src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwLJ_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 113     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 114     __m128i          ewitab;
 115     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 116     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 117     real             *ewtab;
 118     __m128           dummy_mask,cutoff_mask;
 119     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 120     __m128           one     = _mm_set1_ps(1.0);
 121     __m128           two     = _mm_set1_ps(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_ps(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     nvdwtype         = fr->ntype;
 136     vdwparam         = fr->nbfp;
 137     vdwtype          = mdatoms->typeA;
 138
 139     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 140     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 141     beta2            = _mm_mul_ps(beta,beta);
 142     beta3            = _mm_mul_ps(beta,beta2);
 143     ewtab            = fr->ic->tabq_coul_FDV0;
 144     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 145     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 146
 147     /* Setup water-specific parameters */
 148     inr              = nlist->iinr[0];
 149     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 150     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 151     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 152     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 153
 154     jq0              = _mm_set1_ps(charge[inr+0]);
 155     jq1              = _mm_set1_ps(charge[inr+1]);
 156     jq2              = _mm_set1_ps(charge[inr+2]);
 157     vdwjidx0A        = 2*vdwtype[inr+0];
 158     qq00             = _mm_mul_ps(iq0,jq0);
 159     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 160     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 161     qq01             = _mm_mul_ps(iq0,jq1);
 162     qq02             = _mm_mul_ps(iq0,jq2);
 163     qq10             = _mm_mul_ps(iq1,jq0);
 164     qq11             = _mm_mul_ps(iq1,jq1);
 165     qq12             = _mm_mul_ps(iq1,jq2);
 166     qq20             = _mm_mul_ps(iq2,jq0);
 167     qq21             = _mm_mul_ps(iq2,jq1);
 168     qq22             = _mm_mul_ps(iq2,jq2);
 169
 170     /* Avoid stupid compiler warnings */
 171     jnrA = jnrB = jnrC = jnrD = 0;
 172     j_coord_offsetA = 0;
 173     j_coord_offsetB = 0;
 174     j_coord_offsetC = 0;
 175     j_coord_offsetD = 0;
 176
 177     outeriter        = 0;
 178     inneriter        = 0;
 179
 180     for(iidx=0;iidx<4*DIM;iidx++)
 181     {
 182         scratch[iidx] = 0.0;
 183     }
 184
 185     /* Start outer loop over neighborlists */
 186     for(iidx=0; iidx<nri; iidx++)
 187     {
 188         /* Load shift vector for this list */
 189         i_shift_offset   = DIM*shiftidx[iidx];
 190
 191         /* Load limits for loop over neighbors */
 192         j_index_start    = jindex[iidx];
 193         j_index_end      = jindex[iidx+1];
 194
 195         /* Get outer coordinate index */
 196         inr              = iinr[iidx];
 197         i_coord_offset   = DIM*inr;
 198
 199         /* Load i particle coords and add shift vector */
 200         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 201                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 202
 203         fix0             = _mm_setzero_ps();
 204         fiy0             = _mm_setzero_ps();
 205         fiz0             = _mm_setzero_ps();
 206         fix1             = _mm_setzero_ps();
 207         fiy1             = _mm_setzero_ps();
 208         fiz1             = _mm_setzero_ps();
 209         fix2             = _mm_setzero_ps();
 210         fiy2             = _mm_setzero_ps();
 211         fiz2             = _mm_setzero_ps();
 212
 213         /* Reset potential sums */
 214         velecsum         = _mm_setzero_ps();
 215         vvdwsum          = _mm_setzero_ps();
 216
 217         /* Start inner kernel loop */
 218         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 219         {
 220
 221             /* Get j neighbor index, and coordinate index */
 222             jnrA             = jjnr[jidx];
 223             jnrB             = jjnr[jidx+1];
 224             jnrC             = jjnr[jidx+2];
 225             jnrD             = jjnr[jidx+3];
 226             j_coord_offsetA  = DIM*jnrA;
 227             j_coord_offsetB  = DIM*jnrB;
 228             j_coord_offsetC  = DIM*jnrC;
 229             j_coord_offsetD  = DIM*jnrD;
 230
 231             /* load j atom coordinates */
 232             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 233                                               x+j_coord_offsetC,x+j_coord_offsetD,
 234                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 235
 236             /* Calculate displacement vector */
 237             dx00             = _mm_sub_ps(ix0,jx0);
 238             dy00             = _mm_sub_ps(iy0,jy0);
 239             dz00             = _mm_sub_ps(iz0,jz0);
 240             dx01             = _mm_sub_ps(ix0,jx1);
 241             dy01             = _mm_sub_ps(iy0,jy1);
 242             dz01             = _mm_sub_ps(iz0,jz1);
 243             dx02             = _mm_sub_ps(ix0,jx2);
 244             dy02             = _mm_sub_ps(iy0,jy2);
 245             dz02             = _mm_sub_ps(iz0,jz2);
 246             dx10             = _mm_sub_ps(ix1,jx0);
 247             dy10             = _mm_sub_ps(iy1,jy0);
 248             dz10             = _mm_sub_ps(iz1,jz0);
 249             dx11             = _mm_sub_ps(ix1,jx1);
 250             dy11             = _mm_sub_ps(iy1,jy1);
 251             dz11             = _mm_sub_ps(iz1,jz1);
 252             dx12             = _mm_sub_ps(ix1,jx2);
 253             dy12             = _mm_sub_ps(iy1,jy2);
 254             dz12             = _mm_sub_ps(iz1,jz2);
 255             dx20             = _mm_sub_ps(ix2,jx0);
 256             dy20             = _mm_sub_ps(iy2,jy0);
 257             dz20             = _mm_sub_ps(iz2,jz0);
 258             dx21             = _mm_sub_ps(ix2,jx1);
 259             dy21             = _mm_sub_ps(iy2,jy1);
 260             dz21             = _mm_sub_ps(iz2,jz1);
 261             dx22             = _mm_sub_ps(ix2,jx2);
 262             dy22             = _mm_sub_ps(iy2,jy2);
 263             dz22             = _mm_sub_ps(iz2,jz2);
 264
 265             /* Calculate squared distance and things based on it */
 266             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 267             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 268             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 269             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 270             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 271             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 272             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 273             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 274             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 275
 276             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 277             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 278             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 279             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 280             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 281             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 282             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 283             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 284             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 285
 286             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 287             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 288             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 289             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 290             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 291             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 292             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 293             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 294             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 295
 296             fjx0             = _mm_setzero_ps();
 297             fjy0             = _mm_setzero_ps();
 298             fjz0             = _mm_setzero_ps();
 299             fjx1             = _mm_setzero_ps();
 300             fjy1             = _mm_setzero_ps();
 301             fjz1             = _mm_setzero_ps();
 302             fjx2             = _mm_setzero_ps();
 303             fjy2             = _mm_setzero_ps();
 304             fjz2             = _mm_setzero_ps();
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             r00              = _mm_mul_ps(rsq00,rinv00);
 311
 312             /* EWALD ELECTROSTATICS */
 313
 314             /* Analytical PME correction */
 315             zeta2            = _mm_mul_ps(beta2,rsq00);
 316             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 317             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 318             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 319             felec            = _mm_mul_ps(qq00,felec);
 320             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 321             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 322             velec            = _mm_mul_ps(qq00,velec);
 323
 324             /* LENNARD-JONES DISPERSION/REPULSION */
 325
 326             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 327             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 328             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 329             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 330             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 331
 332             /* Update potential sum for this i atom from the interaction with this j atom. */
 333             velecsum         = _mm_add_ps(velecsum,velec);
 334             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 335
 336             fscal            = _mm_add_ps(felec,fvdw);
 337
 338              /* Update vectorial force */
 339             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 340             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 341             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 342
 343             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 344             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 345             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             r01              = _mm_mul_ps(rsq01,rinv01);
 352
 353             /* EWALD ELECTROSTATICS */
 354
 355             /* Analytical PME correction */
 356             zeta2            = _mm_mul_ps(beta2,rsq01);
 357             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 358             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 359             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 360             felec            = _mm_mul_ps(qq01,felec);
 361             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 362             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 363             velec            = _mm_mul_ps(qq01,velec);
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             velecsum         = _mm_add_ps(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370              /* Update vectorial force */
 371             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 372             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 373             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 374
 375             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 376             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 377             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             r02              = _mm_mul_ps(rsq02,rinv02);
 384
 385             /* EWALD ELECTROSTATICS */
 386
 387             /* Analytical PME correction */
 388             zeta2            = _mm_mul_ps(beta2,rsq02);
 389             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 390             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 391             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 392             felec            = _mm_mul_ps(qq02,felec);
 393             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 394             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 395             velec            = _mm_mul_ps(qq02,velec);
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velecsum         = _mm_add_ps(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402              /* Update vectorial force */
 403             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 404             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 405             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 406
 407             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 408             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 409             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             r10              = _mm_mul_ps(rsq10,rinv10);
 416
 417             /* EWALD ELECTROSTATICS */
 418
 419             /* Analytical PME correction */
 420             zeta2            = _mm_mul_ps(beta2,rsq10);
 421             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 422             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 423             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 424             felec            = _mm_mul_ps(qq10,felec);
 425             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 426             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 427             velec            = _mm_mul_ps(qq10,velec);
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velecsum         = _mm_add_ps(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434              /* Update vectorial force */
 435             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 436             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 437             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 438
 439             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 440             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 441             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             r11              = _mm_mul_ps(rsq11,rinv11);
 448
 449             /* EWALD ELECTROSTATICS */
 450
 451             /* Analytical PME correction */
 452             zeta2            = _mm_mul_ps(beta2,rsq11);
 453             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 454             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 455             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 456             felec            = _mm_mul_ps(qq11,felec);
 457             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 458             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 459             velec            = _mm_mul_ps(qq11,velec);
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velecsum         = _mm_add_ps(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466              /* Update vectorial force */
 467             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 468             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 469             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 470
 471             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 472             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 473             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 474
 475             /**************************
 476              * CALCULATE INTERACTIONS *
 477              **************************/
 478
 479             r12              = _mm_mul_ps(rsq12,rinv12);
 480
 481             /* EWALD ELECTROSTATICS */
 482
 483             /* Analytical PME correction */
 484             zeta2            = _mm_mul_ps(beta2,rsq12);
 485             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 486             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 487             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 488             felec            = _mm_mul_ps(qq12,felec);
 489             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 490             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 491             velec            = _mm_mul_ps(qq12,velec);
 492
 493             /* Update potential sum for this i atom from the interaction with this j atom. */
 494             velecsum         = _mm_add_ps(velecsum,velec);
 495
 496             fscal            = felec;
 497
 498              /* Update vectorial force */
 499             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 500             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 501             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 502
 503             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 504             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 505             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 506
 507             /**************************
 508              * CALCULATE INTERACTIONS *
 509              **************************/
 510
 511             r20              = _mm_mul_ps(rsq20,rinv20);
 512
 513             /* EWALD ELECTROSTATICS */
 514
 515             /* Analytical PME correction */
 516             zeta2            = _mm_mul_ps(beta2,rsq20);
 517             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 518             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 519             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 520             felec            = _mm_mul_ps(qq20,felec);
 521             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 522             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 523             velec            = _mm_mul_ps(qq20,velec);
 524
 525             /* Update potential sum for this i atom from the interaction with this j atom. */
 526             velecsum         = _mm_add_ps(velecsum,velec);
 527
 528             fscal            = felec;
 529
 530              /* Update vectorial force */
 531             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 532             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 533             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 534
 535             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 536             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 537             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 538
 539             /**************************
 540              * CALCULATE INTERACTIONS *
 541              **************************/
 542
 543             r21              = _mm_mul_ps(rsq21,rinv21);
 544
 545             /* EWALD ELECTROSTATICS */
 546
 547             /* Analytical PME correction */
 548             zeta2            = _mm_mul_ps(beta2,rsq21);
 549             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 550             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 551             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 552             felec            = _mm_mul_ps(qq21,felec);
 553             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 554             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 555             velec            = _mm_mul_ps(qq21,velec);
 556
 557             /* Update potential sum for this i atom from the interaction with this j atom. */
 558             velecsum         = _mm_add_ps(velecsum,velec);
 559
 560             fscal            = felec;
 561
 562              /* Update vectorial force */
 563             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 564             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 565             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 566
 567             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 568             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 569             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 570
 571             /**************************
 572              * CALCULATE INTERACTIONS *
 573              **************************/
 574
 575             r22              = _mm_mul_ps(rsq22,rinv22);
 576
 577             /* EWALD ELECTROSTATICS */
 578
 579             /* Analytical PME correction */
 580             zeta2            = _mm_mul_ps(beta2,rsq22);
 581             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 582             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 583             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 584             felec            = _mm_mul_ps(qq22,felec);
 585             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 586             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 587             velec            = _mm_mul_ps(qq22,velec);
 588
 589             /* Update potential sum for this i atom from the interaction with this j atom. */
 590             velecsum         = _mm_add_ps(velecsum,velec);
 591
 592             fscal            = felec;
 593
 594              /* Update vectorial force */
 595             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 596             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 597             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 598
 599             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 600             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 601             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 602
 603             fjptrA             = f+j_coord_offsetA;
 604             fjptrB             = f+j_coord_offsetB;
 605             fjptrC             = f+j_coord_offsetC;
 606             fjptrD             = f+j_coord_offsetD;
 607
 608             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 609                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 610
 611             /* Inner loop uses 273 flops */
 612         }
 613
 614         if(jidx<j_index_end)
 615         {
 616
 617             /* Get j neighbor index, and coordinate index */
 618             jnrlistA         = jjnr[jidx];
 619             jnrlistB         = jjnr[jidx+1];
 620             jnrlistC         = jjnr[jidx+2];
 621             jnrlistD         = jjnr[jidx+3];
 622             /* Sign of each element will be negative for non-real atoms.
 623              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 624              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 625              */
 626             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 627             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 628             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 629             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 630             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 631             j_coord_offsetA  = DIM*jnrA;
 632             j_coord_offsetB  = DIM*jnrB;
 633             j_coord_offsetC  = DIM*jnrC;
 634             j_coord_offsetD  = DIM*jnrD;
 635
 636             /* load j atom coordinates */
 637             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 638                                               x+j_coord_offsetC,x+j_coord_offsetD,
 639                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 640
 641             /* Calculate displacement vector */
 642             dx00             = _mm_sub_ps(ix0,jx0);
 643             dy00             = _mm_sub_ps(iy0,jy0);
 644             dz00             = _mm_sub_ps(iz0,jz0);
 645             dx01             = _mm_sub_ps(ix0,jx1);
 646             dy01             = _mm_sub_ps(iy0,jy1);
 647             dz01             = _mm_sub_ps(iz0,jz1);
 648             dx02             = _mm_sub_ps(ix0,jx2);
 649             dy02             = _mm_sub_ps(iy0,jy2);
 650             dz02             = _mm_sub_ps(iz0,jz2);
 651             dx10             = _mm_sub_ps(ix1,jx0);
 652             dy10             = _mm_sub_ps(iy1,jy0);
 653             dz10             = _mm_sub_ps(iz1,jz0);
 654             dx11             = _mm_sub_ps(ix1,jx1);
 655             dy11             = _mm_sub_ps(iy1,jy1);
 656             dz11             = _mm_sub_ps(iz1,jz1);
 657             dx12             = _mm_sub_ps(ix1,jx2);
 658             dy12             = _mm_sub_ps(iy1,jy2);
 659             dz12             = _mm_sub_ps(iz1,jz2);
 660             dx20             = _mm_sub_ps(ix2,jx0);
 661             dy20             = _mm_sub_ps(iy2,jy0);
 662             dz20             = _mm_sub_ps(iz2,jz0);
 663             dx21             = _mm_sub_ps(ix2,jx1);
 664             dy21             = _mm_sub_ps(iy2,jy1);
 665             dz21             = _mm_sub_ps(iz2,jz1);
 666             dx22             = _mm_sub_ps(ix2,jx2);
 667             dy22             = _mm_sub_ps(iy2,jy2);
 668             dz22             = _mm_sub_ps(iz2,jz2);
 669
 670             /* Calculate squared distance and things based on it */
 671             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 672             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 673             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 674             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 675             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 676             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 677             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 678             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 679             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 680
 681             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 682             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 683             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 684             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 685             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 686             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 687             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 688             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 689             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 690
 691             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 692             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 693             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 694             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 695             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 696             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 697             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 698             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 699             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 700
 701             fjx0             = _mm_setzero_ps();
 702             fjy0             = _mm_setzero_ps();
 703             fjz0             = _mm_setzero_ps();
 704             fjx1             = _mm_setzero_ps();
 705             fjy1             = _mm_setzero_ps();
 706             fjz1             = _mm_setzero_ps();
 707             fjx2             = _mm_setzero_ps();
 708             fjy2             = _mm_setzero_ps();
 709             fjz2             = _mm_setzero_ps();
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             r00              = _mm_mul_ps(rsq00,rinv00);
 716             r00              = _mm_andnot_ps(dummy_mask,r00);
 717
 718             /* EWALD ELECTROSTATICS */
 719
 720             /* Analytical PME correction */
 721             zeta2            = _mm_mul_ps(beta2,rsq00);
 722             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 723             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 724             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 725             felec            = _mm_mul_ps(qq00,felec);
 726             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 727             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 728             velec            = _mm_mul_ps(qq00,velec);
 729
 730             /* LENNARD-JONES DISPERSION/REPULSION */
 731
 732             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 733             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 734             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 735             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 736             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             velec            = _mm_andnot_ps(dummy_mask,velec);
 740             velecsum         = _mm_add_ps(velecsum,velec);
 741             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 742             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 743
 744             fscal            = _mm_add_ps(felec,fvdw);
 745
 746             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 747
 748              /* Update vectorial force */
 749             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 750             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 751             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 752
 753             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 754             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 755             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 756
 757             /**************************
 758              * CALCULATE INTERACTIONS *
 759              **************************/
 760
 761             r01              = _mm_mul_ps(rsq01,rinv01);
 762             r01              = _mm_andnot_ps(dummy_mask,r01);
 763
 764             /* EWALD ELECTROSTATICS */
 765
 766             /* Analytical PME correction */
 767             zeta2            = _mm_mul_ps(beta2,rsq01);
 768             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 769             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 770             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 771             felec            = _mm_mul_ps(qq01,felec);
 772             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 773             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 774             velec            = _mm_mul_ps(qq01,velec);
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_andnot_ps(dummy_mask,velec);
 778             velecsum         = _mm_add_ps(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 783
 784              /* Update vectorial force */
 785             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 786             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 787             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 788
 789             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 790             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 791             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             r02              = _mm_mul_ps(rsq02,rinv02);
 798             r02              = _mm_andnot_ps(dummy_mask,r02);
 799
 800             /* EWALD ELECTROSTATICS */
 801
 802             /* Analytical PME correction */
 803             zeta2            = _mm_mul_ps(beta2,rsq02);
 804             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 805             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 806             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 807             felec            = _mm_mul_ps(qq02,felec);
 808             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 809             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 810             velec            = _mm_mul_ps(qq02,velec);
 811
 812             /* Update potential sum for this i atom from the interaction with this j atom. */
 813             velec            = _mm_andnot_ps(dummy_mask,velec);
 814             velecsum         = _mm_add_ps(velecsum,velec);
 815
 816             fscal            = felec;
 817
 818             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 819
 820              /* Update vectorial force */
 821             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 822             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 823             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 824
 825             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 826             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 827             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 828
 829             /**************************
 830              * CALCULATE INTERACTIONS *
 831              **************************/
 832
 833             r10              = _mm_mul_ps(rsq10,rinv10);
 834             r10              = _mm_andnot_ps(dummy_mask,r10);
 835
 836             /* EWALD ELECTROSTATICS */
 837
 838             /* Analytical PME correction */
 839             zeta2            = _mm_mul_ps(beta2,rsq10);
 840             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 841             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 842             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 843             felec            = _mm_mul_ps(qq10,felec);
 844             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 845             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 846             velec            = _mm_mul_ps(qq10,velec);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             velec            = _mm_andnot_ps(dummy_mask,velec);
 850             velecsum         = _mm_add_ps(velecsum,velec);
 851
 852             fscal            = felec;
 853
 854             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 855
 856              /* Update vectorial force */
 857             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 858             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 859             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 860
 861             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 862             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 863             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             r11              = _mm_mul_ps(rsq11,rinv11);
 870             r11              = _mm_andnot_ps(dummy_mask,r11);
 871
 872             /* EWALD ELECTROSTATICS */
 873
 874             /* Analytical PME correction */
 875             zeta2            = _mm_mul_ps(beta2,rsq11);
 876             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 877             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 878             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 879             felec            = _mm_mul_ps(qq11,felec);
 880             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 881             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 882             velec            = _mm_mul_ps(qq11,velec);
 883
 884             /* Update potential sum for this i atom from the interaction with this j atom. */
 885             velec            = _mm_andnot_ps(dummy_mask,velec);
 886             velecsum         = _mm_add_ps(velecsum,velec);
 887
 888             fscal            = felec;
 889
 890             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 891
 892              /* Update vectorial force */
 893             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 894             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 895             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 896
 897             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 898             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 899             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 900
 901             /**************************
 902              * CALCULATE INTERACTIONS *
 903              **************************/
 904
 905             r12              = _mm_mul_ps(rsq12,rinv12);
 906             r12              = _mm_andnot_ps(dummy_mask,r12);
 907
 908             /* EWALD ELECTROSTATICS */
 909
 910             /* Analytical PME correction */
 911             zeta2            = _mm_mul_ps(beta2,rsq12);
 912             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 913             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 914             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 915             felec            = _mm_mul_ps(qq12,felec);
 916             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 917             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 918             velec            = _mm_mul_ps(qq12,velec);
 919
 920             /* Update potential sum for this i atom from the interaction with this j atom. */
 921             velec            = _mm_andnot_ps(dummy_mask,velec);
 922             velecsum         = _mm_add_ps(velecsum,velec);
 923
 924             fscal            = felec;
 925
 926             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 927
 928              /* Update vectorial force */
 929             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 930             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 931             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 932
 933             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 934             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 935             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 936
 937             /**************************
 938              * CALCULATE INTERACTIONS *
 939              **************************/
 940
 941             r20              = _mm_mul_ps(rsq20,rinv20);
 942             r20              = _mm_andnot_ps(dummy_mask,r20);
 943
 944             /* EWALD ELECTROSTATICS */
 945
 946             /* Analytical PME correction */
 947             zeta2            = _mm_mul_ps(beta2,rsq20);
 948             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 949             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 950             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 951             felec            = _mm_mul_ps(qq20,felec);
 952             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 953             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 954             velec            = _mm_mul_ps(qq20,velec);
 955
 956             /* Update potential sum for this i atom from the interaction with this j atom. */
 957             velec            = _mm_andnot_ps(dummy_mask,velec);
 958             velecsum         = _mm_add_ps(velecsum,velec);
 959
 960             fscal            = felec;
 961
 962             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 963
 964              /* Update vectorial force */
 965             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 966             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 967             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 968
 969             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 970             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 971             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 972
 973             /**************************
 974              * CALCULATE INTERACTIONS *
 975              **************************/
 976
 977             r21              = _mm_mul_ps(rsq21,rinv21);
 978             r21              = _mm_andnot_ps(dummy_mask,r21);
 979
 980             /* EWALD ELECTROSTATICS */
 981
 982             /* Analytical PME correction */
 983             zeta2            = _mm_mul_ps(beta2,rsq21);
 984             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 985             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 986             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 987             felec            = _mm_mul_ps(qq21,felec);
 988             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 989             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 990             velec            = _mm_mul_ps(qq21,velec);
 991
 992             /* Update potential sum for this i atom from the interaction with this j atom. */
 993             velec            = _mm_andnot_ps(dummy_mask,velec);
 994             velecsum         = _mm_add_ps(velecsum,velec);
 995
 996             fscal            = felec;
 997
 998             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 999
1000              /* Update vectorial force */
1001             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1002             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1003             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1004
1005             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1006             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1007             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1008
1009             /**************************
1010              * CALCULATE INTERACTIONS *
1011              **************************/
1012
1013             r22              = _mm_mul_ps(rsq22,rinv22);
1014             r22              = _mm_andnot_ps(dummy_mask,r22);
1015
1016             /* EWALD ELECTROSTATICS */
1017
1018             /* Analytical PME correction */
1019             zeta2            = _mm_mul_ps(beta2,rsq22);
1020             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1021             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1022             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1023             felec            = _mm_mul_ps(qq22,felec);
1024             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1025             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
1026             velec            = _mm_mul_ps(qq22,velec);
1027
1028             /* Update potential sum for this i atom from the interaction with this j atom. */
1029             velec            = _mm_andnot_ps(dummy_mask,velec);
1030             velecsum         = _mm_add_ps(velecsum,velec);
1031
1032             fscal            = felec;
1033
1034             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1035
1036              /* Update vectorial force */
1037             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1038             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1039             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1040
1041             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1042             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1043             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1044
1045             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1046             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1047             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1048             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1049
1050             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1051                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1052
1053             /* Inner loop uses 282 flops */
1054         }
1055
1056         /* End of innermost loop */
1057
1058         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1059                                               f+i_coord_offset,fshift+i_shift_offset);
1060
1061         ggid                        = gid[iidx];
1062         /* Update potential energies */
1063         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1064         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1065
1066         /* Increment number of inner iterations */
1067         inneriter                  += j_index_end - j_index_start;
1068
1069         /* Outer loop uses 20 flops */
1070     }
1071
1072     /* Increment number of outer iterations */
1073     outeriter        += nri;
1074
1075     /* Update outer/inner flops */
1076
1077     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*282);
1078 }
1079 /*
1080  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_avx_128_fma_single
1081  * Electrostatics interaction: Ewald
1082  * VdW interaction:            LennardJones
1083  * Geometry:                   Water3-Water3
1084  * Calculate force/pot:        Force
1085  */
1086 void
1087 nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_avx_128_fma_single
1088                     (t_nblist                    * gmx_restrict       nlist,
1089                      rvec                        * gmx_restrict          xx,
1090                      rvec                        * gmx_restrict          ff,
1091                      t_forcerec                  * gmx_restrict          fr,
1092                      t_mdatoms                   * gmx_restrict     mdatoms,
1093                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1094                      t_nrnb                      * gmx_restrict        nrnb)
1095 {
1096     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1097      * just 0 for non-waters.
1098      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1099      * jnr indices corresponding to data put in the four positions in the SIMD register.
1100      */
1101     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1102     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1103     int              jnrA,jnrB,jnrC,jnrD;
1104     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1105     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1106     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1107     real             rcutoff_scalar;
1108     real             *shiftvec,*fshift,*x,*f;
1109     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1110     real             scratch[4*DIM];
1111     __m128           fscal,rcutoff,rcutoff2,jidxall;
1112     int              vdwioffset0;
1113     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1114     int              vdwioffset1;
1115     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1116     int              vdwioffset2;
1117     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1118     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1119     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1120     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1121     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1122     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1123     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1124     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1125     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1126     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1127     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1128     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1129     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1130     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1131     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1132     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1133     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1134     real             *charge;
1135     int              nvdwtype;
1136     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1137     int              *vdwtype;
1138     real             *vdwparam;
1139     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1140     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1141     __m128i          ewitab;
1142     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1143     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1144     real             *ewtab;
1145     __m128           dummy_mask,cutoff_mask;
1146     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1147     __m128           one     = _mm_set1_ps(1.0);
1148     __m128           two     = _mm_set1_ps(2.0);
1149     x                = xx[0];
1150     f                = ff[0];
1151
1152     nri              = nlist->nri;
1153     iinr             = nlist->iinr;
1154     jindex           = nlist->jindex;
1155     jjnr             = nlist->jjnr;
1156     shiftidx         = nlist->shift;
1157     gid              = nlist->gid;
1158     shiftvec         = fr->shift_vec[0];
1159     fshift           = fr->fshift[0];
1160     facel            = _mm_set1_ps(fr->epsfac);
1161     charge           = mdatoms->chargeA;
1162     nvdwtype         = fr->ntype;
1163     vdwparam         = fr->nbfp;
1164     vdwtype          = mdatoms->typeA;
1165
1166     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1167     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1168     beta2            = _mm_mul_ps(beta,beta);
1169     beta3            = _mm_mul_ps(beta,beta2);
1170     ewtab            = fr->ic->tabq_coul_F;
1171     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1172     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1173
1174     /* Setup water-specific parameters */
1175     inr              = nlist->iinr[0];
1176     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1177     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1178     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1179     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1180
1181     jq0              = _mm_set1_ps(charge[inr+0]);
1182     jq1              = _mm_set1_ps(charge[inr+1]);
1183     jq2              = _mm_set1_ps(charge[inr+2]);
1184     vdwjidx0A        = 2*vdwtype[inr+0];
1185     qq00             = _mm_mul_ps(iq0,jq0);
1186     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1187     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1188     qq01             = _mm_mul_ps(iq0,jq1);
1189     qq02             = _mm_mul_ps(iq0,jq2);
1190     qq10             = _mm_mul_ps(iq1,jq0);
1191     qq11             = _mm_mul_ps(iq1,jq1);
1192     qq12             = _mm_mul_ps(iq1,jq2);
1193     qq20             = _mm_mul_ps(iq2,jq0);
1194     qq21             = _mm_mul_ps(iq2,jq1);
1195     qq22             = _mm_mul_ps(iq2,jq2);
1196
1197     /* Avoid stupid compiler warnings */
1198     jnrA = jnrB = jnrC = jnrD = 0;
1199     j_coord_offsetA = 0;
1200     j_coord_offsetB = 0;
1201     j_coord_offsetC = 0;
1202     j_coord_offsetD = 0;
1203
1204     outeriter        = 0;
1205     inneriter        = 0;
1206
1207     for(iidx=0;iidx<4*DIM;iidx++)
1208     {
1209         scratch[iidx] = 0.0;
1210     }
1211
1212     /* Start outer loop over neighborlists */
1213     for(iidx=0; iidx<nri; iidx++)
1214     {
1215         /* Load shift vector for this list */
1216         i_shift_offset   = DIM*shiftidx[iidx];
1217
1218         /* Load limits for loop over neighbors */
1219         j_index_start    = jindex[iidx];
1220         j_index_end      = jindex[iidx+1];
1221
1222         /* Get outer coordinate index */
1223         inr              = iinr[iidx];
1224         i_coord_offset   = DIM*inr;
1225
1226         /* Load i particle coords and add shift vector */
1227         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1228                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1229
1230         fix0             = _mm_setzero_ps();
1231         fiy0             = _mm_setzero_ps();
1232         fiz0             = _mm_setzero_ps();
1233         fix1             = _mm_setzero_ps();
1234         fiy1             = _mm_setzero_ps();
1235         fiz1             = _mm_setzero_ps();
1236         fix2             = _mm_setzero_ps();
1237         fiy2             = _mm_setzero_ps();
1238         fiz2             = _mm_setzero_ps();
1239
1240         /* Start inner kernel loop */
1241         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1242         {
1243
1244             /* Get j neighbor index, and coordinate index */
1245             jnrA             = jjnr[jidx];
1246             jnrB             = jjnr[jidx+1];
1247             jnrC             = jjnr[jidx+2];
1248             jnrD             = jjnr[jidx+3];
1249             j_coord_offsetA  = DIM*jnrA;
1250             j_coord_offsetB  = DIM*jnrB;
1251             j_coord_offsetC  = DIM*jnrC;
1252             j_coord_offsetD  = DIM*jnrD;
1253
1254             /* load j atom coordinates */
1255             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1256                                               x+j_coord_offsetC,x+j_coord_offsetD,
1257                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1258
1259             /* Calculate displacement vector */
1260             dx00             = _mm_sub_ps(ix0,jx0);
1261             dy00             = _mm_sub_ps(iy0,jy0);
1262             dz00             = _mm_sub_ps(iz0,jz0);
1263             dx01             = _mm_sub_ps(ix0,jx1);
1264             dy01             = _mm_sub_ps(iy0,jy1);
1265             dz01             = _mm_sub_ps(iz0,jz1);
1266             dx02             = _mm_sub_ps(ix0,jx2);
1267             dy02             = _mm_sub_ps(iy0,jy2);
1268             dz02             = _mm_sub_ps(iz0,jz2);
1269             dx10             = _mm_sub_ps(ix1,jx0);
1270             dy10             = _mm_sub_ps(iy1,jy0);
1271             dz10             = _mm_sub_ps(iz1,jz0);
1272             dx11             = _mm_sub_ps(ix1,jx1);
1273             dy11             = _mm_sub_ps(iy1,jy1);
1274             dz11             = _mm_sub_ps(iz1,jz1);
1275             dx12             = _mm_sub_ps(ix1,jx2);
1276             dy12             = _mm_sub_ps(iy1,jy2);
1277             dz12             = _mm_sub_ps(iz1,jz2);
1278             dx20             = _mm_sub_ps(ix2,jx0);
1279             dy20             = _mm_sub_ps(iy2,jy0);
1280             dz20             = _mm_sub_ps(iz2,jz0);
1281             dx21             = _mm_sub_ps(ix2,jx1);
1282             dy21             = _mm_sub_ps(iy2,jy1);
1283             dz21             = _mm_sub_ps(iz2,jz1);
1284             dx22             = _mm_sub_ps(ix2,jx2);
1285             dy22             = _mm_sub_ps(iy2,jy2);
1286             dz22             = _mm_sub_ps(iz2,jz2);
1287
1288             /* Calculate squared distance and things based on it */
1289             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1290             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1291             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1292             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1293             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1294             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1295             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1296             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1297             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1298
1299             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1300             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1301             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1302             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1303             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1304             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1305             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1306             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1307             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1308
1309             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1310             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1311             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1312             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1313             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1314             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1315             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1316             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1317             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1318
1319             fjx0             = _mm_setzero_ps();
1320             fjy0             = _mm_setzero_ps();
1321             fjz0             = _mm_setzero_ps();
1322             fjx1             = _mm_setzero_ps();
1323             fjy1             = _mm_setzero_ps();
1324             fjz1             = _mm_setzero_ps();
1325             fjx2             = _mm_setzero_ps();
1326             fjy2             = _mm_setzero_ps();
1327             fjz2             = _mm_setzero_ps();
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             r00              = _mm_mul_ps(rsq00,rinv00);
1334
1335             /* EWALD ELECTROSTATICS */
1336
1337             /* Analytical PME correction */
1338             zeta2            = _mm_mul_ps(beta2,rsq00);
1339             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1340             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1341             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1342             felec            = _mm_mul_ps(qq00,felec);
1343
1344             /* LENNARD-JONES DISPERSION/REPULSION */
1345
1346             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1347             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1348
1349             fscal            = _mm_add_ps(felec,fvdw);
1350
1351              /* Update vectorial force */
1352             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1353             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1354             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1355
1356             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1357             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1358             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1359
1360             /**************************
1361              * CALCULATE INTERACTIONS *
1362              **************************/
1363
1364             r01              = _mm_mul_ps(rsq01,rinv01);
1365
1366             /* EWALD ELECTROSTATICS */
1367
1368             /* Analytical PME correction */
1369             zeta2            = _mm_mul_ps(beta2,rsq01);
1370             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1371             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1372             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1373             felec            = _mm_mul_ps(qq01,felec);
1374
1375             fscal            = felec;
1376
1377              /* Update vectorial force */
1378             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1379             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1380             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1381
1382             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1383             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1384             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             r02              = _mm_mul_ps(rsq02,rinv02);
1391
1392             /* EWALD ELECTROSTATICS */
1393
1394             /* Analytical PME correction */
1395             zeta2            = _mm_mul_ps(beta2,rsq02);
1396             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1397             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1398             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1399             felec            = _mm_mul_ps(qq02,felec);
1400
1401             fscal            = felec;
1402
1403              /* Update vectorial force */
1404             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1405             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1406             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1407
1408             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1409             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1410             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             r10              = _mm_mul_ps(rsq10,rinv10);
1417
1418             /* EWALD ELECTROSTATICS */
1419
1420             /* Analytical PME correction */
1421             zeta2            = _mm_mul_ps(beta2,rsq10);
1422             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1423             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1424             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1425             felec            = _mm_mul_ps(qq10,felec);
1426
1427             fscal            = felec;
1428
1429              /* Update vectorial force */
1430             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1431             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1432             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1433
1434             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1435             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1436             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             r11              = _mm_mul_ps(rsq11,rinv11);
1443
1444             /* EWALD ELECTROSTATICS */
1445
1446             /* Analytical PME correction */
1447             zeta2            = _mm_mul_ps(beta2,rsq11);
1448             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1449             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1450             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1451             felec            = _mm_mul_ps(qq11,felec);
1452
1453             fscal            = felec;
1454
1455              /* Update vectorial force */
1456             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1457             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1458             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1459
1460             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1461             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1462             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             r12              = _mm_mul_ps(rsq12,rinv12);
1469
1470             /* EWALD ELECTROSTATICS */
1471
1472             /* Analytical PME correction */
1473             zeta2            = _mm_mul_ps(beta2,rsq12);
1474             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1475             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1476             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1477             felec            = _mm_mul_ps(qq12,felec);
1478
1479             fscal            = felec;
1480
1481              /* Update vectorial force */
1482             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1483             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1484             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1485
1486             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1487             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1488             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1489
1490             /**************************
1491              * CALCULATE INTERACTIONS *
1492              **************************/
1493
1494             r20              = _mm_mul_ps(rsq20,rinv20);
1495
1496             /* EWALD ELECTROSTATICS */
1497
1498             /* Analytical PME correction */
1499             zeta2            = _mm_mul_ps(beta2,rsq20);
1500             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1501             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1502             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1503             felec            = _mm_mul_ps(qq20,felec);
1504
1505             fscal            = felec;
1506
1507              /* Update vectorial force */
1508             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1509             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1510             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1511
1512             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1513             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1514             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1515
1516             /**************************
1517              * CALCULATE INTERACTIONS *
1518              **************************/
1519
1520             r21              = _mm_mul_ps(rsq21,rinv21);
1521
1522             /* EWALD ELECTROSTATICS */
1523
1524             /* Analytical PME correction */
1525             zeta2            = _mm_mul_ps(beta2,rsq21);
1526             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1527             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1528             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1529             felec            = _mm_mul_ps(qq21,felec);
1530
1531             fscal            = felec;
1532
1533              /* Update vectorial force */
1534             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1535             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1536             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1537
1538             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1539             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1540             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1541
1542             /**************************
1543              * CALCULATE INTERACTIONS *
1544              **************************/
1545
1546             r22              = _mm_mul_ps(rsq22,rinv22);
1547
1548             /* EWALD ELECTROSTATICS */
1549
1550             /* Analytical PME correction */
1551             zeta2            = _mm_mul_ps(beta2,rsq22);
1552             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1553             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1554             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1555             felec            = _mm_mul_ps(qq22,felec);
1556
1557             fscal            = felec;
1558
1559              /* Update vectorial force */
1560             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1561             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1562             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1563
1564             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1565             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1566             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1567
1568             fjptrA             = f+j_coord_offsetA;
1569             fjptrB             = f+j_coord_offsetB;
1570             fjptrC             = f+j_coord_offsetC;
1571             fjptrD             = f+j_coord_offsetD;
1572
1573             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1574                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1575
1576             /* Inner loop uses 259 flops */
1577         }
1578
1579         if(jidx<j_index_end)
1580         {
1581
1582             /* Get j neighbor index, and coordinate index */
1583             jnrlistA         = jjnr[jidx];
1584             jnrlistB         = jjnr[jidx+1];
1585             jnrlistC         = jjnr[jidx+2];
1586             jnrlistD         = jjnr[jidx+3];
1587             /* Sign of each element will be negative for non-real atoms.
1588              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1589              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1590              */
1591             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1592             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1593             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1594             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1595             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1596             j_coord_offsetA  = DIM*jnrA;
1597             j_coord_offsetB  = DIM*jnrB;
1598             j_coord_offsetC  = DIM*jnrC;
1599             j_coord_offsetD  = DIM*jnrD;
1600
1601             /* load j atom coordinates */
1602             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1603                                               x+j_coord_offsetC,x+j_coord_offsetD,
1604                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1605
1606             /* Calculate displacement vector */
1607             dx00             = _mm_sub_ps(ix0,jx0);
1608             dy00             = _mm_sub_ps(iy0,jy0);
1609             dz00             = _mm_sub_ps(iz0,jz0);
1610             dx01             = _mm_sub_ps(ix0,jx1);
1611             dy01             = _mm_sub_ps(iy0,jy1);
1612             dz01             = _mm_sub_ps(iz0,jz1);
1613             dx02             = _mm_sub_ps(ix0,jx2);
1614             dy02             = _mm_sub_ps(iy0,jy2);
1615             dz02             = _mm_sub_ps(iz0,jz2);
1616             dx10             = _mm_sub_ps(ix1,jx0);
1617             dy10             = _mm_sub_ps(iy1,jy0);
1618             dz10             = _mm_sub_ps(iz1,jz0);
1619             dx11             = _mm_sub_ps(ix1,jx1);
1620             dy11             = _mm_sub_ps(iy1,jy1);
1621             dz11             = _mm_sub_ps(iz1,jz1);
1622             dx12             = _mm_sub_ps(ix1,jx2);
1623             dy12             = _mm_sub_ps(iy1,jy2);
1624             dz12             = _mm_sub_ps(iz1,jz2);
1625             dx20             = _mm_sub_ps(ix2,jx0);
1626             dy20             = _mm_sub_ps(iy2,jy0);
1627             dz20             = _mm_sub_ps(iz2,jz0);
1628             dx21             = _mm_sub_ps(ix2,jx1);
1629             dy21             = _mm_sub_ps(iy2,jy1);
1630             dz21             = _mm_sub_ps(iz2,jz1);
1631             dx22             = _mm_sub_ps(ix2,jx2);
1632             dy22             = _mm_sub_ps(iy2,jy2);
1633             dz22             = _mm_sub_ps(iz2,jz2);
1634
1635             /* Calculate squared distance and things based on it */
1636             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1637             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1638             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1639             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1640             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1641             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1642             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1643             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1644             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1645
1646             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1647             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1648             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1649             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1650             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1651             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1652             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1653             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1654             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1655
1656             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1657             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1658             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1659             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1660             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1661             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1662             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1663             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1664             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1665
1666             fjx0             = _mm_setzero_ps();
1667             fjy0             = _mm_setzero_ps();
1668             fjz0             = _mm_setzero_ps();
1669             fjx1             = _mm_setzero_ps();
1670             fjy1             = _mm_setzero_ps();
1671             fjz1             = _mm_setzero_ps();
1672             fjx2             = _mm_setzero_ps();
1673             fjy2             = _mm_setzero_ps();
1674             fjz2             = _mm_setzero_ps();
1675
1676             /**************************
1677              * CALCULATE INTERACTIONS *
1678              **************************/
1679
1680             r00              = _mm_mul_ps(rsq00,rinv00);
1681             r00              = _mm_andnot_ps(dummy_mask,r00);
1682
1683             /* EWALD ELECTROSTATICS */
1684
1685             /* Analytical PME correction */
1686             zeta2            = _mm_mul_ps(beta2,rsq00);
1687             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1688             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1689             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1690             felec            = _mm_mul_ps(qq00,felec);
1691
1692             /* LENNARD-JONES DISPERSION/REPULSION */
1693
1694             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1695             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1696
1697             fscal            = _mm_add_ps(felec,fvdw);
1698
1699             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1700
1701              /* Update vectorial force */
1702             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1703             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1704             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1705
1706             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1707             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1708             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1709
1710             /**************************
1711              * CALCULATE INTERACTIONS *
1712              **************************/
1713
1714             r01              = _mm_mul_ps(rsq01,rinv01);
1715             r01              = _mm_andnot_ps(dummy_mask,r01);
1716
1717             /* EWALD ELECTROSTATICS */
1718
1719             /* Analytical PME correction */
1720             zeta2            = _mm_mul_ps(beta2,rsq01);
1721             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1722             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1723             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1724             felec            = _mm_mul_ps(qq01,felec);
1725
1726             fscal            = felec;
1727
1728             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1729
1730              /* Update vectorial force */
1731             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1732             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1733             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1734
1735             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1736             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1737             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1738
1739             /**************************
1740              * CALCULATE INTERACTIONS *
1741              **************************/
1742
1743             r02              = _mm_mul_ps(rsq02,rinv02);
1744             r02              = _mm_andnot_ps(dummy_mask,r02);
1745
1746             /* EWALD ELECTROSTATICS */
1747
1748             /* Analytical PME correction */
1749             zeta2            = _mm_mul_ps(beta2,rsq02);
1750             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1751             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1752             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1753             felec            = _mm_mul_ps(qq02,felec);
1754
1755             fscal            = felec;
1756
1757             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1758
1759              /* Update vectorial force */
1760             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1761             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1762             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1763
1764             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1765             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1766             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1767
1768             /**************************
1769              * CALCULATE INTERACTIONS *
1770              **************************/
1771
1772             r10              = _mm_mul_ps(rsq10,rinv10);
1773             r10              = _mm_andnot_ps(dummy_mask,r10);
1774
1775             /* EWALD ELECTROSTATICS */
1776
1777             /* Analytical PME correction */
1778             zeta2            = _mm_mul_ps(beta2,rsq10);
1779             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1780             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1781             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1782             felec            = _mm_mul_ps(qq10,felec);
1783
1784             fscal            = felec;
1785
1786             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1787
1788              /* Update vectorial force */
1789             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1790             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1791             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1792
1793             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1794             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1795             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1796
1797             /**************************
1798              * CALCULATE INTERACTIONS *
1799              **************************/
1800
1801             r11              = _mm_mul_ps(rsq11,rinv11);
1802             r11              = _mm_andnot_ps(dummy_mask,r11);
1803
1804             /* EWALD ELECTROSTATICS */
1805
1806             /* Analytical PME correction */
1807             zeta2            = _mm_mul_ps(beta2,rsq11);
1808             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1809             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1810             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1811             felec            = _mm_mul_ps(qq11,felec);
1812
1813             fscal            = felec;
1814
1815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1816
1817              /* Update vectorial force */
1818             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1819             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1820             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1821
1822             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1823             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1824             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1825
1826             /**************************
1827              * CALCULATE INTERACTIONS *
1828              **************************/
1829
1830             r12              = _mm_mul_ps(rsq12,rinv12);
1831             r12              = _mm_andnot_ps(dummy_mask,r12);
1832
1833             /* EWALD ELECTROSTATICS */
1834
1835             /* Analytical PME correction */
1836             zeta2            = _mm_mul_ps(beta2,rsq12);
1837             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1838             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1839             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1840             felec            = _mm_mul_ps(qq12,felec);
1841
1842             fscal            = felec;
1843
1844             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1845
1846              /* Update vectorial force */
1847             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1848             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1849             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1850
1851             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1852             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1853             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1854
1855             /**************************
1856              * CALCULATE INTERACTIONS *
1857              **************************/
1858
1859             r20              = _mm_mul_ps(rsq20,rinv20);
1860             r20              = _mm_andnot_ps(dummy_mask,r20);
1861
1862             /* EWALD ELECTROSTATICS */
1863
1864             /* Analytical PME correction */
1865             zeta2            = _mm_mul_ps(beta2,rsq20);
1866             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1867             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1868             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1869             felec            = _mm_mul_ps(qq20,felec);
1870
1871             fscal            = felec;
1872
1873             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1874
1875              /* Update vectorial force */
1876             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1877             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1878             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1879
1880             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1881             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1882             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1883
1884             /**************************
1885              * CALCULATE INTERACTIONS *
1886              **************************/
1887
1888             r21              = _mm_mul_ps(rsq21,rinv21);
1889             r21              = _mm_andnot_ps(dummy_mask,r21);
1890
1891             /* EWALD ELECTROSTATICS */
1892
1893             /* Analytical PME correction */
1894             zeta2            = _mm_mul_ps(beta2,rsq21);
1895             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1896             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1897             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1898             felec            = _mm_mul_ps(qq21,felec);
1899
1900             fscal            = felec;
1901
1902             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1903
1904              /* Update vectorial force */
1905             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1906             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1907             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1908
1909             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1910             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1911             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1912
1913             /**************************
1914              * CALCULATE INTERACTIONS *
1915              **************************/
1916
1917             r22              = _mm_mul_ps(rsq22,rinv22);
1918             r22              = _mm_andnot_ps(dummy_mask,r22);
1919
1920             /* EWALD ELECTROSTATICS */
1921
1922             /* Analytical PME correction */
1923             zeta2            = _mm_mul_ps(beta2,rsq22);
1924             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1925             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1926             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1927             felec            = _mm_mul_ps(qq22,felec);
1928
1929             fscal            = felec;
1930
1931             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1932
1933              /* Update vectorial force */
1934             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1935             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1936             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1937
1938             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1939             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1940             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1941
1942             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1943             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1944             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1945             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1946
1947             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1948                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1949
1950             /* Inner loop uses 268 flops */
1951         }
1952
1953         /* End of innermost loop */
1954
1955         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1956                                               f+i_coord_offset,fshift+i_shift_offset);
1957
1958         /* Increment number of inner iterations */
1959         inneriter                  += j_index_end - j_index_start;
1960
1961         /* Outer loop uses 18 flops */
1962     }
1963
1964     /* Increment number of outer iterations */
1965     outeriter        += nri;
1966
1967     /* Update outer/inner flops */
1968
1969     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*268);
1970 }