty               = _mm_mul_pd(fscal,dy21);
1992             tz               = _mm_mul_pd(fscal,dz21);
1993
1994             /* Update vectorial force */
1995             fix2             = _mm_add_pd(fix2,tx);
1996             fiy2             = _mm_add_pd(fiy2,ty);
1997             fiz2             = _mm_add_pd(fiz2,tz);
1998
1999             fjx1             = _mm_add_pd(fjx1,tx);
2000             fjy1             = _mm_add_pd(fjy1,ty);
2001             fjz1             = _mm_add_pd(fjz1,tz);
2002
2003             }
2004
2005             /**************************
2006              * CALCULATE INTERACTIONS *
2007              **************************/
2008
2009             if (gmx_mm_any_lt(rsq22,rcutoff2))
2010             {
2011
2012             /* REACTION-FIELD ELECTROSTATICS */
2013             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2014
2015             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2016
2017             fscal            = felec;
2018
2019             fscal            = _mm_and_pd(fscal,cutoff_mask);
2020
2021             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2022
2023             /* Calculate temporary vectorial force */
2024             tx               = _mm_mul_pd(fscal,dx22);
2025             ty               = _mm_mul_pd(fscal,dy22);
2026             tz               = _mm_mul_pd(fscal,dz22);
2027
2028             /* Update vectorial force */
2029             fix2             = _mm_add_pd(fix2,tx);
2030             fiy2             = _mm_add_pd(fiy2,ty);
2031             fiz2             = _mm_add_pd(fiz2,tz);
2032
2033             fjx2             = _mm_add_pd(fjx2,tx);
2034             fjy2             = _mm_add_pd(fjy2,ty);
2035             fjz2             = _mm_add_pd(fjz2,tz);
2036
2037             }
2038
2039             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2040
2041             /* Inner loop uses 277 flops */
2042         }
2043
2044         /* End of innermost loop */
2045
2046         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2047                                               f+i_coord_offset,fshift+i_shift_offset);
2048
2049         /* Increment number of inner iterations */
2050         inneriter                  += j_index_end - j_index_start;
2051
2052         /* Outer loop uses 18 flops */
2053     }
2054
2055     /* Increment number of outer iterations */
2056     outeriter        += nri;
2057
2058     /* Update outer/inner flops */
2059
2060     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
2061 }