src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     int              vdwjidx3A,vdwjidx3B;
  95     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  96     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  97     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 113     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 114     __m128d          dummy_mask,cutoff_mask;
 115     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 116     __m128d          one     = _mm_set1_pd(1.0);
 117     __m128d          two     = _mm_set1_pd(2.0);
 118     x                = xx[0];
 119     f                = ff[0];
 120
 121     nri              = nlist->nri;
 122     iinr             = nlist->iinr;
 123     jindex           = nlist->jindex;
 124     jjnr             = nlist->jjnr;
 125     shiftidx         = nlist->shift;
 126     gid              = nlist->gid;
 127     shiftvec         = fr->shift_vec[0];
 128     fshift           = fr->fshift[0];
 129     facel            = _mm_set1_pd(fr->epsfac);
 130     charge           = mdatoms->chargeA;
 131     krf              = _mm_set1_pd(fr->ic->k_rf);
 132     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 133     crf              = _mm_set1_pd(fr->ic->c_rf);
 134     nvdwtype         = fr->ntype;
 135     vdwparam         = fr->nbfp;
 136     vdwtype          = mdatoms->typeA;
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 141     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 142     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq1              = _mm_set1_pd(charge[inr+1]);
 146     jq2              = _mm_set1_pd(charge[inr+2]);
 147     jq3              = _mm_set1_pd(charge[inr+3]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 150     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 151     qq11             = _mm_mul_pd(iq1,jq1);
 152     qq12             = _mm_mul_pd(iq1,jq2);
 153     qq13             = _mm_mul_pd(iq1,jq3);
 154     qq21             = _mm_mul_pd(iq2,jq1);
 155     qq22             = _mm_mul_pd(iq2,jq2);
 156     qq23             = _mm_mul_pd(iq2,jq3);
 157     qq31             = _mm_mul_pd(iq3,jq1);
 158     qq32             = _mm_mul_pd(iq3,jq2);
 159     qq33             = _mm_mul_pd(iq3,jq3);
 160
 161     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 162     rcutoff_scalar   = fr->rcoulomb;
 163     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 164     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 165
 166     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 167     rvdw             = _mm_set1_pd(fr->rvdw);
 168
 169     /* Avoid stupid compiler warnings */
 170     jnrA = jnrB = 0;
 171     j_coord_offsetA = 0;
 172     j_coord_offsetB = 0;
 173
 174     outeriter        = 0;
 175     inneriter        = 0;
 176
 177     /* Start outer loop over neighborlists */
 178     for(iidx=0; iidx<nri; iidx++)
 179     {
 180         /* Load shift vector for this list */
 181         i_shift_offset   = DIM*shiftidx[iidx];
 182
 183         /* Load limits for loop over neighbors */
 184         j_index_start    = jindex[iidx];
 185         j_index_end      = jindex[iidx+1];
 186
 187         /* Get outer coordinate index */
 188         inr              = iinr[iidx];
 189         i_coord_offset   = DIM*inr;
 190
 191         /* Load i particle coords and add shift vector */
 192         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 193                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 194
 195         fix0             = _mm_setzero_pd();
 196         fiy0             = _mm_setzero_pd();
 197         fiz0             = _mm_setzero_pd();
 198         fix1             = _mm_setzero_pd();
 199         fiy1             = _mm_setzero_pd();
 200         fiz1             = _mm_setzero_pd();
 201         fix2             = _mm_setzero_pd();
 202         fiy2             = _mm_setzero_pd();
 203         fiz2             = _mm_setzero_pd();
 204         fix3             = _mm_setzero_pd();
 205         fiy3             = _mm_setzero_pd();
 206         fiz3             = _mm_setzero_pd();
 207
 208         /* Reset potential sums */
 209         velecsum         = _mm_setzero_pd();
 210         vvdwsum          = _mm_setzero_pd();
 211
 212         /* Start inner kernel loop */
 213         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 214         {
 215
 216             /* Get j neighbor index, and coordinate index */
 217             jnrA             = jjnr[jidx];
 218             jnrB             = jjnr[jidx+1];
 219             j_coord_offsetA  = DIM*jnrA;
 220             j_coord_offsetB  = DIM*jnrB;
 221
 222             /* load j atom coordinates */
 223             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 224                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 225                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 226
 227             /* Calculate displacement vector */
 228             dx00             = _mm_sub_pd(ix0,jx0);
 229             dy00             = _mm_sub_pd(iy0,jy0);
 230             dz00             = _mm_sub_pd(iz0,jz0);
 231             dx11             = _mm_sub_pd(ix1,jx1);
 232             dy11             = _mm_sub_pd(iy1,jy1);
 233             dz11             = _mm_sub_pd(iz1,jz1);
 234             dx12             = _mm_sub_pd(ix1,jx2);
 235             dy12             = _mm_sub_pd(iy1,jy2);
 236             dz12             = _mm_sub_pd(iz1,jz2);
 237             dx13             = _mm_sub_pd(ix1,jx3);
 238             dy13             = _mm_sub_pd(iy1,jy3);
 239             dz13             = _mm_sub_pd(iz1,jz3);
 240             dx21             = _mm_sub_pd(ix2,jx1);
 241             dy21             = _mm_sub_pd(iy2,jy1);
 242             dz21             = _mm_sub_pd(iz2,jz1);
 243             dx22             = _mm_sub_pd(ix2,jx2);
 244             dy22             = _mm_sub_pd(iy2,jy2);
 245             dz22             = _mm_sub_pd(iz2,jz2);
 246             dx23             = _mm_sub_pd(ix2,jx3);
 247             dy23             = _mm_sub_pd(iy2,jy3);
 248             dz23             = _mm_sub_pd(iz2,jz3);
 249             dx31             = _mm_sub_pd(ix3,jx1);
 250             dy31             = _mm_sub_pd(iy3,jy1);
 251             dz31             = _mm_sub_pd(iz3,jz1);
 252             dx32             = _mm_sub_pd(ix3,jx2);
 253             dy32             = _mm_sub_pd(iy3,jy2);
 254             dz32             = _mm_sub_pd(iz3,jz2);
 255             dx33             = _mm_sub_pd(ix3,jx3);
 256             dy33             = _mm_sub_pd(iy3,jy3);
 257             dz33             = _mm_sub_pd(iz3,jz3);
 258
 259             /* Calculate squared distance and things based on it */
 260             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 261             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 262             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 263             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 264             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 265             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 266             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 267             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 268             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 269             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 270
 271             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 272             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 273             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 274             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 275             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 276             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 277             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 278             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 279             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 280
 281             rinvsq00         = gmx_mm_inv_pd(rsq00);
 282             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 283             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 284             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 285             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 286             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 287             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 288             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 289             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 290             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 291
 292             fjx0             = _mm_setzero_pd();
 293             fjy0             = _mm_setzero_pd();
 294             fjz0             = _mm_setzero_pd();
 295             fjx1             = _mm_setzero_pd();
 296             fjy1             = _mm_setzero_pd();
 297             fjz1             = _mm_setzero_pd();
 298             fjx2             = _mm_setzero_pd();
 299             fjy2             = _mm_setzero_pd();
 300             fjz2             = _mm_setzero_pd();
 301             fjx3             = _mm_setzero_pd();
 302             fjy3             = _mm_setzero_pd();
 303             fjz3             = _mm_setzero_pd();
 304
 305             /**************************
 306              * CALCULATE INTERACTIONS *
 307              **************************/
 308
 309             if (gmx_mm_any_lt(rsq00,rcutoff2))
 310             {
 311
 312             /* LENNARD-JONES DISPERSION/REPULSION */
 313
 314             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 315             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 316             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 317             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 318                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 319             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 320
 321             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 322
 323             /* Update potential sum for this i atom from the interaction with this j atom. */
 324             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 325             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 326
 327             fscal            = fvdw;
 328
 329             fscal            = _mm_and_pd(fscal,cutoff_mask);
 330
 331             /* Update vectorial force */
 332             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 333             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 334             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 335
 336             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 337             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 338             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 339
 340             }
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             if (gmx_mm_any_lt(rsq11,rcutoff2))
 347             {
 348
 349             /* REACTION-FIELD ELECTROSTATICS */
 350             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 351             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 352
 353             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velec            = _mm_and_pd(velec,cutoff_mask);
 357             velecsum         = _mm_add_pd(velecsum,velec);
 358
 359             fscal            = felec;
 360
 361             fscal            = _mm_and_pd(fscal,cutoff_mask);
 362
 363             /* Update vectorial force */
 364             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 365             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 366             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 367
 368             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 369             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 370             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 371
 372             }
 373
 374             /**************************
 375              * CALCULATE INTERACTIONS *
 376              **************************/
 377
 378             if (gmx_mm_any_lt(rsq12,rcutoff2))
 379             {
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 383             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 384
 385             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velec            = _mm_and_pd(velec,cutoff_mask);
 389             velecsum         = _mm_add_pd(velecsum,velec);
 390
 391             fscal            = felec;
 392
 393             fscal            = _mm_and_pd(fscal,cutoff_mask);
 394
 395             /* Update vectorial force */
 396             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 397             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 398             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 399
 400             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 401             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 402             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 403
 404             }
 405
 406             /**************************
 407              * CALCULATE INTERACTIONS *
 408              **************************/
 409
 410             if (gmx_mm_any_lt(rsq13,rcutoff2))
 411             {
 412
 413             /* REACTION-FIELD ELECTROSTATICS */
 414             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 415             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 416
 417             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velec            = _mm_and_pd(velec,cutoff_mask);
 421             velecsum         = _mm_add_pd(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425             fscal            = _mm_and_pd(fscal,cutoff_mask);
 426
 427             /* Update vectorial force */
 428             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 429             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 430             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 431
 432             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 433             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 434             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 435
 436             }
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             if (gmx_mm_any_lt(rsq21,rcutoff2))
 443             {
 444
 445             /* REACTION-FIELD ELECTROSTATICS */
 446             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 447             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 448
 449             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 450
 451             /* Update potential sum for this i atom from the interaction with this j atom. */
 452             velec            = _mm_and_pd(velec,cutoff_mask);
 453             velecsum         = _mm_add_pd(velecsum,velec);
 454
 455             fscal            = felec;
 456
 457             fscal            = _mm_and_pd(fscal,cutoff_mask);
 458
 459             /* Update vectorial force */
 460             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 461             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 462             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 463
 464             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 465             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 466             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 467
 468             }
 469
 470             /**************************
 471              * CALCULATE INTERACTIONS *
 472              **************************/
 473
 474             if (gmx_mm_any_lt(rsq22,rcutoff2))
 475             {
 476
 477             /* REACTION-FIELD ELECTROSTATICS */
 478             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 479             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 480
 481             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 482
 483             /* Update potential sum for this i atom from the interaction with this j atom. */
 484             velec            = _mm_and_pd(velec,cutoff_mask);
 485             velecsum         = _mm_add_pd(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             fscal            = _mm_and_pd(fscal,cutoff_mask);
 490
 491             /* Update vectorial force */
 492             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 493             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 494             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 495
 496             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 497             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 498             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 499
 500             }
 501
 502             /**************************
 503              * CALCULATE INTERACTIONS *
 504              **************************/
 505
 506             if (gmx_mm_any_lt(rsq23,rcutoff2))
 507             {
 508
 509             /* REACTION-FIELD ELECTROSTATICS */
 510             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 511             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 512
 513             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 514
 515             /* Update potential sum for this i atom from the interaction with this j atom. */
 516             velec            = _mm_and_pd(velec,cutoff_mask);
 517             velecsum         = _mm_add_pd(velecsum,velec);
 518
 519             fscal            = felec;
 520
 521             fscal            = _mm_and_pd(fscal,cutoff_mask);
 522
 523             /* Update vectorial force */
 524             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 525             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 526             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 527
 528             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 529             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 530             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 531
 532             }
 533
 534             /**************************
 535              * CALCULATE INTERACTIONS *
 536              **************************/
 537
 538             if (gmx_mm_any_lt(rsq31,rcutoff2))
 539             {
 540
 541             /* REACTION-FIELD ELECTROSTATICS */
 542             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 543             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 544
 545             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 546
 547             /* Update potential sum for this i atom from the interaction with this j atom. */
 548             velec            = _mm_and_pd(velec,cutoff_mask);
 549             velecsum         = _mm_add_pd(velecsum,velec);
 550
 551             fscal            = felec;
 552
 553             fscal            = _mm_and_pd(fscal,cutoff_mask);
 554
 555             /* Update vectorial force */
 556             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 557             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 558             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 559
 560             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 561             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 562             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 563
 564             }
 565
 566             /**************************
 567              * CALCULATE INTERACTIONS *
 568              **************************/
 569
 570             if (gmx_mm_any_lt(rsq32,rcutoff2))
 571             {
 572
 573             /* REACTION-FIELD ELECTROSTATICS */
 574             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 575             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 576
 577             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 578
 579             /* Update potential sum for this i atom from the interaction with this j atom. */
 580             velec            = _mm_and_pd(velec,cutoff_mask);
 581             velecsum         = _mm_add_pd(velecsum,velec);
 582
 583             fscal            = felec;
 584
 585             fscal            = _mm_and_pd(fscal,cutoff_mask);
 586
 587             /* Update vectorial force */
 588             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 589             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 590             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 591
 592             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 593             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 594             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 595
 596             }
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             if (gmx_mm_any_lt(rsq33,rcutoff2))
 603             {
 604
 605             /* REACTION-FIELD ELECTROSTATICS */
 606             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 607             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 608
 609             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 610
 611             /* Update potential sum for this i atom from the interaction with this j atom. */
 612             velec            = _mm_and_pd(velec,cutoff_mask);
 613             velecsum         = _mm_add_pd(velecsum,velec);
 614
 615             fscal            = felec;
 616
 617             fscal            = _mm_and_pd(fscal,cutoff_mask);
 618
 619             /* Update vectorial force */
 620             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 621             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 622             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 623
 624             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 625             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 626             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 627
 628             }
 629
 630             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 631
 632             /* Inner loop uses 398 flops */
 633         }
 634
 635         if(jidx<j_index_end)
 636         {
 637
 638             jnrA             = jjnr[jidx];
 639             j_coord_offsetA  = DIM*jnrA;
 640
 641             /* load j atom coordinates */
 642             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 643                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 644                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 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             dx11             = _mm_sub_pd(ix1,jx1);
 651             dy11             = _mm_sub_pd(iy1,jy1);
 652             dz11             = _mm_sub_pd(iz1,jz1);
 653             dx12             = _mm_sub_pd(ix1,jx2);
 654             dy12             = _mm_sub_pd(iy1,jy2);
 655             dz12             = _mm_sub_pd(iz1,jz2);
 656             dx13             = _mm_sub_pd(ix1,jx3);
 657             dy13             = _mm_sub_pd(iy1,jy3);
 658             dz13             = _mm_sub_pd(iz1,jz3);
 659             dx21             = _mm_sub_pd(ix2,jx1);
 660             dy21             = _mm_sub_pd(iy2,jy1);
 661             dz21             = _mm_sub_pd(iz2,jz1);
 662             dx22             = _mm_sub_pd(ix2,jx2);
 663             dy22             = _mm_sub_pd(iy2,jy2);
 664             dz22             = _mm_sub_pd(iz2,jz2);
 665             dx23             = _mm_sub_pd(ix2,jx3);
 666             dy23             = _mm_sub_pd(iy2,jy3);
 667             dz23             = _mm_sub_pd(iz2,jz3);
 668             dx31             = _mm_sub_pd(ix3,jx1);
 669             dy31             = _mm_sub_pd(iy3,jy1);
 670             dz31             = _mm_sub_pd(iz3,jz1);
 671             dx32             = _mm_sub_pd(ix3,jx2);
 672             dy32             = _mm_sub_pd(iy3,jy2);
 673             dz32             = _mm_sub_pd(iz3,jz2);
 674             dx33             = _mm_sub_pd(ix3,jx3);
 675             dy33             = _mm_sub_pd(iy3,jy3);
 676             dz33             = _mm_sub_pd(iz3,jz3);
 677
 678             /* Calculate squared distance and things based on it */
 679             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 680             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 681             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 682             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 683             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 684             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 685             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 686             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 687             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 688             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 689
 690             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 691             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 692             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 693             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 694             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 695             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 696             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 697             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 698             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 699
 700             rinvsq00         = gmx_mm_inv_pd(rsq00);
 701             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 702             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 703             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 704             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 705             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 706             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 707             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 708             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 709             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 710
 711             fjx0             = _mm_setzero_pd();
 712             fjy0             = _mm_setzero_pd();
 713             fjz0             = _mm_setzero_pd();
 714             fjx1             = _mm_setzero_pd();
 715             fjy1             = _mm_setzero_pd();
 716             fjz1             = _mm_setzero_pd();
 717             fjx2             = _mm_setzero_pd();
 718             fjy2             = _mm_setzero_pd();
 719             fjz2             = _mm_setzero_pd();
 720             fjx3             = _mm_setzero_pd();
 721             fjy3             = _mm_setzero_pd();
 722             fjz3             = _mm_setzero_pd();
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             if (gmx_mm_any_lt(rsq00,rcutoff2))
 729             {
 730
 731             /* LENNARD-JONES DISPERSION/REPULSION */
 732
 733             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 734             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 735             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 736             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 737                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 738             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 739
 740             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 741
 742             /* Update potential sum for this i atom from the interaction with this j atom. */
 743             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 744             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 745             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 746
 747             fscal            = fvdw;
 748
 749             fscal            = _mm_and_pd(fscal,cutoff_mask);
 750
 751             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 752
 753             /* Update vectorial force */
 754             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 755             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 756             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 757
 758             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 759             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 760             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 761
 762             }
 763
 764             /**************************
 765              * CALCULATE INTERACTIONS *
 766              **************************/
 767
 768             if (gmx_mm_any_lt(rsq11,rcutoff2))
 769             {
 770
 771             /* REACTION-FIELD ELECTROSTATICS */
 772             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 773             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 774
 775             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_and_pd(velec,cutoff_mask);
 779             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 780             velecsum         = _mm_add_pd(velecsum,velec);
 781
 782             fscal            = felec;
 783
 784             fscal            = _mm_and_pd(fscal,cutoff_mask);
 785
 786             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 787
 788             /* Update vectorial force */
 789             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 790             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 791             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 792
 793             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 794             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 795             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 796
 797             }
 798
 799             /**************************
 800              * CALCULATE INTERACTIONS *
 801              **************************/
 802
 803             if (gmx_mm_any_lt(rsq12,rcutoff2))
 804             {
 805
 806             /* REACTION-FIELD ELECTROSTATICS */
 807             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 808             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 809
 810             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 811
 812             /* Update potential sum for this i atom from the interaction with this j atom. */
 813             velec            = _mm_and_pd(velec,cutoff_mask);
 814             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 815             velecsum         = _mm_add_pd(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _mm_and_pd(fscal,cutoff_mask);
 820
 821             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 822
 823             /* Update vectorial force */
 824             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 825             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 826             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 827
 828             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 829             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 830             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 831
 832             }
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             if (gmx_mm_any_lt(rsq13,rcutoff2))
 839             {
 840
 841             /* REACTION-FIELD ELECTROSTATICS */
 842             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 843             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 844
 845             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 846
 847             /* Update potential sum for this i atom from the interaction with this j atom. */
 848             velec            = _mm_and_pd(velec,cutoff_mask);
 849             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 850             velecsum         = _mm_add_pd(velecsum,velec);
 851
 852             fscal            = felec;
 853
 854             fscal            = _mm_and_pd(fscal,cutoff_mask);
 855
 856             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 857
 858             /* Update vectorial force */
 859             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 860             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 861             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 862
 863             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 864             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 865             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 866
 867             }
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             if (gmx_mm_any_lt(rsq21,rcutoff2))
 874             {
 875
 876             /* REACTION-FIELD ELECTROSTATICS */
 877             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 878             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 879
 880             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 881
 882             /* Update potential sum for this i atom from the interaction with this j atom. */
 883             velec            = _mm_and_pd(velec,cutoff_mask);
 884             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 885             velecsum         = _mm_add_pd(velecsum,velec);
 886
 887             fscal            = felec;
 888
 889             fscal            = _mm_and_pd(fscal,cutoff_mask);
 890
 891             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 892
 893             /* Update vectorial force */
 894             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 895             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 896             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 897
 898             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 899             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 900             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 901
 902             }
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             if (gmx_mm_any_lt(rsq22,rcutoff2))
 909             {
 910
 911             /* REACTION-FIELD ELECTROSTATICS */
 912             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 913             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 914
 915             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 916
 917             /* Update potential sum for this i atom from the interaction with this j atom. */
 918             velec            = _mm_and_pd(velec,cutoff_mask);
 919             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 920             velecsum         = _mm_add_pd(velecsum,velec);
 921
 922             fscal            = felec;
 923
 924             fscal            = _mm_and_pd(fscal,cutoff_mask);
 925
 926             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 927
 928             /* Update vectorial force */
 929             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 930             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 931             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 932
 933             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 934             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 935             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 936
 937             }
 938
 939             /**************************
 940              * CALCULATE INTERACTIONS *
 941              **************************/
 942
 943             if (gmx_mm_any_lt(rsq23,rcutoff2))
 944             {
 945
 946             /* REACTION-FIELD ELECTROSTATICS */
 947             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 948             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 949
 950             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 951
 952             /* Update potential sum for this i atom from the interaction with this j atom. */
 953             velec            = _mm_and_pd(velec,cutoff_mask);
 954             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 955             velecsum         = _mm_add_pd(velecsum,velec);
 956
 957             fscal            = felec;
 958
 959             fscal            = _mm_and_pd(fscal,cutoff_mask);
 960
 961             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 962
 963             /* Update vectorial force */
 964             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 965             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 966             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 967
 968             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 969             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 970             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 971
 972             }
 973
 974             /**************************
 975              * CALCULATE INTERACTIONS *
 976              **************************/
 977
 978             if (gmx_mm_any_lt(rsq31,rcutoff2))
 979             {
 980
 981             /* REACTION-FIELD ELECTROSTATICS */
 982             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 983             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 984
 985             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 986
 987             /* Update potential sum for this i atom from the interaction with this j atom. */
 988             velec            = _mm_and_pd(velec,cutoff_mask);
 989             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 990             velecsum         = _mm_add_pd(velecsum,velec);
 991
 992             fscal            = felec;
 993
 994             fscal            = _mm_and_pd(fscal,cutoff_mask);
 995
 996             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 997
 998             /* Update vectorial force */
 999             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1000             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1001             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1002
1003             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1004             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1005             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1006
1007             }
1008
1009             /**************************
1010              * CALCULATE INTERACTIONS *
1011              **************************/
1012
1013             if (gmx_mm_any_lt(rsq32,rcutoff2))
1014             {
1015
1016             /* REACTION-FIELD ELECTROSTATICS */
1017             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
1018             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1019
1020             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1021
1022             /* Update potential sum for this i atom from the interaction with this j atom. */
1023             velec            = _mm_and_pd(velec,cutoff_mask);
1024             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1025             velecsum         = _mm_add_pd(velecsum,velec);
1026
1027             fscal            = felec;
1028
1029             fscal            = _mm_and_pd(fscal,cutoff_mask);
1030
1031             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1032
1033             /* Update vectorial force */
1034             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1035             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1036             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1037
1038             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1039             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1040             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1041
1042             }
1043
1044             /**************************
1045              * CALCULATE INTERACTIONS *
1046              **************************/
1047
1048             if (gmx_mm_any_lt(rsq33,rcutoff2))
1049             {
1050
1051             /* REACTION-FIELD ELECTROSTATICS */
1052             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
1053             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1054
1055             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1056
1057             /* Update potential sum for this i atom from the interaction with this j atom. */
1058             velec            = _mm_and_pd(velec,cutoff_mask);
1059             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1060             velecsum         = _mm_add_pd(velecsum,velec);
1061
1062             fscal            = felec;
1063
1064             fscal            = _mm_and_pd(fscal,cutoff_mask);
1065
1066             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1067
1068             /* Update vectorial force */
1069             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1070             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1071             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1072
1073             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1074             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1075             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1076
1077             }
1078
1079             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1080
1081             /* Inner loop uses 398 flops */
1082         }
1083
1084         /* End of innermost loop */
1085
1086         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1087                                               f+i_coord_offset,fshift+i_shift_offset);
1088
1089         ggid                        = gid[iidx];
1090         /* Update potential energies */
1091         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1092         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1093
1094         /* Increment number of inner iterations */
1095         inneriter                  += j_index_end - j_index_start;
1096
1097         /* Outer loop uses 26 flops */
1098     }
1099
1100     /* Increment number of outer iterations */
1101     outeriter        += nri;
1102
1103     /* Update outer/inner flops */
1104
1105     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*398);
1106 }
1107 /*
1108  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_double
1109  * Electrostatics interaction: ReactionField
1110  * VdW interaction:            LennardJones
1111  * Geometry:                   Water4-Water4
1112  * Calculate force/pot:        Force
1113  */
1114 void
1115 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_double
1116                     (t_nblist                    * gmx_restrict       nlist,
1117                      rvec                        * gmx_restrict          xx,
1118                      rvec                        * gmx_restrict          ff,
1119                      t_forcerec                  * gmx_restrict          fr,
1120                      t_mdatoms                   * gmx_restrict     mdatoms,
1121                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1122                      t_nrnb                      * gmx_restrict        nrnb)
1123 {
1124     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1125      * just 0 for non-waters.
1126      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1127      * jnr indices corresponding to data put in the four positions in the SIMD register.
1128      */
1129     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1130     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1131     int              jnrA,jnrB;
1132     int              j_coord_offsetA,j_coord_offsetB;
1133     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1134     real             rcutoff_scalar;
1135     real             *shiftvec,*fshift,*x,*f;
1136     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1137     int              vdwioffset0;
1138     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1139     int              vdwioffset1;
1140     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1141     int              vdwioffset2;
1142     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1143     int              vdwioffset3;
1144     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1145     int              vdwjidx0A,vdwjidx0B;
1146     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1147     int              vdwjidx1A,vdwjidx1B;
1148     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1149     int              vdwjidx2A,vdwjidx2B;
1150     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1151     int              vdwjidx3A,vdwjidx3B;
1152     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1153     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1154     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1155     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1156     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1157     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1158     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1159     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1160     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1161     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1162     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1163     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1164     real             *charge;
1165     int              nvdwtype;
1166     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1167     int              *vdwtype;
1168     real             *vdwparam;
1169     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1170     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1171     __m128d          dummy_mask,cutoff_mask;
1172     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1173     __m128d          one     = _mm_set1_pd(1.0);
1174     __m128d          two     = _mm_set1_pd(2.0);
1175     x                = xx[0];
1176     f                = ff[0];
1177
1178     nri              = nlist->nri;
1179     iinr             = nlist->iinr;
1180     jindex           = nlist->jindex;
1181     jjnr             = nlist->jjnr;
1182     shiftidx         = nlist->shift;
1183     gid              = nlist->gid;
1184     shiftvec         = fr->shift_vec[0];
1185     fshift           = fr->fshift[0];
1186     facel            = _mm_set1_pd(fr->epsfac);
1187     charge           = mdatoms->chargeA;
1188     krf              = _mm_set1_pd(fr->ic->k_rf);
1189     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1190     crf              = _mm_set1_pd(fr->ic->c_rf);
1191     nvdwtype         = fr->ntype;
1192     vdwparam         = fr->nbfp;
1193     vdwtype          = mdatoms->typeA;
1194
1195     /* Setup water-specific parameters */
1196     inr              = nlist->iinr[0];
1197     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1198     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1199     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1200     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1201
1202     jq1              = _mm_set1_pd(charge[inr+1]);
1203     jq2              = _mm_set1_pd(charge[inr+2]);
1204     jq3              = _mm_set1_pd(charge[inr+3]);
1205     vdwjidx0A        = 2*vdwtype[inr+0];
1206     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1207     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1208     qq11             = _mm_mul_pd(iq1,jq1);
1209     qq12             = _mm_mul_pd(iq1,jq2);
1210     qq13             = _mm_mul_pd(iq1,jq3);
1211     qq21             = _mm_mul_pd(iq2,jq1);
1212     qq22             = _mm_mul_pd(iq2,jq2);
1213     qq23             = _mm_mul_pd(iq2,jq3);
1214     qq31             = _mm_mul_pd(iq3,jq1);
1215     qq32             = _mm_mul_pd(iq3,jq2);
1216     qq33             = _mm_mul_pd(iq3,jq3);
1217
1218     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1219     rcutoff_scalar   = fr->rcoulomb;
1220     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1221     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1222
1223     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1224     rvdw             = _mm_set1_pd(fr->rvdw);
1225
1226     /* Avoid stupid compiler warnings */
1227     jnrA = jnrB = 0;
1228     j_coord_offsetA = 0;
1229     j_coord_offsetB = 0;
1230
1231     outeriter        = 0;
1232     inneriter        = 0;
1233
1234     /* Start outer loop over neighborlists */
1235     for(iidx=0; iidx<nri; iidx++)
1236     {
1237         /* Load shift vector for this list */
1238         i_shift_offset   = DIM*shiftidx[iidx];
1239
1240         /* Load limits for loop over neighbors */
1241         j_index_start    = jindex[iidx];
1242         j_index_end      = jindex[iidx+1];
1243
1244         /* Get outer coordinate index */
1245         inr              = iinr[iidx];
1246         i_coord_offset   = DIM*inr;
1247
1248         /* Load i particle coords and add shift vector */
1249         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1250                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1251
1252         fix0             = _mm_setzero_pd();
1253         fiy0             = _mm_setzero_pd();
1254         fiz0             = _mm_setzero_pd();
1255         fix1             = _mm_setzero_pd();
1256         fiy1             = _mm_setzero_pd();
1257         fiz1             = _mm_setzero_pd();
1258         fix2             = _mm_setzero_pd();
1259         fiy2             = _mm_setzero_pd();
1260         fiz2             = _mm_setzero_pd();
1261         fix3             = _mm_setzero_pd();
1262         fiy3             = _mm_setzero_pd();
1263         fiz3             = _mm_setzero_pd();
1264
1265         /* Start inner kernel loop */
1266         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1267         {
1268
1269             /* Get j neighbor index, and coordinate index */
1270             jnrA             = jjnr[jidx];
1271             jnrB             = jjnr[jidx+1];
1272             j_coord_offsetA  = DIM*jnrA;
1273             j_coord_offsetB  = DIM*jnrB;
1274
1275             /* load j atom coordinates */
1276             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1277                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1278                                               &jy2,&jz2,&jx3,&jy3,&jz3);
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             dx11             = _mm_sub_pd(ix1,jx1);
1285             dy11             = _mm_sub_pd(iy1,jy1);
1286             dz11             = _mm_sub_pd(iz1,jz1);
1287             dx12             = _mm_sub_pd(ix1,jx2);
1288             dy12             = _mm_sub_pd(iy1,jy2);
1289             dz12             = _mm_sub_pd(iz1,jz2);
1290             dx13             = _mm_sub_pd(ix1,jx3);
1291             dy13             = _mm_sub_pd(iy1,jy3);
1292             dz13             = _mm_sub_pd(iz1,jz3);
1293             dx21             = _mm_sub_pd(ix2,jx1);
1294             dy21             = _mm_sub_pd(iy2,jy1);
1295             dz21             = _mm_sub_pd(iz2,jz1);
1296             dx22             = _mm_sub_pd(ix2,jx2);
1297             dy22             = _mm_sub_pd(iy2,jy2);
1298             dz22             = _mm_sub_pd(iz2,jz2);
1299             dx23             = _mm_sub_pd(ix2,jx3);
1300             dy23             = _mm_sub_pd(iy2,jy3);
1301             dz23             = _mm_sub_pd(iz2,jz3);
1302             dx31             = _mm_sub_pd(ix3,jx1);
1303             dy31             = _mm_sub_pd(iy3,jy1);
1304             dz31             = _mm_sub_pd(iz3,jz1);
1305             dx32             = _mm_sub_pd(ix3,jx2);
1306             dy32             = _mm_sub_pd(iy3,jy2);
1307             dz32             = _mm_sub_pd(iz3,jz2);
1308             dx33             = _mm_sub_pd(ix3,jx3);
1309             dy33             = _mm_sub_pd(iy3,jy3);
1310             dz33             = _mm_sub_pd(iz3,jz3);
1311
1312             /* Calculate squared distance and things based on it */
1313             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1314             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1315             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1316             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1317             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1318             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1319             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1320             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1321             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1322             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1323
1324             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1325             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1326             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1327             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1328             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1329             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1330             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1331             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1332             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1333
1334             rinvsq00         = gmx_mm_inv_pd(rsq00);
1335             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1336             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1337             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1338             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1339             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1340             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1341             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1342             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1343             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1344
1345             fjx0             = _mm_setzero_pd();
1346             fjy0             = _mm_setzero_pd();
1347             fjz0             = _mm_setzero_pd();
1348             fjx1             = _mm_setzero_pd();
1349             fjy1             = _mm_setzero_pd();
1350             fjz1             = _mm_setzero_pd();
1351             fjx2             = _mm_setzero_pd();
1352             fjy2             = _mm_setzero_pd();
1353             fjz2             = _mm_setzero_pd();
1354             fjx3             = _mm_setzero_pd();
1355             fjy3             = _mm_setzero_pd();
1356             fjz3             = _mm_setzero_pd();
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             if (gmx_mm_any_lt(rsq00,rcutoff2))
1363             {
1364
1365             /* LENNARD-JONES DISPERSION/REPULSION */
1366
1367             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1368             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1369
1370             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1371
1372             fscal            = fvdw;
1373
1374             fscal            = _mm_and_pd(fscal,cutoff_mask);
1375
1376             /* Update vectorial force */
1377             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1378             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1379             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1380
1381             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1382             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1383             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1384
1385             }
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             if (gmx_mm_any_lt(rsq11,rcutoff2))
1392             {
1393
1394             /* REACTION-FIELD ELECTROSTATICS */
1395             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1396
1397             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1398
1399             fscal            = felec;
1400
1401             fscal            = _mm_and_pd(fscal,cutoff_mask);
1402
1403             /* Update vectorial force */
1404             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1405             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1406             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1407
1408             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1409             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1410             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1411
1412             }
1413
1414             /**************************
1415              * CALCULATE INTERACTIONS *
1416              **************************/
1417
1418             if (gmx_mm_any_lt(rsq12,rcutoff2))
1419             {
1420
1421             /* REACTION-FIELD ELECTROSTATICS */
1422             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1423
1424             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1425
1426             fscal            = felec;
1427
1428             fscal            = _mm_and_pd(fscal,cutoff_mask);
1429
1430             /* Update vectorial force */
1431             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1432             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1433             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1434
1435             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1436             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1437             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1438
1439             }
1440
1441             /**************************
1442              * CALCULATE INTERACTIONS *
1443              **************************/
1444
1445             if (gmx_mm_any_lt(rsq13,rcutoff2))
1446             {
1447
1448             /* REACTION-FIELD ELECTROSTATICS */
1449             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1450
1451             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1452
1453             fscal            = felec;
1454
1455             fscal            = _mm_and_pd(fscal,cutoff_mask);
1456
1457             /* Update vectorial force */
1458             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1459             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1460             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1461
1462             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1463             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1464             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1465
1466             }
1467
1468             /**************************
1469              * CALCULATE INTERACTIONS *
1470              **************************/
1471
1472             if (gmx_mm_any_lt(rsq21,rcutoff2))
1473             {
1474
1475             /* REACTION-FIELD ELECTROSTATICS */
1476             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1477
1478             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1479
1480             fscal            = felec;
1481
1482             fscal            = _mm_and_pd(fscal,cutoff_mask);
1483
1484             /* Update vectorial force */
1485             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1486             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1487             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1488
1489             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1490             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1491             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1492
1493             }
1494
1495             /**************************
1496              * CALCULATE INTERACTIONS *
1497              **************************/
1498
1499             if (gmx_mm_any_lt(rsq22,rcutoff2))
1500             {
1501
1502             /* REACTION-FIELD ELECTROSTATICS */
1503             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1504
1505             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1506
1507             fscal            = felec;
1508
1509             fscal            = _mm_and_pd(fscal,cutoff_mask);
1510
1511             /* Update vectorial force */
1512             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1513             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1514             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1515
1516             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1517             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1518             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1519
1520             }
1521
1522             /**************************
1523              * CALCULATE INTERACTIONS *
1524              **************************/
1525
1526             if (gmx_mm_any_lt(rsq23,rcutoff2))
1527             {
1528
1529             /* REACTION-FIELD ELECTROSTATICS */
1530             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1531
1532             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1533
1534             fscal            = felec;
1535
1536             fscal            = _mm_and_pd(fscal,cutoff_mask);
1537
1538             /* Update vectorial force */
1539             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1540             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1541             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1542
1543             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1544             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1545             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1546
1547             }
1548
1549             /**************************
1550              * CALCULATE INTERACTIONS *
1551              **************************/
1552
1553             if (gmx_mm_any_lt(rsq31,rcutoff2))
1554             {
1555
1556             /* REACTION-FIELD ELECTROSTATICS */
1557             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1558
1559             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1560
1561             fscal            = felec;
1562
1563             fscal            = _mm_and_pd(fscal,cutoff_mask);
1564
1565             /* Update vectorial force */
1566             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1567             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1568             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1569
1570             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1571             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1572             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1573
1574             }
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             if (gmx_mm_any_lt(rsq32,rcutoff2))
1581             {
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1585
1586             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1587
1588             fscal            = felec;
1589
1590             fscal            = _mm_and_pd(fscal,cutoff_mask);
1591
1592             /* Update vectorial force */
1593             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1594             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1595             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1596
1597             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1598             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1599             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1600
1601             }
1602
1603             /**************************
1604              * CALCULATE INTERACTIONS *
1605              **************************/
1606
1607             if (gmx_mm_any_lt(rsq33,rcutoff2))
1608             {
1609
1610             /* REACTION-FIELD ELECTROSTATICS */
1611             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1612
1613             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1614
1615             fscal            = felec;
1616
1617             fscal            = _mm_and_pd(fscal,cutoff_mask);
1618
1619             /* Update vectorial force */
1620             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1621             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1622             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1623
1624             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1625             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1626             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1627
1628             }
1629
1630             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1631
1632             /* Inner loop uses 333 flops */
1633         }
1634
1635         if(jidx<j_index_end)
1636         {
1637
1638             jnrA             = jjnr[jidx];
1639             j_coord_offsetA  = DIM*jnrA;
1640
1641             /* load j atom coordinates */
1642             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1643                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1644                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1645
1646             /* Calculate displacement vector */
1647             dx00             = _mm_sub_pd(ix0,jx0);
1648             dy00             = _mm_sub_pd(iy0,jy0);
1649             dz00             = _mm_sub_pd(iz0,jz0);
1650             dx11             = _mm_sub_pd(ix1,jx1);
1651             dy11             = _mm_sub_pd(iy1,jy1);
1652             dz11             = _mm_sub_pd(iz1,jz1);
1653             dx12             = _mm_sub_pd(ix1,jx2);
1654             dy12             = _mm_sub_pd(iy1,jy2);
1655             dz12             = _mm_sub_pd(iz1,jz2);
1656             dx13             = _mm_sub_pd(ix1,jx3);
1657             dy13             = _mm_sub_pd(iy1,jy3);
1658             dz13             = _mm_sub_pd(iz1,jz3);
1659             dx21             = _mm_sub_pd(ix2,jx1);
1660             dy21             = _mm_sub_pd(iy2,jy1);
1661             dz21             = _mm_sub_pd(iz2,jz1);
1662             dx22             = _mm_sub_pd(ix2,jx2);
1663             dy22             = _mm_sub_pd(iy2,jy2);
1664             dz22             = _mm_sub_pd(iz2,jz2);
1665             dx23             = _mm_sub_pd(ix2,jx3);
1666             dy23             = _mm_sub_pd(iy2,jy3);
1667             dz23             = _mm_sub_pd(iz2,jz3);
1668             dx31             = _mm_sub_pd(ix3,jx1);
1669             dy31             = _mm_sub_pd(iy3,jy1);
1670             dz31             = _mm_sub_pd(iz3,jz1);
1671             dx32             = _mm_sub_pd(ix3,jx2);
1672             dy32             = _mm_sub_pd(iy3,jy2);
1673             dz32             = _mm_sub_pd(iz3,jz2);
1674             dx33             = _mm_sub_pd(ix3,jx3);
1675             dy33             = _mm_sub_pd(iy3,jy3);
1676             dz33             = _mm_sub_pd(iz3,jz3);
1677
1678             /* Calculate squared distance and things based on it */
1679             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1680             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1681             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1682             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1683             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1684             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1685             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1686             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1687             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1688             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1689
1690             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1691             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1692             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1693             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1694             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1695             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1696             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1697             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1698             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1699
1700             rinvsq00         = gmx_mm_inv_pd(rsq00);
1701             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1702             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1703             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1704             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1705             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1706             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1707             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1708             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1709             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1710
1711             fjx0             = _mm_setzero_pd();
1712             fjy0             = _mm_setzero_pd();
1713             fjz0             = _mm_setzero_pd();
1714             fjx1             = _mm_setzero_pd();
1715             fjy1             = _mm_setzero_pd();
1716             fjz1             = _mm_setzero_pd();
1717             fjx2             = _mm_setzero_pd();
1718             fjy2             = _mm_setzero_pd();
1719             fjz2             = _mm_setzero_pd();
1720             fjx3             = _mm_setzero_pd();
1721             fjy3             = _mm_setzero_pd();
1722             fjz3             = _mm_setzero_pd();
1723
1724             /**************************
1725              * CALCULATE INTERACTIONS *
1726              **************************/
1727
1728             if (gmx_mm_any_lt(rsq00,rcutoff2))
1729             {
1730
1731             /* LENNARD-JONES DISPERSION/REPULSION */
1732
1733             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1734             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1735
1736             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1737
1738             fscal            = fvdw;
1739
1740             fscal            = _mm_and_pd(fscal,cutoff_mask);
1741
1742             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1743
1744             /* Update vectorial force */
1745             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1746             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1747             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1748
1749             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1750             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1751             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1752
1753             }
1754
1755             /**************************
1756              * CALCULATE INTERACTIONS *
1757              **************************/
1758
1759             if (gmx_mm_any_lt(rsq11,rcutoff2))
1760             {
1761
1762             /* REACTION-FIELD ELECTROSTATICS */
1763             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1764
1765             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1766
1767             fscal            = felec;
1768
1769             fscal            = _mm_and_pd(fscal,cutoff_mask);
1770
1771             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1772
1773             /* Update vectorial force */
1774             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1775             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1776             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1777
1778             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1779             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1780             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1781
1782             }
1783
1784             /**************************
1785              * CALCULATE INTERACTIONS *
1786              **************************/
1787
1788             if (gmx_mm_any_lt(rsq12,rcutoff2))
1789             {
1790
1791             /* REACTION-FIELD ELECTROSTATICS */
1792             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1793
1794             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1795
1796             fscal            = felec;
1797
1798             fscal            = _mm_and_pd(fscal,cutoff_mask);
1799
1800             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1801
1802             /* Update vectorial force */
1803             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1804             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1805             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1806
1807             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1808             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1809             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1810
1811             }
1812
1813             /**************************
1814              * CALCULATE INTERACTIONS *
1815              **************************/
1816
1817             if (gmx_mm_any_lt(rsq13,rcutoff2))
1818             {
1819
1820             /* REACTION-FIELD ELECTROSTATICS */
1821             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1822
1823             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1824
1825             fscal            = felec;
1826
1827             fscal            = _mm_and_pd(fscal,cutoff_mask);
1828
1829             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1830
1831             /* Update vectorial force */
1832             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1833             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1834             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1835
1836             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1837             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1838             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1839
1840             }
1841
1842             /**************************
1843              * CALCULATE INTERACTIONS *
1844              **************************/
1845
1846             if (gmx_mm_any_lt(rsq21,rcutoff2))
1847             {
1848
1849             /* REACTION-FIELD ELECTROSTATICS */
1850             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1851
1852             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1853
1854             fscal            = felec;
1855
1856             fscal            = _mm_and_pd(fscal,cutoff_mask);
1857
1858             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1859
1860             /* Update vectorial force */
1861             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1862             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1863             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1864
1865             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1866             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1867             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1868
1869             }
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             if (gmx_mm_any_lt(rsq22,rcutoff2))
1876             {
1877
1878             /* REACTION-FIELD ELECTROSTATICS */
1879             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1880
1881             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1882
1883             fscal            = felec;
1884
1885             fscal            = _mm_and_pd(fscal,cutoff_mask);
1886
1887             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1888
1889             /* Update vectorial force */
1890             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1891             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1892             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1893
1894             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1895             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1896             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1897
1898             }
1899
1900             /**************************
1901              * CALCULATE INTERACTIONS *
1902              **************************/
1903
1904             if (gmx_mm_any_lt(rsq23,rcutoff2))
1905             {
1906
1907             /* REACTION-FIELD ELECTROSTATICS */
1908             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1909
1910             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1911
1912             fscal            = felec;
1913
1914             fscal            = _mm_and_pd(fscal,cutoff_mask);
1915
1916             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1917
1918             /* Update vectorial force */
1919             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1920             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1921             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1922
1923             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1924             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1925             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1926
1927             }
1928
1929             /**************************
1930              * CALCULATE INTERACTIONS *
1931              **************************/
1932
1933             if (gmx_mm_any_lt(rsq31,rcutoff2))
1934             {
1935
1936             /* REACTION-FIELD ELECTROSTATICS */
1937             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1938
1939             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1940
1941             fscal            = felec;
1942
1943             fscal            = _mm_and_pd(fscal,cutoff_mask);
1944
1945             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1946
1947             /* Update vectorial force */
1948             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1949             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1950             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1951
1952             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1953             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1954             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1955
1956             }
1957
1958             /**************************
1959              * CALCULATE INTERACTIONS *
1960              **************************/
1961
1962             if (gmx_mm_any_lt(rsq32,rcutoff2))
1963             {
1964
1965             /* REACTION-FIELD ELECTROSTATICS */
1966             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1967
1968             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1969
1970             fscal            = felec;
1971
1972             fscal            = _mm_and_pd(fscal,cutoff_mask);
1973
1974             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1975
1976             /* Update vectorial force */
1977             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1978             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1979             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1980
1981             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1982             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1983             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1984
1985             }
1986
1987             /**************************
1988              * CALCULATE INTERACTIONS *
1989              **************************/
1990
1991             if (gmx_mm_any_lt(rsq33,rcutoff2))
1992             {
1993
1994             /* REACTION-FIELD ELECTROSTATICS */
1995             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1996
1997             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1998
1999             fscal            = felec;
2000
2001             fscal            = _mm_and_pd(fscal,cutoff_mask);
2002
2003             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2004
2005             /* Update vectorial force */
2006             fix3             = _mm_macc_pd(dx33,fscal,fix3);
2007             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
2008             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
2009
2010             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
2011             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
2012             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
2013
2014             }
2015
2016             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2017
2018             /* Inner loop uses 333 flops */
2019         }
2020
2021         /* End of innermost loop */
2022
2023         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2024                                               f+i_coord_offset,fshift+i_shift_offset);
2025
2026         /* Increment number of inner iterations */
2027         inneriter                  += j_index_end - j_index_start;
2028
2029         /* Outer loop uses 24 flops */
2030     }
2031
2032     /* Increment number of outer iterations */
2033     outeriter        += nri;
2034
2035     /* Update outer/inner flops */
2036
2037     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*333);
2038 }