src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3W3_avx_128_fma_single.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_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LJEwald
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 113     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 114     __m128           c6grid_00;
 115     __m128           c6grid_01;
 116     __m128           c6grid_02;
 117     __m128           c6grid_10;
 118     __m128           c6grid_11;
 119     __m128           c6grid_12;
 120     __m128           c6grid_20;
 121     __m128           c6grid_21;
 122     __m128           c6grid_22;
 123     real             *vdwgridparam;
 124     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 125     __m128           one_half = _mm_set1_ps(0.5);
 126     __m128           minus_one = _mm_set1_ps(-1.0);
 127     __m128i          ewitab;
 128     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 129     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 130     real             *ewtab;
 131     __m128           dummy_mask,cutoff_mask;
 132     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 133     __m128           one     = _mm_set1_ps(1.0);
 134     __m128           two     = _mm_set1_ps(2.0);
 135     x                = xx[0];
 136     f                = ff[0];
 137
 138     nri              = nlist->nri;
 139     iinr             = nlist->iinr;
 140     jindex           = nlist->jindex;
 141     jjnr             = nlist->jjnr;
 142     shiftidx         = nlist->shift;
 143     gid              = nlist->gid;
 144     shiftvec         = fr->shift_vec[0];
 145     fshift           = fr->fshift[0];
 146     facel            = _mm_set1_ps(fr->epsfac);
 147     charge           = mdatoms->chargeA;
 148     nvdwtype         = fr->ntype;
 149     vdwparam         = fr->nbfp;
 150     vdwtype          = mdatoms->typeA;
 151     vdwgridparam     = fr->ljpme_c6grid;
 152     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 153     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 154     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 155
 156     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 157     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 158     beta2            = _mm_mul_ps(beta,beta);
 159     beta3            = _mm_mul_ps(beta,beta2);
 160     ewtab            = fr->ic->tabq_coul_FDV0;
 161     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 162     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 163
 164     /* Setup water-specific parameters */
 165     inr              = nlist->iinr[0];
 166     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 167     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 168     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 169     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 170
 171     jq0              = _mm_set1_ps(charge[inr+0]);
 172     jq1              = _mm_set1_ps(charge[inr+1]);
 173     jq2              = _mm_set1_ps(charge[inr+2]);
 174     vdwjidx0A        = 2*vdwtype[inr+0];
 175     qq00             = _mm_mul_ps(iq0,jq0);
 176     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 177     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 178     c6grid_00        = _mm_set1_ps(vdwgridparam[vdwioffset0+vdwjidx0A]);
 179     qq01             = _mm_mul_ps(iq0,jq1);
 180     qq02             = _mm_mul_ps(iq0,jq2);
 181     qq10             = _mm_mul_ps(iq1,jq0);
 182     qq11             = _mm_mul_ps(iq1,jq1);
 183     qq12             = _mm_mul_ps(iq1,jq2);
 184     qq20             = _mm_mul_ps(iq2,jq0);
 185     qq21             = _mm_mul_ps(iq2,jq1);
 186     qq22             = _mm_mul_ps(iq2,jq2);
 187
 188     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 189     rcutoff_scalar   = fr->rcoulomb;
 190     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 191     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 192
 193     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 194     rvdw             = _mm_set1_ps(fr->rvdw);
 195
 196     /* Avoid stupid compiler warnings */
 197     jnrA = jnrB = jnrC = jnrD = 0;
 198     j_coord_offsetA = 0;
 199     j_coord_offsetB = 0;
 200     j_coord_offsetC = 0;
 201     j_coord_offsetD = 0;
 202
 203     outeriter        = 0;
 204     inneriter        = 0;
 205
 206     for(iidx=0;iidx<4*DIM;iidx++)
 207     {
 208         scratch[iidx] = 0.0;
 209     }
 210
 211     /* Start outer loop over neighborlists */
 212     for(iidx=0; iidx<nri; iidx++)
 213     {
 214         /* Load shift vector for this list */
 215         i_shift_offset   = DIM*shiftidx[iidx];
 216
 217         /* Load limits for loop over neighbors */
 218         j_index_start    = jindex[iidx];
 219         j_index_end      = jindex[iidx+1];
 220
 221         /* Get outer coordinate index */
 222         inr              = iinr[iidx];
 223         i_coord_offset   = DIM*inr;
 224
 225         /* Load i particle coords and add shift vector */
 226         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 227                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 228
 229         fix0             = _mm_setzero_ps();
 230         fiy0             = _mm_setzero_ps();
 231         fiz0             = _mm_setzero_ps();
 232         fix1             = _mm_setzero_ps();
 233         fiy1             = _mm_setzero_ps();
 234         fiz1             = _mm_setzero_ps();
 235         fix2             = _mm_setzero_ps();
 236         fiy2             = _mm_setzero_ps();
 237         fiz2             = _mm_setzero_ps();
 238
 239         /* Reset potential sums */
 240         velecsum         = _mm_setzero_ps();
 241         vvdwsum          = _mm_setzero_ps();
 242
 243         /* Start inner kernel loop */
 244         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 245         {
 246
 247             /* Get j neighbor index, and coordinate index */
 248             jnrA             = jjnr[jidx];
 249             jnrB             = jjnr[jidx+1];
 250             jnrC             = jjnr[jidx+2];
 251             jnrD             = jjnr[jidx+3];
 252             j_coord_offsetA  = DIM*jnrA;
 253             j_coord_offsetB  = DIM*jnrB;
 254             j_coord_offsetC  = DIM*jnrC;
 255             j_coord_offsetD  = DIM*jnrD;
 256
 257             /* load j atom coordinates */
 258             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 259                                               x+j_coord_offsetC,x+j_coord_offsetD,
 260                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 261
 262             /* Calculate displacement vector */
 263             dx00             = _mm_sub_ps(ix0,jx0);
 264             dy00             = _mm_sub_ps(iy0,jy0);
 265             dz00             = _mm_sub_ps(iz0,jz0);
 266             dx01             = _mm_sub_ps(ix0,jx1);
 267             dy01             = _mm_sub_ps(iy0,jy1);
 268             dz01             = _mm_sub_ps(iz0,jz1);
 269             dx02             = _mm_sub_ps(ix0,jx2);
 270             dy02             = _mm_sub_ps(iy0,jy2);
 271             dz02             = _mm_sub_ps(iz0,jz2);
 272             dx10             = _mm_sub_ps(ix1,jx0);
 273             dy10             = _mm_sub_ps(iy1,jy0);
 274             dz10             = _mm_sub_ps(iz1,jz0);
 275             dx11             = _mm_sub_ps(ix1,jx1);
 276             dy11             = _mm_sub_ps(iy1,jy1);
 277             dz11             = _mm_sub_ps(iz1,jz1);
 278             dx12             = _mm_sub_ps(ix1,jx2);
 279             dy12             = _mm_sub_ps(iy1,jy2);
 280             dz12             = _mm_sub_ps(iz1,jz2);
 281             dx20             = _mm_sub_ps(ix2,jx0);
 282             dy20             = _mm_sub_ps(iy2,jy0);
 283             dz20             = _mm_sub_ps(iz2,jz0);
 284             dx21             = _mm_sub_ps(ix2,jx1);
 285             dy21             = _mm_sub_ps(iy2,jy1);
 286             dz21             = _mm_sub_ps(iz2,jz1);
 287             dx22             = _mm_sub_ps(ix2,jx2);
 288             dy22             = _mm_sub_ps(iy2,jy2);
 289             dz22             = _mm_sub_ps(iz2,jz2);
 290
 291             /* Calculate squared distance and things based on it */
 292             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 293             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 294             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 295             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 296             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 297             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 298             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 299             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 300             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 301
 302             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 303             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 304             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 305             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 306             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 307             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 308             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 309             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 310             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 311
 312             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 313             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 314             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 315             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 316             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 317             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 318             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 319             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 320             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 321
 322             fjx0             = _mm_setzero_ps();
 323             fjy0             = _mm_setzero_ps();
 324             fjz0             = _mm_setzero_ps();
 325             fjx1             = _mm_setzero_ps();
 326             fjy1             = _mm_setzero_ps();
 327             fjz1             = _mm_setzero_ps();
 328             fjx2             = _mm_setzero_ps();
 329             fjy2             = _mm_setzero_ps();
 330             fjz2             = _mm_setzero_ps();
 331
 332             /**************************
 333              * CALCULATE INTERACTIONS *
 334              **************************/
 335
 336             if (gmx_mm_any_lt(rsq00,rcutoff2))
 337             {
 338
 339             r00              = _mm_mul_ps(rsq00,rinv00);
 340
 341             /* EWALD ELECTROSTATICS */
 342
 343             /* Analytical PME correction */
 344             zeta2            = _mm_mul_ps(beta2,rsq00);
 345             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 346             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 347             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 348             felec            = _mm_mul_ps(qq00,felec);
 349             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 350             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 351             velec            = _mm_mul_ps(qq00,velec);
 352
 353             /* Analytical LJ-PME */
 354             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 355             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 356             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 357             exponent         = gmx_simd_exp_r(ewcljrsq);
 358             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 359             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 360             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 361             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 362             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 363             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 364                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 365             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 366             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 367
 368             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velec            = _mm_and_ps(velec,cutoff_mask);
 372             velecsum         = _mm_add_ps(velecsum,velec);
 373             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 374             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 375
 376             fscal            = _mm_add_ps(felec,fvdw);
 377
 378             fscal            = _mm_and_ps(fscal,cutoff_mask);
 379
 380              /* Update vectorial force */
 381             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 382             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 383             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 384
 385             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 386             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 387             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 388
 389             }
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             if (gmx_mm_any_lt(rsq01,rcutoff2))
 396             {
 397
 398             r01              = _mm_mul_ps(rsq01,rinv01);
 399
 400             /* EWALD ELECTROSTATICS */
 401
 402             /* Analytical PME correction */
 403             zeta2            = _mm_mul_ps(beta2,rsq01);
 404             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 405             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 406             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 407             felec            = _mm_mul_ps(qq01,felec);
 408             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 409             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 410             velec            = _mm_mul_ps(qq01,velec);
 411
 412             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_and_ps(velec,cutoff_mask);
 416             velecsum         = _mm_add_ps(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             fscal            = _mm_and_ps(fscal,cutoff_mask);
 421
 422              /* Update vectorial force */
 423             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 424             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 425             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 426
 427             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 428             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 429             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 430
 431             }
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             if (gmx_mm_any_lt(rsq02,rcutoff2))
 438             {
 439
 440             r02              = _mm_mul_ps(rsq02,rinv02);
 441
 442             /* EWALD ELECTROSTATICS */
 443
 444             /* Analytical PME correction */
 445             zeta2            = _mm_mul_ps(beta2,rsq02);
 446             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 447             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 448             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 449             felec            = _mm_mul_ps(qq02,felec);
 450             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 451             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 452             velec            = _mm_mul_ps(qq02,velec);
 453
 454             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velec            = _mm_and_ps(velec,cutoff_mask);
 458             velecsum         = _mm_add_ps(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             fscal            = _mm_and_ps(fscal,cutoff_mask);
 463
 464              /* Update vectorial force */
 465             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 466             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 467             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 468
 469             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 470             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 471             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 472
 473             }
 474
 475             /**************************
 476              * CALCULATE INTERACTIONS *
 477              **************************/
 478
 479             if (gmx_mm_any_lt(rsq10,rcutoff2))
 480             {
 481
 482             r10              = _mm_mul_ps(rsq10,rinv10);
 483
 484             /* EWALD ELECTROSTATICS */
 485
 486             /* Analytical PME correction */
 487             zeta2            = _mm_mul_ps(beta2,rsq10);
 488             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 489             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 490             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 491             felec            = _mm_mul_ps(qq10,felec);
 492             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 493             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 494             velec            = _mm_mul_ps(qq10,velec);
 495
 496             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 497
 498             /* Update potential sum for this i atom from the interaction with this j atom. */
 499             velec            = _mm_and_ps(velec,cutoff_mask);
 500             velecsum         = _mm_add_ps(velecsum,velec);
 501
 502             fscal            = felec;
 503
 504             fscal            = _mm_and_ps(fscal,cutoff_mask);
 505
 506              /* Update vectorial force */
 507             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 508             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 509             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 510
 511             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 512             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 513             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 514
 515             }
 516
 517             /**************************
 518              * CALCULATE INTERACTIONS *
 519              **************************/
 520
 521             if (gmx_mm_any_lt(rsq11,rcutoff2))
 522             {
 523
 524             r11              = _mm_mul_ps(rsq11,rinv11);
 525
 526             /* EWALD ELECTROSTATICS */
 527
 528             /* Analytical PME correction */
 529             zeta2            = _mm_mul_ps(beta2,rsq11);
 530             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 531             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 532             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 533             felec            = _mm_mul_ps(qq11,felec);
 534             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 535             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 536             velec            = _mm_mul_ps(qq11,velec);
 537
 538             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 539
 540             /* Update potential sum for this i atom from the interaction with this j atom. */
 541             velec            = _mm_and_ps(velec,cutoff_mask);
 542             velecsum         = _mm_add_ps(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546             fscal            = _mm_and_ps(fscal,cutoff_mask);
 547
 548              /* Update vectorial force */
 549             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 550             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 551             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 552
 553             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 554             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 555             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 556
 557             }
 558
 559             /**************************
 560              * CALCULATE INTERACTIONS *
 561              **************************/
 562
 563             if (gmx_mm_any_lt(rsq12,rcutoff2))
 564             {
 565
 566             r12              = _mm_mul_ps(rsq12,rinv12);
 567
 568             /* EWALD ELECTROSTATICS */
 569
 570             /* Analytical PME correction */
 571             zeta2            = _mm_mul_ps(beta2,rsq12);
 572             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 573             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 574             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 575             felec            = _mm_mul_ps(qq12,felec);
 576             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 577             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
 578             velec            = _mm_mul_ps(qq12,velec);
 579
 580             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 581
 582             /* Update potential sum for this i atom from the interaction with this j atom. */
 583             velec            = _mm_and_ps(velec,cutoff_mask);
 584             velecsum         = _mm_add_ps(velecsum,velec);
 585
 586             fscal            = felec;
 587
 588             fscal            = _mm_and_ps(fscal,cutoff_mask);
 589
 590              /* Update vectorial force */
 591             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 592             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 593             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 594
 595             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 596             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 597             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 598
 599             }
 600
 601             /**************************
 602              * CALCULATE INTERACTIONS *
 603              **************************/
 604
 605             if (gmx_mm_any_lt(rsq20,rcutoff2))
 606             {
 607
 608             r20              = _mm_mul_ps(rsq20,rinv20);
 609
 610             /* EWALD ELECTROSTATICS */
 611
 612             /* Analytical PME correction */
 613             zeta2            = _mm_mul_ps(beta2,rsq20);
 614             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 615             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 616             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 617             felec            = _mm_mul_ps(qq20,felec);
 618             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 619             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 620             velec            = _mm_mul_ps(qq20,velec);
 621
 622             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 623
 624             /* Update potential sum for this i atom from the interaction with this j atom. */
 625             velec            = _mm_and_ps(velec,cutoff_mask);
 626             velecsum         = _mm_add_ps(velecsum,velec);
 627
 628             fscal            = felec;
 629
 630             fscal            = _mm_and_ps(fscal,cutoff_mask);
 631
 632              /* Update vectorial force */
 633             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 634             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 635             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 636
 637             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 638             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 639             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 640
 641             }
 642
 643             /**************************
 644              * CALCULATE INTERACTIONS *
 645              **************************/
 646
 647             if (gmx_mm_any_lt(rsq21,rcutoff2))
 648             {
 649
 650             r21              = _mm_mul_ps(rsq21,rinv21);
 651
 652             /* EWALD ELECTROSTATICS */
 653
 654             /* Analytical PME correction */
 655             zeta2            = _mm_mul_ps(beta2,rsq21);
 656             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 657             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 658             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 659             felec            = _mm_mul_ps(qq21,felec);
 660             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 661             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
 662             velec            = _mm_mul_ps(qq21,velec);
 663
 664             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _mm_and_ps(velec,cutoff_mask);
 668             velecsum         = _mm_add_ps(velecsum,velec);
 669
 670             fscal            = felec;
 671
 672             fscal            = _mm_and_ps(fscal,cutoff_mask);
 673
 674              /* Update vectorial force */
 675             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 676             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 677             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 678
 679             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 680             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 681             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 682
 683             }
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             if (gmx_mm_any_lt(rsq22,rcutoff2))
 690             {
 691
 692             r22              = _mm_mul_ps(rsq22,rinv22);
 693
 694             /* EWALD ELECTROSTATICS */
 695
 696             /* Analytical PME correction */
 697             zeta2            = _mm_mul_ps(beta2,rsq22);
 698             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 699             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 700             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 701             felec            = _mm_mul_ps(qq22,felec);
 702             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 703             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
 704             velec            = _mm_mul_ps(qq22,velec);
 705
 706             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 707
 708             /* Update potential sum for this i atom from the interaction with this j atom. */
 709             velec            = _mm_and_ps(velec,cutoff_mask);
 710             velecsum         = _mm_add_ps(velecsum,velec);
 711
 712             fscal            = felec;
 713
 714             fscal            = _mm_and_ps(fscal,cutoff_mask);
 715
 716              /* Update vectorial force */
 717             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 718             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 719             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 720
 721             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 722             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 723             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 724
 725             }
 726
 727             fjptrA             = f+j_coord_offsetA;
 728             fjptrB             = f+j_coord_offsetB;
 729             fjptrC             = f+j_coord_offsetC;
 730             fjptrD             = f+j_coord_offsetD;
 731
 732             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 733                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 734
 735             /* Inner loop uses 327 flops */
 736         }
 737
 738         if(jidx<j_index_end)
 739         {
 740
 741             /* Get j neighbor index, and coordinate index */
 742             jnrlistA         = jjnr[jidx];
 743             jnrlistB         = jjnr[jidx+1];
 744             jnrlistC         = jjnr[jidx+2];
 745             jnrlistD         = jjnr[jidx+3];
 746             /* Sign of each element will be negative for non-real atoms.
 747              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 748              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 749              */
 750             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 751             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 752             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 753             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 754             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 755             j_coord_offsetA  = DIM*jnrA;
 756             j_coord_offsetB  = DIM*jnrB;
 757             j_coord_offsetC  = DIM*jnrC;
 758             j_coord_offsetD  = DIM*jnrD;
 759
 760             /* load j atom coordinates */
 761             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 762                                               x+j_coord_offsetC,x+j_coord_offsetD,
 763                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 764
 765             /* Calculate displacement vector */
 766             dx00             = _mm_sub_ps(ix0,jx0);
 767             dy00             = _mm_sub_ps(iy0,jy0);
 768             dz00             = _mm_sub_ps(iz0,jz0);
 769             dx01             = _mm_sub_ps(ix0,jx1);
 770             dy01             = _mm_sub_ps(iy0,jy1);
 771             dz01             = _mm_sub_ps(iz0,jz1);
 772             dx02             = _mm_sub_ps(ix0,jx2);
 773             dy02             = _mm_sub_ps(iy0,jy2);
 774             dz02             = _mm_sub_ps(iz0,jz2);
 775             dx10             = _mm_sub_ps(ix1,jx0);
 776             dy10             = _mm_sub_ps(iy1,jy0);
 777             dz10             = _mm_sub_ps(iz1,jz0);
 778             dx11             = _mm_sub_ps(ix1,jx1);
 779             dy11             = _mm_sub_ps(iy1,jy1);
 780             dz11             = _mm_sub_ps(iz1,jz1);
 781             dx12             = _mm_sub_ps(ix1,jx2);
 782             dy12             = _mm_sub_ps(iy1,jy2);
 783             dz12             = _mm_sub_ps(iz1,jz2);
 784             dx20             = _mm_sub_ps(ix2,jx0);
 785             dy20             = _mm_sub_ps(iy2,jy0);
 786             dz20             = _mm_sub_ps(iz2,jz0);
 787             dx21             = _mm_sub_ps(ix2,jx1);
 788             dy21             = _mm_sub_ps(iy2,jy1);
 789             dz21             = _mm_sub_ps(iz2,jz1);
 790             dx22             = _mm_sub_ps(ix2,jx2);
 791             dy22             = _mm_sub_ps(iy2,jy2);
 792             dz22             = _mm_sub_ps(iz2,jz2);
 793
 794             /* Calculate squared distance and things based on it */
 795             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 796             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 797             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 798             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 799             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 800             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 801             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 802             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 803             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 804
 805             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 806             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 807             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 808             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 809             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 810             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 811             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 812             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 813             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 814
 815             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 816             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 817             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 818             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 819             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 820             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 821             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 822             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 823             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 824
 825             fjx0             = _mm_setzero_ps();
 826             fjy0             = _mm_setzero_ps();
 827             fjz0             = _mm_setzero_ps();
 828             fjx1             = _mm_setzero_ps();
 829             fjy1             = _mm_setzero_ps();
 830             fjz1             = _mm_setzero_ps();
 831             fjx2             = _mm_setzero_ps();
 832             fjy2             = _mm_setzero_ps();
 833             fjz2             = _mm_setzero_ps();
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             if (gmx_mm_any_lt(rsq00,rcutoff2))
 840             {
 841
 842             r00              = _mm_mul_ps(rsq00,rinv00);
 843             r00              = _mm_andnot_ps(dummy_mask,r00);
 844
 845             /* EWALD ELECTROSTATICS */
 846
 847             /* Analytical PME correction */
 848             zeta2            = _mm_mul_ps(beta2,rsq00);
 849             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 850             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 851             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 852             felec            = _mm_mul_ps(qq00,felec);
 853             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 854             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 855             velec            = _mm_mul_ps(qq00,velec);
 856
 857             /* Analytical LJ-PME */
 858             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 859             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 860             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 861             exponent         = gmx_simd_exp_r(ewcljrsq);
 862             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 863             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 864             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 865             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 866             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 867             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 868                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 869             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 870             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 871
 872             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 873
 874             /* Update potential sum for this i atom from the interaction with this j atom. */
 875             velec            = _mm_and_ps(velec,cutoff_mask);
 876             velec            = _mm_andnot_ps(dummy_mask,velec);
 877             velecsum         = _mm_add_ps(velecsum,velec);
 878             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 879             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 880             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 881
 882             fscal            = _mm_add_ps(felec,fvdw);
 883
 884             fscal            = _mm_and_ps(fscal,cutoff_mask);
 885
 886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 887
 888              /* Update vectorial force */
 889             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 890             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 891             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 892
 893             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 894             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 895             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 896
 897             }
 898
 899             /**************************
 900              * CALCULATE INTERACTIONS *
 901              **************************/
 902
 903             if (gmx_mm_any_lt(rsq01,rcutoff2))
 904             {
 905
 906             r01              = _mm_mul_ps(rsq01,rinv01);
 907             r01              = _mm_andnot_ps(dummy_mask,r01);
 908
 909             /* EWALD ELECTROSTATICS */
 910
 911             /* Analytical PME correction */
 912             zeta2            = _mm_mul_ps(beta2,rsq01);
 913             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 914             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 915             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 916             felec            = _mm_mul_ps(qq01,felec);
 917             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 918             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 919             velec            = _mm_mul_ps(qq01,velec);
 920
 921             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_and_ps(velec,cutoff_mask);
 925             velec            = _mm_andnot_ps(dummy_mask,velec);
 926             velecsum         = _mm_add_ps(velecsum,velec);
 927
 928             fscal            = felec;
 929
 930             fscal            = _mm_and_ps(fscal,cutoff_mask);
 931
 932             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 933
 934              /* Update vectorial force */
 935             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 936             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 937             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 938
 939             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 940             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 941             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 942
 943             }
 944
 945             /**************************
 946              * CALCULATE INTERACTIONS *
 947              **************************/
 948
 949             if (gmx_mm_any_lt(rsq02,rcutoff2))
 950             {
 951
 952             r02              = _mm_mul_ps(rsq02,rinv02);
 953             r02              = _mm_andnot_ps(dummy_mask,r02);
 954
 955             /* EWALD ELECTROSTATICS */
 956
 957             /* Analytical PME correction */
 958             zeta2            = _mm_mul_ps(beta2,rsq02);
 959             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 960             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 961             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 962             felec            = _mm_mul_ps(qq02,felec);
 963             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 964             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 965             velec            = _mm_mul_ps(qq02,velec);
 966
 967             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 968
 969             /* Update potential sum for this i atom from the interaction with this j atom. */
 970             velec            = _mm_and_ps(velec,cutoff_mask);
 971             velec            = _mm_andnot_ps(dummy_mask,velec);
 972             velecsum         = _mm_add_ps(velecsum,velec);
 973
 974             fscal            = felec;
 975
 976             fscal            = _mm_and_ps(fscal,cutoff_mask);
 977
 978             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 979
 980              /* Update vectorial force */
 981             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 982             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 983             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 984
 985             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 986             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 987             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 988
 989             }
 990
 991             /**************************
 992              * CALCULATE INTERACTIONS *
 993              **************************/
 994
 995             if (gmx_mm_any_lt(rsq10,rcutoff2))
 996             {
 997
 998             r10              = _mm_mul_ps(rsq10,rinv10);
 999             r10              = _mm_andnot_ps(dummy_mask,r10);
1000
1001             /* EWALD ELECTROSTATICS */
1002
1003             /* Analytical PME correction */
1004             zeta2            = _mm_mul_ps(beta2,rsq10);
1005             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1006             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1007             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1008             felec            = _mm_mul_ps(qq10,felec);
1009             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1010             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
1011             velec            = _mm_mul_ps(qq10,velec);
1012
1013             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1014
1015             /* Update potential sum for this i atom from the interaction with this j atom. */
1016             velec            = _mm_and_ps(velec,cutoff_mask);
1017             velec            = _mm_andnot_ps(dummy_mask,velec);
1018             velecsum         = _mm_add_ps(velecsum,velec);
1019
1020             fscal            = felec;
1021
1022             fscal            = _mm_and_ps(fscal,cutoff_mask);
1023
1024             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1025
1026              /* Update vectorial force */
1027             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1028             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1029             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1030
1031             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1032             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1033             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1034
1035             }
1036
1037             /**************************
1038              * CALCULATE INTERACTIONS *
1039              **************************/
1040
1041             if (gmx_mm_any_lt(rsq11,rcutoff2))
1042             {
1043
1044             r11              = _mm_mul_ps(rsq11,rinv11);
1045             r11              = _mm_andnot_ps(dummy_mask,r11);
1046
1047             /* EWALD ELECTROSTATICS */
1048
1049             /* Analytical PME correction */
1050             zeta2            = _mm_mul_ps(beta2,rsq11);
1051             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1052             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1053             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1054             felec            = _mm_mul_ps(qq11,felec);
1055             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1056             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
1057             velec            = _mm_mul_ps(qq11,velec);
1058
1059             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1060
1061             /* Update potential sum for this i atom from the interaction with this j atom. */
1062             velec            = _mm_and_ps(velec,cutoff_mask);
1063             velec            = _mm_andnot_ps(dummy_mask,velec);
1064             velecsum         = _mm_add_ps(velecsum,velec);
1065
1066             fscal            = felec;
1067
1068             fscal            = _mm_and_ps(fscal,cutoff_mask);
1069
1070             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1071
1072              /* Update vectorial force */
1073             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1074             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1075             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1076
1077             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1078             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1079             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1080
1081             }
1082
1083             /**************************
1084              * CALCULATE INTERACTIONS *
1085              **************************/
1086
1087             if (gmx_mm_any_lt(rsq12,rcutoff2))
1088             {
1089
1090             r12              = _mm_mul_ps(rsq12,rinv12);
1091             r12              = _mm_andnot_ps(dummy_mask,r12);
1092
1093             /* EWALD ELECTROSTATICS */
1094
1095             /* Analytical PME correction */
1096             zeta2            = _mm_mul_ps(beta2,rsq12);
1097             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1098             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1099             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1100             felec            = _mm_mul_ps(qq12,felec);
1101             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1102             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
1103             velec            = _mm_mul_ps(qq12,velec);
1104
1105             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1106
1107             /* Update potential sum for this i atom from the interaction with this j atom. */
1108             velec            = _mm_and_ps(velec,cutoff_mask);
1109             velec            = _mm_andnot_ps(dummy_mask,velec);
1110             velecsum         = _mm_add_ps(velecsum,velec);
1111
1112             fscal            = felec;
1113
1114             fscal            = _mm_and_ps(fscal,cutoff_mask);
1115
1116             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1117
1118              /* Update vectorial force */
1119             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1120             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1121             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1122
1123             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1124             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1125             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1126
1127             }
1128
1129             /**************************
1130              * CALCULATE INTERACTIONS *
1131              **************************/
1132
1133             if (gmx_mm_any_lt(rsq20,rcutoff2))
1134             {
1135
1136             r20              = _mm_mul_ps(rsq20,rinv20);
1137             r20              = _mm_andnot_ps(dummy_mask,r20);
1138
1139             /* EWALD ELECTROSTATICS */
1140
1141             /* Analytical PME correction */
1142             zeta2            = _mm_mul_ps(beta2,rsq20);
1143             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1144             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1145             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1146             felec            = _mm_mul_ps(qq20,felec);
1147             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1148             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
1149             velec            = _mm_mul_ps(qq20,velec);
1150
1151             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1152
1153             /* Update potential sum for this i atom from the interaction with this j atom. */
1154             velec            = _mm_and_ps(velec,cutoff_mask);
1155             velec            = _mm_andnot_ps(dummy_mask,velec);
1156             velecsum         = _mm_add_ps(velecsum,velec);
1157
1158             fscal            = felec;
1159
1160             fscal            = _mm_and_ps(fscal,cutoff_mask);
1161
1162             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1163
1164              /* Update vectorial force */
1165             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1166             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1167             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1168
1169             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1170             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1171             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1172
1173             }
1174
1175             /**************************
1176              * CALCULATE INTERACTIONS *
1177              **************************/
1178
1179             if (gmx_mm_any_lt(rsq21,rcutoff2))
1180             {
1181
1182             r21              = _mm_mul_ps(rsq21,rinv21);
1183             r21              = _mm_andnot_ps(dummy_mask,r21);
1184
1185             /* EWALD ELECTROSTATICS */
1186
1187             /* Analytical PME correction */
1188             zeta2            = _mm_mul_ps(beta2,rsq21);
1189             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1190             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1191             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1192             felec            = _mm_mul_ps(qq21,felec);
1193             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1194             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
1195             velec            = _mm_mul_ps(qq21,velec);
1196
1197             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1198
1199             /* Update potential sum for this i atom from the interaction with this j atom. */
1200             velec            = _mm_and_ps(velec,cutoff_mask);
1201             velec            = _mm_andnot_ps(dummy_mask,velec);
1202             velecsum         = _mm_add_ps(velecsum,velec);
1203
1204             fscal            = felec;
1205
1206             fscal            = _mm_and_ps(fscal,cutoff_mask);
1207
1208             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1209
1210              /* Update vectorial force */
1211             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1212             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1213             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1214
1215             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1216             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1217             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1218
1219             }
1220
1221             /**************************
1222              * CALCULATE INTERACTIONS *
1223              **************************/
1224
1225             if (gmx_mm_any_lt(rsq22,rcutoff2))
1226             {
1227
1228             r22              = _mm_mul_ps(rsq22,rinv22);
1229             r22              = _mm_andnot_ps(dummy_mask,r22);
1230
1231             /* EWALD ELECTROSTATICS */
1232
1233             /* Analytical PME correction */
1234             zeta2            = _mm_mul_ps(beta2,rsq22);
1235             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1236             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1237             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1238             felec            = _mm_mul_ps(qq22,felec);
1239             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1240             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
1241             velec            = _mm_mul_ps(qq22,velec);
1242
1243             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1244
1245             /* Update potential sum for this i atom from the interaction with this j atom. */
1246             velec            = _mm_and_ps(velec,cutoff_mask);
1247             velec            = _mm_andnot_ps(dummy_mask,velec);
1248             velecsum         = _mm_add_ps(velecsum,velec);
1249
1250             fscal            = felec;
1251
1252             fscal            = _mm_and_ps(fscal,cutoff_mask);
1253
1254             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1255
1256              /* Update vectorial force */
1257             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1258             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1259             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1260
1261             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1262             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1263             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1264
1265             }
1266
1267             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1268             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1269             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1270             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1271
1272             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1273                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1274
1275             /* Inner loop uses 336 flops */
1276         }
1277
1278         /* End of innermost loop */
1279
1280         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1281                                               f+i_coord_offset,fshift+i_shift_offset);
1282
1283         ggid                        = gid[iidx];
1284         /* Update potential energies */
1285         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1286         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1287
1288         /* Increment number of inner iterations */
1289         inneriter                  += j_index_end - j_index_start;
1290
1291         /* Outer loop uses 20 flops */
1292     }
1293
1294     /* Increment number of outer iterations */
1295     outeriter        += nri;
1296
1297     /* Update outer/inner flops */
1298
1299     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*336);
1300 }
1301 /*
1302  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3W3_F_avx_128_fma_single
1303  * Electrostatics interaction: Ewald
1304  * VdW interaction:            LJEwald
1305  * Geometry:                   Water3-Water3
1306  * Calculate force/pot:        Force
1307  */
1308 void
1309 nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3W3_F_avx_128_fma_single
1310                     (t_nblist                    * gmx_restrict       nlist,
1311                      rvec                        * gmx_restrict          xx,
1312                      rvec                        * gmx_restrict          ff,
1313                      t_forcerec                  * gmx_restrict          fr,
1314                      t_mdatoms                   * gmx_restrict     mdatoms,
1315                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1316                      t_nrnb                      * gmx_restrict        nrnb)
1317 {
1318     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1319      * just 0 for non-waters.
1320      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1321      * jnr indices corresponding to data put in the four positions in the SIMD register.
1322      */
1323     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1324     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1325     int              jnrA,jnrB,jnrC,jnrD;
1326     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1327     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1328     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1329     real             rcutoff_scalar;
1330     real             *shiftvec,*fshift,*x,*f;
1331     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1332     real             scratch[4*DIM];
1333     __m128           fscal,rcutoff,rcutoff2,jidxall;
1334     int              vdwioffset0;
1335     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1336     int              vdwioffset1;
1337     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1338     int              vdwioffset2;
1339     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1340     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1341     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1342     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1343     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1344     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1345     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1346     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1347     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1348     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1349     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1350     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1351     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1352     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1353     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1354     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1355     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1356     real             *charge;
1357     int              nvdwtype;
1358     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1359     int              *vdwtype;
1360     real             *vdwparam;
1361     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1362     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1363     __m128           c6grid_00;
1364     __m128           c6grid_01;
1365     __m128           c6grid_02;
1366     __m128           c6grid_10;
1367     __m128           c6grid_11;
1368     __m128           c6grid_12;
1369     __m128           c6grid_20;
1370     __m128           c6grid_21;
1371     __m128           c6grid_22;
1372     real             *vdwgridparam;
1373     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
1374     __m128           one_half = _mm_set1_ps(0.5);
1375     __m128           minus_one = _mm_set1_ps(-1.0);
1376     __m128i          ewitab;
1377     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1378     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1379     real             *ewtab;
1380     __m128           dummy_mask,cutoff_mask;
1381     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1382     __m128           one     = _mm_set1_ps(1.0);
1383     __m128           two     = _mm_set1_ps(2.0);
1384     x                = xx[0];
1385     f                = ff[0];
1386
1387     nri              = nlist->nri;
1388     iinr             = nlist->iinr;
1389     jindex           = nlist->jindex;
1390     jjnr             = nlist->jjnr;
1391     shiftidx         = nlist->shift;
1392     gid              = nlist->gid;
1393     shiftvec         = fr->shift_vec[0];
1394     fshift           = fr->fshift[0];
1395     facel            = _mm_set1_ps(fr->epsfac);
1396     charge           = mdatoms->chargeA;
1397     nvdwtype         = fr->ntype;
1398     vdwparam         = fr->nbfp;
1399     vdwtype          = mdatoms->typeA;
1400     vdwgridparam     = fr->ljpme_c6grid;
1401     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
1402     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
1403     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
1404
1405     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1406     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1407     beta2            = _mm_mul_ps(beta,beta);
1408     beta3            = _mm_mul_ps(beta,beta2);
1409     ewtab            = fr->ic->tabq_coul_F;
1410     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1411     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1412
1413     /* Setup water-specific parameters */
1414     inr              = nlist->iinr[0];
1415     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1416     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1417     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1418     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1419
1420     jq0              = _mm_set1_ps(charge[inr+0]);
1421     jq1              = _mm_set1_ps(charge[inr+1]);
1422     jq2              = _mm_set1_ps(charge[inr+2]);
1423     vdwjidx0A        = 2*vdwtype[inr+0];
1424     qq00             = _mm_mul_ps(iq0,jq0);
1425     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1426     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1427     c6grid_00        = _mm_set1_ps(vdwgridparam[vdwioffset0+vdwjidx0A]);
1428     qq01             = _mm_mul_ps(iq0,jq1);
1429     qq02             = _mm_mul_ps(iq0,jq2);
1430     qq10             = _mm_mul_ps(iq1,jq0);
1431     qq11             = _mm_mul_ps(iq1,jq1);
1432     qq12             = _mm_mul_ps(iq1,jq2);
1433     qq20             = _mm_mul_ps(iq2,jq0);
1434     qq21             = _mm_mul_ps(iq2,jq1);
1435     qq22             = _mm_mul_ps(iq2,jq2);
1436
1437     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1438     rcutoff_scalar   = fr->rcoulomb;
1439     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1440     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1441
1442     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1443     rvdw             = _mm_set1_ps(fr->rvdw);
1444
1445     /* Avoid stupid compiler warnings */
1446     jnrA = jnrB = jnrC = jnrD = 0;
1447     j_coord_offsetA = 0;
1448     j_coord_offsetB = 0;
1449     j_coord_offsetC = 0;
1450     j_coord_offsetD = 0;
1451
1452     outeriter        = 0;
1453     inneriter        = 0;
1454
1455     for(iidx=0;iidx<4*DIM;iidx++)
1456     {
1457         scratch[iidx] = 0.0;
1458     }
1459
1460     /* Start outer loop over neighborlists */
1461     for(iidx=0; iidx<nri; iidx++)
1462     {
1463         /* Load shift vector for this list */
1464         i_shift_offset   = DIM*shiftidx[iidx];
1465
1466         /* Load limits for loop over neighbors */
1467         j_index_start    = jindex[iidx];
1468         j_index_end      = jindex[iidx+1];
1469
1470         /* Get outer coordinate index */
1471         inr              = iinr[iidx];
1472         i_coord_offset   = DIM*inr;
1473
1474         /* Load i particle coords and add shift vector */
1475         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1476                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1477
1478         fix0             = _mm_setzero_ps();
1479         fiy0             = _mm_setzero_ps();
1480         fiz0             = _mm_setzero_ps();
1481         fix1             = _mm_setzero_ps();
1482         fiy1             = _mm_setzero_ps();
1483         fiz1             = _mm_setzero_ps();
1484         fix2             = _mm_setzero_ps();
1485         fiy2             = _mm_setzero_ps();
1486         fiz2             = _mm_setzero_ps();
1487
1488         /* Start inner kernel loop */
1489         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1490         {
1491
1492             /* Get j neighbor index, and coordinate index */
1493             jnrA             = jjnr[jidx];
1494             jnrB             = jjnr[jidx+1];
1495             jnrC             = jjnr[jidx+2];
1496             jnrD             = jjnr[jidx+3];
1497             j_coord_offsetA  = DIM*jnrA;
1498             j_coord_offsetB  = DIM*jnrB;
1499             j_coord_offsetC  = DIM*jnrC;
1500             j_coord_offsetD  = DIM*jnrD;
1501
1502             /* load j atom coordinates */
1503             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1504                                               x+j_coord_offsetC,x+j_coord_offsetD,
1505                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1506
1507             /* Calculate displacement vector */
1508             dx00             = _mm_sub_ps(ix0,jx0);
1509             dy00             = _mm_sub_ps(iy0,jy0);
1510             dz00             = _mm_sub_ps(iz0,jz0);
1511             dx01             = _mm_sub_ps(ix0,jx1);
1512             dy01             = _mm_sub_ps(iy0,jy1);
1513             dz01             = _mm_sub_ps(iz0,jz1);
1514             dx02             = _mm_sub_ps(ix0,jx2);
1515             dy02             = _mm_sub_ps(iy0,jy2);
1516             dz02             = _mm_sub_ps(iz0,jz2);
1517             dx10             = _mm_sub_ps(ix1,jx0);
1518             dy10             = _mm_sub_ps(iy1,jy0);
1519             dz10             = _mm_sub_ps(iz1,jz0);
1520             dx11             = _mm_sub_ps(ix1,jx1);
1521             dy11             = _mm_sub_ps(iy1,jy1);
1522             dz11             = _mm_sub_ps(iz1,jz1);
1523             dx12             = _mm_sub_ps(ix1,jx2);
1524             dy12             = _mm_sub_ps(iy1,jy2);
1525             dz12             = _mm_sub_ps(iz1,jz2);
1526             dx20             = _mm_sub_ps(ix2,jx0);
1527             dy20             = _mm_sub_ps(iy2,jy0);
1528             dz20             = _mm_sub_ps(iz2,jz0);
1529             dx21             = _mm_sub_ps(ix2,jx1);
1530             dy21             = _mm_sub_ps(iy2,jy1);
1531             dz21             = _mm_sub_ps(iz2,jz1);
1532             dx22             = _mm_sub_ps(ix2,jx2);
1533             dy22             = _mm_sub_ps(iy2,jy2);
1534             dz22             = _mm_sub_ps(iz2,jz2);
1535
1536             /* Calculate squared distance and things based on it */
1537             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1538             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1539             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1540             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1541             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1542             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1543             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1544             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1545             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1546
1547             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1548             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1549             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1550             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1551             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1552             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1553             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1554             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1555             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1556
1557             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1558             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1559             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1560             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1561             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1562             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1563             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1564             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1565             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1566
1567             fjx0             = _mm_setzero_ps();
1568             fjy0             = _mm_setzero_ps();
1569             fjz0             = _mm_setzero_ps();
1570             fjx1             = _mm_setzero_ps();
1571             fjy1             = _mm_setzero_ps();
1572             fjz1             = _mm_setzero_ps();
1573             fjx2             = _mm_setzero_ps();
1574             fjy2             = _mm_setzero_ps();
1575             fjz2             = _mm_setzero_ps();
1576
1577             /**************************
1578              * CALCULATE INTERACTIONS *
1579              **************************/
1580
1581             if (gmx_mm_any_lt(rsq00,rcutoff2))
1582             {
1583
1584             r00              = _mm_mul_ps(rsq00,rinv00);
1585
1586             /* EWALD ELECTROSTATICS */
1587
1588             /* Analytical PME correction */
1589             zeta2            = _mm_mul_ps(beta2,rsq00);
1590             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1591             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1592             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1593             felec            = _mm_mul_ps(qq00,felec);
1594
1595             /* Analytical LJ-PME */
1596             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1597             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
1598             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
1599             exponent         = gmx_simd_exp_r(ewcljrsq);
1600             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1601             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
1602             /* f6A = 6 * C6grid * (1 - poly) */
1603             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
1604             /* f6B = C6grid * exponent * beta^6 */
1605             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
1606             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1607             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1608
1609             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1610
1611             fscal            = _mm_add_ps(felec,fvdw);
1612
1613             fscal            = _mm_and_ps(fscal,cutoff_mask);
1614
1615              /* Update vectorial force */
1616             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1617             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1618             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1619
1620             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1621             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1622             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1623
1624             }
1625
1626             /**************************
1627              * CALCULATE INTERACTIONS *
1628              **************************/
1629
1630             if (gmx_mm_any_lt(rsq01,rcutoff2))
1631             {
1632
1633             r01              = _mm_mul_ps(rsq01,rinv01);
1634
1635             /* EWALD ELECTROSTATICS */
1636
1637             /* Analytical PME correction */
1638             zeta2            = _mm_mul_ps(beta2,rsq01);
1639             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1640             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1641             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1642             felec            = _mm_mul_ps(qq01,felec);
1643
1644             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1645
1646             fscal            = felec;
1647
1648             fscal            = _mm_and_ps(fscal,cutoff_mask);
1649
1650              /* Update vectorial force */
1651             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1652             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1653             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1654
1655             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1656             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1657             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1658
1659             }
1660
1661             /**************************
1662              * CALCULATE INTERACTIONS *
1663              **************************/
1664
1665             if (gmx_mm_any_lt(rsq02,rcutoff2))
1666             {
1667
1668             r02              = _mm_mul_ps(rsq02,rinv02);
1669
1670             /* EWALD ELECTROSTATICS */
1671
1672             /* Analytical PME correction */
1673             zeta2            = _mm_mul_ps(beta2,rsq02);
1674             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1675             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1676             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1677             felec            = _mm_mul_ps(qq02,felec);
1678
1679             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1680
1681             fscal            = felec;
1682
1683             fscal            = _mm_and_ps(fscal,cutoff_mask);
1684
1685              /* Update vectorial force */
1686             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1687             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1688             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1689
1690             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1691             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1692             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1693
1694             }
1695
1696             /**************************
1697              * CALCULATE INTERACTIONS *
1698              **************************/
1699
1700             if (gmx_mm_any_lt(rsq10,rcutoff2))
1701             {
1702
1703             r10              = _mm_mul_ps(rsq10,rinv10);
1704
1705             /* EWALD ELECTROSTATICS */
1706
1707             /* Analytical PME correction */
1708             zeta2            = _mm_mul_ps(beta2,rsq10);
1709             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1710             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1711             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1712             felec            = _mm_mul_ps(qq10,felec);
1713
1714             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1715
1716             fscal            = felec;
1717
1718             fscal            = _mm_and_ps(fscal,cutoff_mask);
1719
1720              /* Update vectorial force */
1721             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1722             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1723             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1724
1725             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1726             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1727             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1728
1729             }
1730
1731             /**************************
1732              * CALCULATE INTERACTIONS *
1733              **************************/
1734
1735             if (gmx_mm_any_lt(rsq11,rcutoff2))
1736             {
1737
1738             r11              = _mm_mul_ps(rsq11,rinv11);
1739
1740             /* EWALD ELECTROSTATICS */
1741
1742             /* Analytical PME correction */
1743             zeta2            = _mm_mul_ps(beta2,rsq11);
1744             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1745             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1746             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1747             felec            = _mm_mul_ps(qq11,felec);
1748
1749             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1750
1751             fscal            = felec;
1752
1753             fscal            = _mm_and_ps(fscal,cutoff_mask);
1754
1755              /* Update vectorial force */
1756             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1757             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1758             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1759
1760             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1761             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1762             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1763
1764             }
1765
1766             /**************************
1767              * CALCULATE INTERACTIONS *
1768              **************************/
1769
1770             if (gmx_mm_any_lt(rsq12,rcutoff2))
1771             {
1772
1773             r12              = _mm_mul_ps(rsq12,rinv12);
1774
1775             /* EWALD ELECTROSTATICS */
1776
1777             /* Analytical PME correction */
1778             zeta2            = _mm_mul_ps(beta2,rsq12);
1779             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1780             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1781             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1782             felec            = _mm_mul_ps(qq12,felec);
1783
1784             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1785
1786             fscal            = felec;
1787
1788             fscal            = _mm_and_ps(fscal,cutoff_mask);
1789
1790              /* Update vectorial force */
1791             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1792             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1793             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1794
1795             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1796             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1797             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1798
1799             }
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             if (gmx_mm_any_lt(rsq20,rcutoff2))
1806             {
1807
1808             r20              = _mm_mul_ps(rsq20,rinv20);
1809
1810             /* EWALD ELECTROSTATICS */
1811
1812             /* Analytical PME correction */
1813             zeta2            = _mm_mul_ps(beta2,rsq20);
1814             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1815             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1816             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1817             felec            = _mm_mul_ps(qq20,felec);
1818
1819             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1820
1821             fscal            = felec;
1822
1823             fscal            = _mm_and_ps(fscal,cutoff_mask);
1824
1825              /* Update vectorial force */
1826             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1827             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1828             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1829
1830             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1831             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1832             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1833
1834             }
1835
1836             /**************************
1837              * CALCULATE INTERACTIONS *
1838              **************************/
1839
1840             if (gmx_mm_any_lt(rsq21,rcutoff2))
1841             {
1842
1843             r21              = _mm_mul_ps(rsq21,rinv21);
1844
1845             /* EWALD ELECTROSTATICS */
1846
1847             /* Analytical PME correction */
1848             zeta2            = _mm_mul_ps(beta2,rsq21);
1849             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1850             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1851             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1852             felec            = _mm_mul_ps(qq21,felec);
1853
1854             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1855
1856             fscal            = felec;
1857
1858             fscal            = _mm_and_ps(fscal,cutoff_mask);
1859
1860              /* Update vectorial force */
1861             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1862             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1863             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1864
1865             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1866             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1867             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1868
1869             }
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             if (gmx_mm_any_lt(rsq22,rcutoff2))
1876             {
1877
1878             r22              = _mm_mul_ps(rsq22,rinv22);
1879
1880             /* EWALD ELECTROSTATICS */
1881
1882             /* Analytical PME correction */
1883             zeta2            = _mm_mul_ps(beta2,rsq22);
1884             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1885             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1886             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1887             felec            = _mm_mul_ps(qq22,felec);
1888
1889             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1890
1891             fscal            = felec;
1892
1893             fscal            = _mm_and_ps(fscal,cutoff_mask);
1894
1895              /* Update vectorial force */
1896             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1897             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1898             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1899
1900             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1901             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1902             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1903
1904             }
1905
1906             fjptrA             = f+j_coord_offsetA;
1907             fjptrB             = f+j_coord_offsetB;
1908             fjptrC             = f+j_coord_offsetC;
1909             fjptrD             = f+j_coord_offsetD;
1910
1911             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1912                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1913
1914             /* Inner loop uses 300 flops */
1915         }
1916
1917         if(jidx<j_index_end)
1918         {
1919
1920             /* Get j neighbor index, and coordinate index */
1921             jnrlistA         = jjnr[jidx];
1922             jnrlistB         = jjnr[jidx+1];
1923             jnrlistC         = jjnr[jidx+2];
1924             jnrlistD         = jjnr[jidx+3];
1925             /* Sign of each element will be negative for non-real atoms.
1926              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1927              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1928              */
1929             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1930             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1931             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1932             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1933             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1934             j_coord_offsetA  = DIM*jnrA;
1935             j_coord_offsetB  = DIM*jnrB;
1936             j_coord_offsetC  = DIM*jnrC;
1937             j_coord_offsetD  = DIM*jnrD;
1938
1939             /* load j atom coordinates */
1940             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1941                                               x+j_coord_offsetC,x+j_coord_offsetD,
1942                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1943
1944             /* Calculate displacement vector */
1945             dx00             = _mm_sub_ps(ix0,jx0);
1946             dy00             = _mm_sub_ps(iy0,jy0);
1947             dz00             = _mm_sub_ps(iz0,jz0);
1948             dx01             = _mm_sub_ps(ix0,jx1);
1949             dy01             = _mm_sub_ps(iy0,jy1);
1950             dz01             = _mm_sub_ps(iz0,jz1);
1951             dx02             = _mm_sub_ps(ix0,jx2);
1952             dy02             = _mm_sub_ps(iy0,jy2);
1953             dz02             = _mm_sub_ps(iz0,jz2);
1954             dx10             = _mm_sub_ps(ix1,jx0);
1955             dy10             = _mm_sub_ps(iy1,jy0);
1956             dz10             = _mm_sub_ps(iz1,jz0);
1957             dx11             = _mm_sub_ps(ix1,jx1);
1958             dy11             = _mm_sub_ps(iy1,jy1);
1959             dz11             = _mm_sub_ps(iz1,jz1);
1960             dx12             = _mm_sub_ps(ix1,jx2);
1961             dy12             = _mm_sub_ps(iy1,jy2);
1962             dz12             = _mm_sub_ps(iz1,jz2);
1963             dx20             = _mm_sub_ps(ix2,jx0);
1964             dy20             = _mm_sub_ps(iy2,jy0);
1965             dz20             = _mm_sub_ps(iz2,jz0);
1966             dx21             = _mm_sub_ps(ix2,jx1);
1967             dy21             = _mm_sub_ps(iy2,jy1);
1968             dz21             = _mm_sub_ps(iz2,jz1);
1969             dx22             = _mm_sub_ps(ix2,jx2);
1970             dy22             = _mm_sub_ps(iy2,jy2);
1971             dz22             = _mm_sub_ps(iz2,jz2);
1972
1973             /* Calculate squared distance and things based on it */
1974             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1975             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1976             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1977             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1978             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1979             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1980             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1981             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1982             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1983
1984             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1985             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1986             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1987             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1988             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1989             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1990             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1991             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1992             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1993
1994             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1995             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1996             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1997             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1998             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1999             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
2000             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
2001             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
2002             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
2003
2004             fjx0             = _mm_setzero_ps();
2005             fjy0             = _mm_setzero_ps();
2006             fjz0             = _mm_setzero_ps();
2007             fjx1             = _mm_setzero_ps();
2008             fjy1             = _mm_setzero_ps();
2009             fjz1             = _mm_setzero_ps();
2010             fjx2             = _mm_setzero_ps();
2011             fjy2             = _mm_setzero_ps();
2012             fjz2             = _mm_setzero_ps();
2013
2014             /**************************
2015              * CALCULATE INTERACTIONS *
2016              **************************/
2017
2018             if (gmx_mm_any_lt(rsq00,rcutoff2))
2019             {
2020
2021             r00              = _mm_mul_ps(rsq00,rinv00);
2022             r00              = _mm_andnot_ps(dummy_mask,r00);
2023
2024             /* EWALD ELECTROSTATICS */
2025
2026             /* Analytical PME correction */
2027             zeta2            = _mm_mul_ps(beta2,rsq00);
2028             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
2029             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2030             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2031             felec            = _mm_mul_ps(qq00,felec);
2032
2033             /* Analytical LJ-PME */
2034             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
2035             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
2036             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
2037             exponent         = gmx_simd_exp_r(ewcljrsq);
2038             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
2039             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
2040             /* f6A = 6 * C6grid * (1 - poly) */
2041             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
2042             /* f6B = C6grid * exponent * beta^6 */
2043             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
2044             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
2045             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
2046
2047             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
2048
2049             fscal            = _mm_add_ps(felec,fvdw);
2050
2051             fscal            = _mm_and_ps(fscal,cutoff_mask);
2052
2053             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2054
2055              /* Update vectorial force */
2056             fix0             = _mm_macc_ps(dx00,fscal,fix0);
2057             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
2058             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
2059
2060             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
2061             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
2062             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
2063
2064             }
2065
2066             /**************************
2067              * CALCULATE INTERACTIONS *
2068              **************************/
2069
2070             if (gmx_mm_any_lt(rsq01,rcutoff2))
2071             {
2072
2073             r01              = _mm_mul_ps(rsq01,rinv01);
2074             r01              = _mm_andnot_ps(dummy_mask,r01);
2075
2076             /* EWALD ELECTROSTATICS */
2077
2078             /* Analytical PME correction */
2079             zeta2            = _mm_mul_ps(beta2,rsq01);
2080             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
2081             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2082             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2083             felec            = _mm_mul_ps(qq01,felec);
2084
2085             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
2086
2087             fscal            = felec;
2088
2089             fscal            = _mm_and_ps(fscal,cutoff_mask);
2090
2091             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2092
2093              /* Update vectorial force */
2094             fix0             = _mm_macc_ps(dx01,fscal,fix0);
2095             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
2096             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
2097
2098             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
2099             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
2100             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
2101
2102             }
2103
2104             /**************************
2105              * CALCULATE INTERACTIONS *
2106              **************************/
2107
2108             if (gmx_mm_any_lt(rsq02,rcutoff2))
2109             {
2110
2111             r02              = _mm_mul_ps(rsq02,rinv02);
2112             r02              = _mm_andnot_ps(dummy_mask,r02);
2113
2114             /* EWALD ELECTROSTATICS */
2115
2116             /* Analytical PME correction */
2117             zeta2            = _mm_mul_ps(beta2,rsq02);
2118             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
2119             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2120             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2121             felec            = _mm_mul_ps(qq02,felec);
2122
2123             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
2124
2125             fscal            = felec;
2126
2127             fscal            = _mm_and_ps(fscal,cutoff_mask);
2128
2129             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2130
2131              /* Update vectorial force */
2132             fix0             = _mm_macc_ps(dx02,fscal,fix0);
2133             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
2134             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
2135
2136             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
2137             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
2138             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
2139
2140             }
2141
2142             /**************************
2143              * CALCULATE INTERACTIONS *
2144              **************************/
2145
2146             if (gmx_mm_any_lt(rsq10,rcutoff2))
2147             {
2148
2149             r10              = _mm_mul_ps(rsq10,rinv10);
2150             r10              = _mm_andnot_ps(dummy_mask,r10);
2151
2152             /* EWALD ELECTROSTATICS */
2153
2154             /* Analytical PME correction */
2155             zeta2            = _mm_mul_ps(beta2,rsq10);
2156             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
2157             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2158             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2159             felec            = _mm_mul_ps(qq10,felec);
2160
2161             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
2162
2163             fscal            = felec;
2164
2165             fscal            = _mm_and_ps(fscal,cutoff_mask);
2166
2167             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2168
2169              /* Update vectorial force */
2170             fix1             = _mm_macc_ps(dx10,fscal,fix1);
2171             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
2172             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
2173
2174             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
2175             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
2176             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
2177
2178             }
2179
2180             /**************************
2181              * CALCULATE INTERACTIONS *
2182              **************************/
2183
2184             if (gmx_mm_any_lt(rsq11,rcutoff2))
2185             {
2186
2187             r11              = _mm_mul_ps(rsq11,rinv11);
2188             r11              = _mm_andnot_ps(dummy_mask,r11);
2189
2190             /* EWALD ELECTROSTATICS */
2191
2192             /* Analytical PME correction */
2193             zeta2            = _mm_mul_ps(beta2,rsq11);
2194             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
2195             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2196             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2197             felec            = _mm_mul_ps(qq11,felec);
2198
2199             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2200
2201             fscal            = felec;
2202
2203             fscal            = _mm_and_ps(fscal,cutoff_mask);
2204
2205             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2206
2207              /* Update vectorial force */
2208             fix1             = _mm_macc_ps(dx11,fscal,fix1);
2209             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
2210             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
2211
2212             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
2213             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
2214             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
2215
2216             }
2217
2218             /**************************
2219              * CALCULATE INTERACTIONS *
2220              **************************/
2221
2222             if (gmx_mm_any_lt(rsq12,rcutoff2))
2223             {
2224
2225             r12              = _mm_mul_ps(rsq12,rinv12);
2226             r12              = _mm_andnot_ps(dummy_mask,r12);
2227
2228             /* EWALD ELECTROSTATICS */
2229
2230             /* Analytical PME correction */
2231             zeta2            = _mm_mul_ps(beta2,rsq12);
2232             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
2233             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2234             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2235             felec            = _mm_mul_ps(qq12,felec);
2236
2237             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2238
2239             fscal            = felec;
2240
2241             fscal            = _mm_and_ps(fscal,cutoff_mask);
2242
2243             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2244
2245              /* Update vectorial force */
2246             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2247             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2248             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2249
2250             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2251             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2252             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2253
2254             }
2255
2256             /**************************
2257              * CALCULATE INTERACTIONS *
2258              **************************/
2259
2260             if (gmx_mm_any_lt(rsq20,rcutoff2))
2261             {
2262
2263             r20              = _mm_mul_ps(rsq20,rinv20);
2264             r20              = _mm_andnot_ps(dummy_mask,r20);
2265
2266             /* EWALD ELECTROSTATICS */
2267
2268             /* Analytical PME correction */
2269             zeta2            = _mm_mul_ps(beta2,rsq20);
2270             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
2271             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2272             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2273             felec            = _mm_mul_ps(qq20,felec);
2274
2275             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2276
2277             fscal            = felec;
2278
2279             fscal            = _mm_and_ps(fscal,cutoff_mask);
2280
2281             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2282
2283              /* Update vectorial force */
2284             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2285             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2286             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2287
2288             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2289             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2290             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2291
2292             }
2293
2294             /**************************
2295              * CALCULATE INTERACTIONS *
2296              **************************/
2297
2298             if (gmx_mm_any_lt(rsq21,rcutoff2))
2299             {
2300
2301             r21              = _mm_mul_ps(rsq21,rinv21);
2302             r21              = _mm_andnot_ps(dummy_mask,r21);
2303
2304             /* EWALD ELECTROSTATICS */
2305
2306             /* Analytical PME correction */
2307             zeta2            = _mm_mul_ps(beta2,rsq21);
2308             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
2309             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2310             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2311             felec            = _mm_mul_ps(qq21,felec);
2312
2313             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2314
2315             fscal            = felec;
2316
2317             fscal            = _mm_and_ps(fscal,cutoff_mask);
2318
2319             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2320
2321              /* Update vectorial force */
2322             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2323             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2324             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2325
2326             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2327             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2328             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2329
2330             }
2331
2332             /**************************
2333              * CALCULATE INTERACTIONS *
2334              **************************/
2335
2336             if (gmx_mm_any_lt(rsq22,rcutoff2))
2337             {
2338
2339             r22              = _mm_mul_ps(rsq22,rinv22);
2340             r22              = _mm_andnot_ps(dummy_mask,r22);
2341
2342             /* EWALD ELECTROSTATICS */
2343
2344             /* Analytical PME correction */
2345             zeta2            = _mm_mul_ps(beta2,rsq22);
2346             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
2347             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2348             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2349             felec            = _mm_mul_ps(qq22,felec);
2350
2351             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2352
2353             fscal            = felec;
2354
2355             fscal            = _mm_and_ps(fscal,cutoff_mask);
2356
2357             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2358
2359              /* Update vectorial force */
2360             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2361             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2362             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2363
2364             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2365             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2366             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2367
2368             }
2369
2370             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2371             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2372             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2373             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2374
2375             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2376                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2377
2378             /* Inner loop uses 309 flops */
2379         }
2380
2381         /* End of innermost loop */
2382
2383         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2384                                               f+i_coord_offset,fshift+i_shift_offset);
2385
2386         /* Increment number of inner iterations */
2387         inneriter                  += j_index_end - j_index_start;
2388
2389         /* Outer loop uses 18 flops */
2390     }
2391
2392     /* Increment number of outer iterations */
2393     outeriter        += nri;
2394
2395     /* Update outer/inner flops */
2396
2397     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*309);
2398 }