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