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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEw_VdwLJ_GeomW3W3_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  97     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  98     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     int              nvdwtype;
 107     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 108     int              *vdwtype;
 109     real             *vdwparam;
 110     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 111     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 112     __m128i          ewitab;
 113     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 114     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 115     real             *ewtab;
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 138     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 139     beta2            = _mm_mul_ps(beta,beta);
 140     beta3            = _mm_mul_ps(beta,beta2);
 141     ewtab            = fr->ic->tabq_coul_FDV0;
 142     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 143     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 148     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 149     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     jq0              = _mm_set1_ps(charge[inr+0]);
 153     jq1              = _mm_set1_ps(charge[inr+1]);
 154     jq2              = _mm_set1_ps(charge[inr+2]);
 155     vdwjidx0A        = 2*vdwtype[inr+0];
 156     qq00             = _mm_mul_ps(iq0,jq0);
 157     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 158     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 159     qq01             = _mm_mul_ps(iq0,jq1);
 160     qq02             = _mm_mul_ps(iq0,jq2);
 161     qq10             = _mm_mul_ps(iq1,jq0);
 162     qq11             = _mm_mul_ps(iq1,jq1);
 163     qq12             = _mm_mul_ps(iq1,jq2);
 164     qq20             = _mm_mul_ps(iq2,jq0);
 165     qq21             = _mm_mul_ps(iq2,jq1);
 166     qq22             = _mm_mul_ps(iq2,jq2);
 167
 168     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = jnrC = jnrD = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172     j_coord_offsetC = 0;
 173     j_coord_offsetD = 0;
 174
 175     outeriter        = 0;
 176     inneriter        = 0;
 177
 178     for(iidx=0;iidx<4*DIM;iidx++)
 179     {
 180         scratch[iidx] = 0.0;
 181     }
 182
 183     /* Start outer loop over neighborlists */
 184     for(iidx=0; iidx<nri; iidx++)
 185     {
 186         /* Load shift vector for this list */
 187         i_shift_offset   = DIM*shiftidx[iidx];
 188
 189         /* Load limits for loop over neighbors */
 190         j_index_start    = jindex[iidx];
 191         j_index_end      = jindex[iidx+1];
 192
 193         /* Get outer coordinate index */
 194         inr              = iinr[iidx];
 195         i_coord_offset   = DIM*inr;
 196
 197         /* Load i particle coords and add shift vector */
 198         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 199                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 200
 201         fix0             = _mm_setzero_ps();
 202         fiy0             = _mm_setzero_ps();
 203         fiz0             = _mm_setzero_ps();
 204         fix1             = _mm_setzero_ps();
 205         fiy1             = _mm_setzero_ps();
 206         fiz1             = _mm_setzero_ps();
 207         fix2             = _mm_setzero_ps();
 208         fiy2             = _mm_setzero_ps();
 209         fiz2             = _mm_setzero_ps();
 210
 211         /* Reset potential sums */
 212         velecsum         = _mm_setzero_ps();
 213         vvdwsum          = _mm_setzero_ps();
 214
 215         /* Start inner kernel loop */
 216         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 217         {
 218
 219             /* Get j neighbor index, and coordinate index */
 220             jnrA             = jjnr[jidx];
 221             jnrB             = jjnr[jidx+1];
 222             jnrC             = jjnr[jidx+2];
 223             jnrD             = jjnr[jidx+3];
 224             j_coord_offsetA  = DIM*jnrA;
 225             j_coord_offsetB  = DIM*jnrB;
 226             j_coord_offsetC  = DIM*jnrC;
 227             j_coord_offsetD  = DIM*jnrD;
 228
 229             /* load j atom coordinates */
 230             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 231                                               x+j_coord_offsetC,x+j_coord_offsetD,
 232                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 233
 234             /* Calculate displacement vector */
 235             dx00             = _mm_sub_ps(ix0,jx0);
 236             dy00             = _mm_sub_ps(iy0,jy0);
 237             dz00             = _mm_sub_ps(iz0,jz0);
 238             dx01             = _mm_sub_ps(ix0,jx1);
 239             dy01             = _mm_sub_ps(iy0,jy1);
 240             dz01             = _mm_sub_ps(iz0,jz1);
 241             dx02             = _mm_sub_ps(ix0,jx2);
 242             dy02             = _mm_sub_ps(iy0,jy2);
 243             dz02             = _mm_sub_ps(iz0,jz2);
 244             dx10             = _mm_sub_ps(ix1,jx0);
 245             dy10             = _mm_sub_ps(iy1,jy0);
 246             dz10             = _mm_sub_ps(iz1,jz0);
 247             dx11             = _mm_sub_ps(ix1,jx1);
 248             dy11             = _mm_sub_ps(iy1,jy1);
 249             dz11             = _mm_sub_ps(iz1,jz1);
 250             dx12             = _mm_sub_ps(ix1,jx2);
 251             dy12             = _mm_sub_ps(iy1,jy2);
 252             dz12             = _mm_sub_ps(iz1,jz2);
 253             dx20             = _mm_sub_ps(ix2,jx0);
 254             dy20             = _mm_sub_ps(iy2,jy0);
 255             dz20             = _mm_sub_ps(iz2,jz0);
 256             dx21             = _mm_sub_ps(ix2,jx1);
 257             dy21             = _mm_sub_ps(iy2,jy1);
 258             dz21             = _mm_sub_ps(iz2,jz1);
 259             dx22             = _mm_sub_ps(ix2,jx2);
 260             dy22             = _mm_sub_ps(iy2,jy2);
 261             dz22             = _mm_sub_ps(iz2,jz2);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 265             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 266             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 267             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 268             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 269             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 270             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 271             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 272             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 273
 274             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 275             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 276             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 277             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 278             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 279             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 280             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 281             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 282             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 283
 284             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 285             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 286             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 287             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 288             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 289             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 290             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 291             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 292             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 293
 294             fjx0             = _mm_setzero_ps();
 295             fjy0             = _mm_setzero_ps();
 296             fjz0             = _mm_setzero_ps();
 297             fjx1             = _mm_setzero_ps();
 298             fjy1             = _mm_setzero_ps();
 299             fjz1             = _mm_setzero_ps();
 300             fjx2             = _mm_setzero_ps();
 301             fjy2             = _mm_setzero_ps();
 302             fjz2             = _mm_setzero_ps();
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             r00              = _mm_mul_ps(rsq00,rinv00);
 309
 310             /* EWALD ELECTROSTATICS */
 311
 312             /* Analytical PME correction */
 313             zeta2            = _mm_mul_ps(beta2,rsq00);
 314             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 315             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 316             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 317             felec            = _mm_mul_ps(qq00,felec);
 318             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 319             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 320             velec            = _mm_mul_ps(qq00,velec);
 321
 322             /* LENNARD-JONES DISPERSION/REPULSION */
 323
 324             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 325             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 326             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 327             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 328             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             velecsum         = _mm_add_ps(velecsum,velec);
 332             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 333
 334             fscal            = _mm_add_ps(felec,fvdw);
 335
 336              /* Update vectorial force */
 337             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 338             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 339             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 340
 341             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 342             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 343             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             r01              = _mm_mul_ps(rsq01,rinv01);
 350
 351             /* EWALD ELECTROSTATICS */
 352
 353             /* Analytical PME correction */
 354             zeta2            = _mm_mul_ps(beta2,rsq01);
 355             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 356             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 357             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 358             felec            = _mm_mul_ps(qq01,felec);
 359             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 360             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 361             velec            = _mm_mul_ps(qq01,velec);
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368              /* Update vectorial force */
 369             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 370             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 371             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 372
 373             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 374             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 375             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 376
 377             /**************************
 378              * CALCULATE INTERACTIONS *
 379              **************************/
 380
 381             r02              = _mm_mul_ps(rsq02,rinv02);
 382
 383             /* EWALD ELECTROSTATICS */
 384
 385             /* Analytical PME correction */
 386             zeta2            = _mm_mul_ps(beta2,rsq02);
 387             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 388             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 389             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 390             felec            = _mm_mul_ps(qq02,felec);
 391             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 392             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 393             velec            = _mm_mul_ps(qq02,velec);
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velecsum         = _mm_add_ps(velecsum,velec);
 397
 398             fscal            = felec;
 399
 400              /* Update vectorial force */
 401             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 402             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 403             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 404
 405             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 406             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 407             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 408
 409             /**************************
 410              * CALCULATE INTERACTIONS *
 411              **************************/
 412
 413             r10              = _mm_mul_ps(rsq10,rinv10);
 414
 415             /* EWALD ELECTROSTATICS */
 416
 417             /* Analytical PME correction */
 418             zeta2            = _mm_mul_ps(beta2,rsq10);
 419             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 420             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 421             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 422             felec            = _mm_mul_ps(qq10,felec);
 423             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 424             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 425             velec            = _mm_mul_ps(qq10,velec);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velecsum         = _mm_add_ps(velecsum,velec);
 429
 430             fscal            = felec;
 431
 432              /* Update vectorial force */
 433             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 434             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 435             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 436
 437             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 438             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 439             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 440
 441             /**************************
 442              * CALCULATE INTERACTIONS *
 443              **************************/
 444
 445             r11              = _mm_mul_ps(rsq11,rinv11);
 446
 447             /* EWALD ELECTROSTATICS */
 448
 449             /* Analytical PME correction */
 450             zeta2            = _mm_mul_ps(beta2,rsq11);
 451             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 452             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 453             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 454             felec            = _mm_mul_ps(qq11,felec);
 455             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 456             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 457             velec            = _mm_mul_ps(qq11,velec);
 458
 459             /* Update potential sum for this i atom from the interaction with this j atom. */
 460             velecsum         = _mm_add_ps(velecsum,velec);
 461
 462             fscal            = felec;
 463
 464              /* Update vectorial force */
 465             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 466             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 467             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 468
 469             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 470             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 471             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 472
 473             /**************************
 474              * CALCULATE INTERACTIONS *
 475              **************************/
 476
 477             r12              = _mm_mul_ps(rsq12,rinv12);
 478
 479             /* EWALD ELECTROSTATICS */
 480
 481             /* Analytical PME correction */
 482             zeta2            = _mm_mul_ps(beta2,rsq12);
 483             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 484             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 485             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 486             felec            = _mm_mul_ps(qq12,felec);
 487             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 488             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 489             velec            = _mm_mul_ps(qq12,velec);
 490
 491             /* Update potential sum for this i atom from the interaction with this j atom. */
 492             velecsum         = _mm_add_ps(velecsum,velec);
 493
 494             fscal            = felec;
 495
 496              /* Update vectorial force */
 497             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 498             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 499             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 500
 501             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 502             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 503             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 504
 505             /**************************
 506              * CALCULATE INTERACTIONS *
 507              **************************/
 508
 509             r20              = _mm_mul_ps(rsq20,rinv20);
 510
 511             /* EWALD ELECTROSTATICS */
 512
 513             /* Analytical PME correction */
 514             zeta2            = _mm_mul_ps(beta2,rsq20);
 515             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 516             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 517             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 518             felec            = _mm_mul_ps(qq20,felec);
 519             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 520             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 521             velec            = _mm_mul_ps(qq20,velec);
 522
 523             /* Update potential sum for this i atom from the interaction with this j atom. */
 524             velecsum         = _mm_add_ps(velecsum,velec);
 525
 526             fscal            = felec;
 527
 528              /* Update vectorial force */
 529             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 530             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 531             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 532
 533             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 534             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 535             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 536
 537             /**************************
 538              * CALCULATE INTERACTIONS *
 539              **************************/
 540
 541             r21              = _mm_mul_ps(rsq21,rinv21);
 542
 543             /* EWALD ELECTROSTATICS */
 544
 545             /* Analytical PME correction */
 546             zeta2            = _mm_mul_ps(beta2,rsq21);
 547             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 548             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 549             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 550             felec            = _mm_mul_ps(qq21,felec);
 551             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 552             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 553             velec            = _mm_mul_ps(qq21,velec);
 554
 555             /* Update potential sum for this i atom from the interaction with this j atom. */
 556             velecsum         = _mm_add_ps(velecsum,velec);
 557
 558             fscal            = felec;
 559
 560              /* Update vectorial force */
 561             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 562             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 563             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 564
 565             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 566             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 567             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 568
 569             /**************************
 570              * CALCULATE INTERACTIONS *
 571              **************************/
 572
 573             r22              = _mm_mul_ps(rsq22,rinv22);
 574
 575             /* EWALD ELECTROSTATICS */
 576
 577             /* Analytical PME correction */
 578             zeta2            = _mm_mul_ps(beta2,rsq22);
 579             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 580             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 581             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 582             felec            = _mm_mul_ps(qq22,felec);
 583             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 584             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 585             velec            = _mm_mul_ps(qq22,velec);
 586
 587             /* Update potential sum for this i atom from the interaction with this j atom. */
 588             velecsum         = _mm_add_ps(velecsum,velec);
 589
 590             fscal            = felec;
 591
 592              /* Update vectorial force */
 593             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 594             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 595             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 596
 597             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 598             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 599             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 600
 601             fjptrA             = f+j_coord_offsetA;
 602             fjptrB             = f+j_coord_offsetB;
 603             fjptrC             = f+j_coord_offsetC;
 604             fjptrD             = f+j_coord_offsetD;
 605
 606             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 607                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 608
 609             /* Inner loop uses 273 flops */
 610         }
 611
 612         if(jidx<j_index_end)
 613         {
 614
 615             /* Get j neighbor index, and coordinate index */
 616             jnrlistA         = jjnr[jidx];
 617             jnrlistB         = jjnr[jidx+1];
 618             jnrlistC         = jjnr[jidx+2];
 619             jnrlistD         = jjnr[jidx+3];
 620             /* Sign of each element will be negative for non-real atoms.
 621              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 622              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 623              */
 624             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 625             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 626             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 627             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 628             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 629             j_coord_offsetA  = DIM*jnrA;
 630             j_coord_offsetB  = DIM*jnrB;
 631             j_coord_offsetC  = DIM*jnrC;
 632             j_coord_offsetD  = DIM*jnrD;
 633
 634             /* load j atom coordinates */
 635             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 636                                               x+j_coord_offsetC,x+j_coord_offsetD,
 637                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 638
 639             /* Calculate displacement vector */
 640             dx00             = _mm_sub_ps(ix0,jx0);
 641             dy00             = _mm_sub_ps(iy0,jy0);
 642             dz00             = _mm_sub_ps(iz0,jz0);
 643             dx01             = _mm_sub_ps(ix0,jx1);
 644             dy01             = _mm_sub_ps(iy0,jy1);
 645             dz01             = _mm_sub_ps(iz0,jz1);
 646             dx02             = _mm_sub_ps(ix0,jx2);
 647             dy02             = _mm_sub_ps(iy0,jy2);
 648             dz02             = _mm_sub_ps(iz0,jz2);
 649             dx10             = _mm_sub_ps(ix1,jx0);
 650             dy10             = _mm_sub_ps(iy1,jy0);
 651             dz10             = _mm_sub_ps(iz1,jz0);
 652             dx11             = _mm_sub_ps(ix1,jx1);
 653             dy11             = _mm_sub_ps(iy1,jy1);
 654             dz11             = _mm_sub_ps(iz1,jz1);
 655             dx12             = _mm_sub_ps(ix1,jx2);
 656             dy12             = _mm_sub_ps(iy1,jy2);
 657             dz12             = _mm_sub_ps(iz1,jz2);
 658             dx20             = _mm_sub_ps(ix2,jx0);
 659             dy20             = _mm_sub_ps(iy2,jy0);
 660             dz20             = _mm_sub_ps(iz2,jz0);
 661             dx21             = _mm_sub_ps(ix2,jx1);
 662             dy21             = _mm_sub_ps(iy2,jy1);
 663             dz21             = _mm_sub_ps(iz2,jz1);
 664             dx22             = _mm_sub_ps(ix2,jx2);
 665             dy22             = _mm_sub_ps(iy2,jy2);
 666             dz22             = _mm_sub_ps(iz2,jz2);
 667
 668             /* Calculate squared distance and things based on it */
 669             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 670             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 671             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 672             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 673             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 674             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 675             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 676             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 677             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 678
 679             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 680             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 681             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 682             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 683             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 684             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 685             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 686             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 687             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 688
 689             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 690             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 691             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 692             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 693             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 694             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 695             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 696             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 697             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 698
 699             fjx0             = _mm_setzero_ps();
 700             fjy0             = _mm_setzero_ps();
 701             fjz0             = _mm_setzero_ps();
 702             fjx1             = _mm_setzero_ps();
 703             fjy1             = _mm_setzero_ps();
 704             fjz1             = _mm_setzero_ps();
 705             fjx2             = _mm_setzero_ps();
 706             fjy2             = _mm_setzero_ps();
 707             fjz2             = _mm_setzero_ps();
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             r00              = _mm_mul_ps(rsq00,rinv00);
 714             r00              = _mm_andnot_ps(dummy_mask,r00);
 715
 716             /* EWALD ELECTROSTATICS */
 717
 718             /* Analytical PME correction */
 719             zeta2            = _mm_mul_ps(beta2,rsq00);
 720             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 721             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 722             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 723             felec            = _mm_mul_ps(qq00,felec);
 724             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 725             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 726             velec            = _mm_mul_ps(qq00,velec);
 727
 728             /* LENNARD-JONES DISPERSION/REPULSION */
 729
 730             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 731             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 732             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 733             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 734             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 735
 736             /* Update potential sum for this i atom from the interaction with this j atom. */
 737             velec            = _mm_andnot_ps(dummy_mask,velec);
 738             velecsum         = _mm_add_ps(velecsum,velec);
 739             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 740             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 741
 742             fscal            = _mm_add_ps(felec,fvdw);
 743
 744             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 745
 746              /* Update vectorial force */
 747             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 748             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 749             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 750
 751             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 752             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 753             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             r01              = _mm_mul_ps(rsq01,rinv01);
 760             r01              = _mm_andnot_ps(dummy_mask,r01);
 761
 762             /* EWALD ELECTROSTATICS */
 763
 764             /* Analytical PME correction */
 765             zeta2            = _mm_mul_ps(beta2,rsq01);
 766             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 767             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 768             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 769             felec            = _mm_mul_ps(qq01,felec);
 770             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 771             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv01);
 772             velec            = _mm_mul_ps(qq01,velec);
 773
 774             /* Update potential sum for this i atom from the interaction with this j atom. */
 775             velec            = _mm_andnot_ps(dummy_mask,velec);
 776             velecsum         = _mm_add_ps(velecsum,velec);
 777
 778             fscal            = felec;
 779
 780             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 781
 782              /* Update vectorial force */
 783             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 784             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 785             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 786
 787             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 788             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 789             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 790
 791             /**************************
 792              * CALCULATE INTERACTIONS *
 793              **************************/
 794
 795             r02              = _mm_mul_ps(rsq02,rinv02);
 796             r02              = _mm_andnot_ps(dummy_mask,r02);
 797
 798             /* EWALD ELECTROSTATICS */
 799
 800             /* Analytical PME correction */
 801             zeta2            = _mm_mul_ps(beta2,rsq02);
 802             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 803             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 804             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 805             felec            = _mm_mul_ps(qq02,felec);
 806             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 807             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv02);
 808             velec            = _mm_mul_ps(qq02,velec);
 809
 810             /* Update potential sum for this i atom from the interaction with this j atom. */
 811             velec            = _mm_andnot_ps(dummy_mask,velec);
 812             velecsum         = _mm_add_ps(velecsum,velec);
 813
 814             fscal            = felec;
 815
 816             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 817
 818              /* Update vectorial force */
 819             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 820             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 821             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 822
 823             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 824             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 825             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             r10              = _mm_mul_ps(rsq10,rinv10);
 832             r10              = _mm_andnot_ps(dummy_mask,r10);
 833
 834             /* EWALD ELECTROSTATICS */
 835
 836             /* Analytical PME correction */
 837             zeta2            = _mm_mul_ps(beta2,rsq10);
 838             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 839             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 840             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 841             felec            = _mm_mul_ps(qq10,felec);
 842             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 843             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 844             velec            = _mm_mul_ps(qq10,velec);
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             velec            = _mm_andnot_ps(dummy_mask,velec);
 848             velecsum         = _mm_add_ps(velecsum,velec);
 849
 850             fscal            = felec;
 851
 852             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 853
 854              /* Update vectorial force */
 855             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 856             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 857             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 858
 859             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 860             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 861             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 862
 863             /**************************
 864              * CALCULATE INTERACTIONS *
 865              **************************/
 866
 867             r11              = _mm_mul_ps(rsq11,rinv11);
 868             r11              = _mm_andnot_ps(dummy_mask,r11);
 869
 870             /* EWALD ELECTROSTATICS */
 871
 872             /* Analytical PME correction */
 873             zeta2            = _mm_mul_ps(beta2,rsq11);
 874             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 875             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 876             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 877             felec            = _mm_mul_ps(qq11,felec);
 878             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 879             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 880             velec            = _mm_mul_ps(qq11,velec);
 881
 882             /* Update potential sum for this i atom from the interaction with this j atom. */
 883             velec            = _mm_andnot_ps(dummy_mask,velec);
 884             velecsum         = _mm_add_ps(velecsum,velec);
 885
 886             fscal            = felec;
 887
 888             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 889
 890              /* Update vectorial force */
 891             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 892             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 893             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 894
 895             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 896             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 897             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 898
 899             /**************************
 900              * CALCULATE INTERACTIONS *
 901              **************************/
 902
 903             r12              = _mm_mul_ps(rsq12,rinv12);
 904             r12              = _mm_andnot_ps(dummy_mask,r12);
 905
 906             /* EWALD ELECTROSTATICS */
 907
 908             /* Analytical PME correction */
 909             zeta2            = _mm_mul_ps(beta2,rsq12);
 910             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 911             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 912             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 913             felec            = _mm_mul_ps(qq12,felec);
 914             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 915             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 916             velec            = _mm_mul_ps(qq12,velec);
 917
 918             /* Update potential sum for this i atom from the interaction with this j atom. */
 919             velec            = _mm_andnot_ps(dummy_mask,velec);
 920             velecsum         = _mm_add_ps(velecsum,velec);
 921
 922             fscal            = felec;
 923
 924             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 925
 926              /* Update vectorial force */
 927             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 928             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 929             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 930
 931             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 932             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 933             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             r20              = _mm_mul_ps(rsq20,rinv20);
 940             r20              = _mm_andnot_ps(dummy_mask,r20);
 941
 942             /* EWALD ELECTROSTATICS */
 943
 944             /* Analytical PME correction */
 945             zeta2            = _mm_mul_ps(beta2,rsq20);
 946             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 947             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 948             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 949             felec            = _mm_mul_ps(qq20,felec);
 950             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 951             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 952             velec            = _mm_mul_ps(qq20,velec);
 953
 954             /* Update potential sum for this i atom from the interaction with this j atom. */
 955             velec            = _mm_andnot_ps(dummy_mask,velec);
 956             velecsum         = _mm_add_ps(velecsum,velec);
 957
 958             fscal            = felec;
 959
 960             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 961
 962              /* Update vectorial force */
 963             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 964             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 965             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 966
 967             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 968             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 969             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 970
 971             /**************************
 972              * CALCULATE INTERACTIONS *
 973              **************************/
 974
 975             r21              = _mm_mul_ps(rsq21,rinv21);
 976             r21              = _mm_andnot_ps(dummy_mask,r21);
 977
 978             /* EWALD ELECTROSTATICS */
 979
 980             /* Analytical PME correction */
 981             zeta2            = _mm_mul_ps(beta2,rsq21);
 982             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 983             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 984             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 985             felec            = _mm_mul_ps(qq21,felec);
 986             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 987             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 988             velec            = _mm_mul_ps(qq21,velec);
 989
 990             /* Update potential sum for this i atom from the interaction with this j atom. */
 991             velec            = _mm_andnot_ps(dummy_mask,velec);
 992             velecsum         = _mm_add_ps(velecsum,velec);
 993
 994             fscal            = felec;
 995
 996             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 997
 998              /* Update vectorial force */
 999             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1000             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1001             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1002
1003             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1004             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1005             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1006
1007             /**************************
1008              * CALCULATE INTERACTIONS *
1009              **************************/
1010
1011             r22              = _mm_mul_ps(rsq22,rinv22);
1012             r22              = _mm_andnot_ps(dummy_mask,r22);
1013
1014             /* EWALD ELECTROSTATICS */
1015
1016             /* Analytical PME correction */
1017             zeta2            = _mm_mul_ps(beta2,rsq22);
1018             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1019             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1020             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1021             felec            = _mm_mul_ps(qq22,felec);
1022             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1023             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
1024             velec            = _mm_mul_ps(qq22,velec);
1025
1026             /* Update potential sum for this i atom from the interaction with this j atom. */
1027             velec            = _mm_andnot_ps(dummy_mask,velec);
1028             velecsum         = _mm_add_ps(velecsum,velec);
1029
1030             fscal            = felec;
1031
1032             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1033
1034              /* Update vectorial force */
1035             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1036             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1037             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1038
1039             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1040             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1041             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1042
1043             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1044             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1045             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1046             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1047
1048             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1049                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1050
1051             /* Inner loop uses 282 flops */
1052         }
1053
1054         /* End of innermost loop */
1055
1056         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1057                                               f+i_coord_offset,fshift+i_shift_offset);
1058
1059         ggid                        = gid[iidx];
1060         /* Update potential energies */
1061         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1062         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1063
1064         /* Increment number of inner iterations */
1065         inneriter                  += j_index_end - j_index_start;
1066
1067         /* Outer loop uses 20 flops */
1068     }
1069
1070     /* Increment number of outer iterations */
1071     outeriter        += nri;
1072
1073     /* Update outer/inner flops */
1074
1075     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*282);
1076 }
1077 /*
1078  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_avx_128_fma_single
1079  * Electrostatics interaction: Ewald
1080  * VdW interaction:            LennardJones
1081  * Geometry:                   Water3-Water3
1082  * Calculate force/pot:        Force
1083  */
1084 void
1085 nb_kernel_ElecEw_VdwLJ_GeomW3W3_F_avx_128_fma_single
1086                     (t_nblist                    * gmx_restrict       nlist,
1087                      rvec                        * gmx_restrict          xx,
1088                      rvec                        * gmx_restrict          ff,
1089                      t_forcerec                  * gmx_restrict          fr,
1090                      t_mdatoms                   * gmx_restrict     mdatoms,
1091                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1092                      t_nrnb                      * gmx_restrict        nrnb)
1093 {
1094     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1095      * just 0 for non-waters.
1096      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1097      * jnr indices corresponding to data put in the four positions in the SIMD register.
1098      */
1099     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1100     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1101     int              jnrA,jnrB,jnrC,jnrD;
1102     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1103     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1104     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1105     real             rcutoff_scalar;
1106     real             *shiftvec,*fshift,*x,*f;
1107     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1108     real             scratch[4*DIM];
1109     __m128           fscal,rcutoff,rcutoff2,jidxall;
1110     int              vdwioffset0;
1111     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1112     int              vdwioffset1;
1113     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1114     int              vdwioffset2;
1115     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1116     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1117     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1118     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1119     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1120     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1121     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1122     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1123     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1124     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1125     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1126     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1127     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1128     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1129     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1130     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1131     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1132     real             *charge;
1133     int              nvdwtype;
1134     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1135     int              *vdwtype;
1136     real             *vdwparam;
1137     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1138     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1139     __m128i          ewitab;
1140     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1141     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1142     real             *ewtab;
1143     __m128           dummy_mask,cutoff_mask;
1144     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1145     __m128           one     = _mm_set1_ps(1.0);
1146     __m128           two     = _mm_set1_ps(2.0);
1147     x                = xx[0];
1148     f                = ff[0];
1149
1150     nri              = nlist->nri;
1151     iinr             = nlist->iinr;
1152     jindex           = nlist->jindex;
1153     jjnr             = nlist->jjnr;
1154     shiftidx         = nlist->shift;
1155     gid              = nlist->gid;
1156     shiftvec         = fr->shift_vec[0];
1157     fshift           = fr->fshift[0];
1158     facel            = _mm_set1_ps(fr->epsfac);
1159     charge           = mdatoms->chargeA;
1160     nvdwtype         = fr->ntype;
1161     vdwparam         = fr->nbfp;
1162     vdwtype          = mdatoms->typeA;
1163
1164     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1165     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1166     beta2            = _mm_mul_ps(beta,beta);
1167     beta3            = _mm_mul_ps(beta,beta2);
1168     ewtab            = fr->ic->tabq_coul_F;
1169     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1170     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1171
1172     /* Setup water-specific parameters */
1173     inr              = nlist->iinr[0];
1174     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1175     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1176     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1177     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1178
1179     jq0              = _mm_set1_ps(charge[inr+0]);
1180     jq1              = _mm_set1_ps(charge[inr+1]);
1181     jq2              = _mm_set1_ps(charge[inr+2]);
1182     vdwjidx0A        = 2*vdwtype[inr+0];
1183     qq00             = _mm_mul_ps(iq0,jq0);
1184     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1185     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1186     qq01             = _mm_mul_ps(iq0,jq1);
1187     qq02             = _mm_mul_ps(iq0,jq2);
1188     qq10             = _mm_mul_ps(iq1,jq0);
1189     qq11             = _mm_mul_ps(iq1,jq1);
1190     qq12             = _mm_mul_ps(iq1,jq2);
1191     qq20             = _mm_mul_ps(iq2,jq0);
1192     qq21             = _mm_mul_ps(iq2,jq1);
1193     qq22             = _mm_mul_ps(iq2,jq2);
1194
1195     /* Avoid stupid compiler warnings */
1196     jnrA = jnrB = jnrC = jnrD = 0;
1197     j_coord_offsetA = 0;
1198     j_coord_offsetB = 0;
1199     j_coord_offsetC = 0;
1200     j_coord_offsetD = 0;
1201
1202     outeriter        = 0;
1203     inneriter        = 0;
1204
1205     for(iidx=0;iidx<4*DIM;iidx++)
1206     {
1207         scratch[iidx] = 0.0;
1208     }
1209
1210     /* Start outer loop over neighborlists */
1211     for(iidx=0; iidx<nri; iidx++)
1212     {
1213         /* Load shift vector for this list */
1214         i_shift_offset   = DIM*shiftidx[iidx];
1215
1216         /* Load limits for loop over neighbors */
1217         j_index_start    = jindex[iidx];
1218         j_index_end      = jindex[iidx+1];
1219
1220         /* Get outer coordinate index */
1221         inr              = iinr[iidx];
1222         i_coord_offset   = DIM*inr;
1223
1224         /* Load i particle coords and add shift vector */
1225         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1226                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1227
1228         fix0             = _mm_setzero_ps();
1229         fiy0             = _mm_setzero_ps();
1230         fiz0             = _mm_setzero_ps();
1231         fix1             = _mm_setzero_ps();
1232         fiy1             = _mm_setzero_ps();
1233         fiz1             = _mm_setzero_ps();
1234         fix2             = _mm_setzero_ps();
1235         fiy2             = _mm_setzero_ps();
1236         fiz2             = _mm_setzero_ps();
1237
1238         /* Start inner kernel loop */
1239         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1240         {
1241
1242             /* Get j neighbor index, and coordinate index */
1243             jnrA             = jjnr[jidx];
1244             jnrB             = jjnr[jidx+1];
1245             jnrC             = jjnr[jidx+2];
1246             jnrD             = jjnr[jidx+3];
1247             j_coord_offsetA  = DIM*jnrA;
1248             j_coord_offsetB  = DIM*jnrB;
1249             j_coord_offsetC  = DIM*jnrC;
1250             j_coord_offsetD  = DIM*jnrD;
1251
1252             /* load j atom coordinates */
1253             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1254                                               x+j_coord_offsetC,x+j_coord_offsetD,
1255                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1256
1257             /* Calculate displacement vector */
1258             dx00             = _mm_sub_ps(ix0,jx0);
1259             dy00             = _mm_sub_ps(iy0,jy0);
1260             dz00             = _mm_sub_ps(iz0,jz0);
1261             dx01             = _mm_sub_ps(ix0,jx1);
1262             dy01             = _mm_sub_ps(iy0,jy1);
1263             dz01             = _mm_sub_ps(iz0,jz1);
1264             dx02             = _mm_sub_ps(ix0,jx2);
1265             dy02             = _mm_sub_ps(iy0,jy2);
1266             dz02             = _mm_sub_ps(iz0,jz2);
1267             dx10             = _mm_sub_ps(ix1,jx0);
1268             dy10             = _mm_sub_ps(iy1,jy0);
1269             dz10             = _mm_sub_ps(iz1,jz0);
1270             dx11             = _mm_sub_ps(ix1,jx1);
1271             dy11             = _mm_sub_ps(iy1,jy1);
1272             dz11             = _mm_sub_ps(iz1,jz1);
1273             dx12             = _mm_sub_ps(ix1,jx2);
1274             dy12             = _mm_sub_ps(iy1,jy2);
1275             dz12             = _mm_sub_ps(iz1,jz2);
1276             dx20             = _mm_sub_ps(ix2,jx0);
1277             dy20             = _mm_sub_ps(iy2,jy0);
1278             dz20             = _mm_sub_ps(iz2,jz0);
1279             dx21             = _mm_sub_ps(ix2,jx1);
1280             dy21             = _mm_sub_ps(iy2,jy1);
1281             dz21             = _mm_sub_ps(iz2,jz1);
1282             dx22             = _mm_sub_ps(ix2,jx2);
1283             dy22             = _mm_sub_ps(iy2,jy2);
1284             dz22             = _mm_sub_ps(iz2,jz2);
1285
1286             /* Calculate squared distance and things based on it */
1287             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1288             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1289             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1290             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1291             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1292             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1293             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1294             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1295             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1296
1297             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1298             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1299             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1300             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1301             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1302             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1303             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1304             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1305             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1306
1307             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1308             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1309             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1310             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1311             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1312             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1313             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1314             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1315             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1316
1317             fjx0             = _mm_setzero_ps();
1318             fjy0             = _mm_setzero_ps();
1319             fjz0             = _mm_setzero_ps();
1320             fjx1             = _mm_setzero_ps();
1321             fjy1             = _mm_setzero_ps();
1322             fjz1             = _mm_setzero_ps();
1323             fjx2             = _mm_setzero_ps();
1324             fjy2             = _mm_setzero_ps();
1325             fjz2             = _mm_setzero_ps();
1326
1327             /**************************
1328              * CALCULATE INTERACTIONS *
1329              **************************/
1330
1331             r00              = _mm_mul_ps(rsq00,rinv00);
1332
1333             /* EWALD ELECTROSTATICS */
1334
1335             /* Analytical PME correction */
1336             zeta2            = _mm_mul_ps(beta2,rsq00);
1337             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1338             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1339             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1340             felec            = _mm_mul_ps(qq00,felec);
1341
1342             /* LENNARD-JONES DISPERSION/REPULSION */
1343
1344             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1345             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1346
1347             fscal            = _mm_add_ps(felec,fvdw);
1348
1349              /* Update vectorial force */
1350             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1351             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1352             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1353
1354             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1355             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1356             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             r01              = _mm_mul_ps(rsq01,rinv01);
1363
1364             /* EWALD ELECTROSTATICS */
1365
1366             /* Analytical PME correction */
1367             zeta2            = _mm_mul_ps(beta2,rsq01);
1368             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1369             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1370             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1371             felec            = _mm_mul_ps(qq01,felec);
1372
1373             fscal            = felec;
1374
1375              /* Update vectorial force */
1376             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1377             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1378             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1379
1380             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1381             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1382             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             r02              = _mm_mul_ps(rsq02,rinv02);
1389
1390             /* EWALD ELECTROSTATICS */
1391
1392             /* Analytical PME correction */
1393             zeta2            = _mm_mul_ps(beta2,rsq02);
1394             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1395             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1396             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1397             felec            = _mm_mul_ps(qq02,felec);
1398
1399             fscal            = felec;
1400
1401              /* Update vectorial force */
1402             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1403             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1404             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1405
1406             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1407             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1408             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1409
1410             /**************************
1411              * CALCULATE INTERACTIONS *
1412              **************************/
1413
1414             r10              = _mm_mul_ps(rsq10,rinv10);
1415
1416             /* EWALD ELECTROSTATICS */
1417
1418             /* Analytical PME correction */
1419             zeta2            = _mm_mul_ps(beta2,rsq10);
1420             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1421             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1422             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1423             felec            = _mm_mul_ps(qq10,felec);
1424
1425             fscal            = felec;
1426
1427              /* Update vectorial force */
1428             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1429             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1430             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1431
1432             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1433             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1434             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1435
1436             /**************************
1437              * CALCULATE INTERACTIONS *
1438              **************************/
1439
1440             r11              = _mm_mul_ps(rsq11,rinv11);
1441
1442             /* EWALD ELECTROSTATICS */
1443
1444             /* Analytical PME correction */
1445             zeta2            = _mm_mul_ps(beta2,rsq11);
1446             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1447             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1448             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1449             felec            = _mm_mul_ps(qq11,felec);
1450
1451             fscal            = felec;
1452
1453              /* Update vectorial force */
1454             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1455             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1456             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1457
1458             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1459             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1460             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             r12              = _mm_mul_ps(rsq12,rinv12);
1467
1468             /* EWALD ELECTROSTATICS */
1469
1470             /* Analytical PME correction */
1471             zeta2            = _mm_mul_ps(beta2,rsq12);
1472             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1473             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1474             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1475             felec            = _mm_mul_ps(qq12,felec);
1476
1477             fscal            = felec;
1478
1479              /* Update vectorial force */
1480             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1481             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1482             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1483
1484             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1485             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1486             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1487
1488             /**************************
1489              * CALCULATE INTERACTIONS *
1490              **************************/
1491
1492             r20              = _mm_mul_ps(rsq20,rinv20);
1493
1494             /* EWALD ELECTROSTATICS */
1495
1496             /* Analytical PME correction */
1497             zeta2            = _mm_mul_ps(beta2,rsq20);
1498             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1499             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1500             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1501             felec            = _mm_mul_ps(qq20,felec);
1502
1503             fscal            = felec;
1504
1505              /* Update vectorial force */
1506             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1507             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1508             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1509
1510             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1511             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1512             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1513
1514             /**************************
1515              * CALCULATE INTERACTIONS *
1516              **************************/
1517
1518             r21              = _mm_mul_ps(rsq21,rinv21);
1519
1520             /* EWALD ELECTROSTATICS */
1521
1522             /* Analytical PME correction */
1523             zeta2            = _mm_mul_ps(beta2,rsq21);
1524             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1525             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1526             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1527             felec            = _mm_mul_ps(qq21,felec);
1528
1529             fscal            = felec;
1530
1531              /* Update vectorial force */
1532             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1533             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1534             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1535
1536             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1537             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1538             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1539
1540             /**************************
1541              * CALCULATE INTERACTIONS *
1542              **************************/
1543
1544             r22              = _mm_mul_ps(rsq22,rinv22);
1545
1546             /* EWALD ELECTROSTATICS */
1547
1548             /* Analytical PME correction */
1549             zeta2            = _mm_mul_ps(beta2,rsq22);
1550             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1551             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1552             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1553             felec            = _mm_mul_ps(qq22,felec);
1554
1555             fscal            = felec;
1556
1557              /* Update vectorial force */
1558             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1559             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1560             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1561
1562             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1563             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1564             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1565
1566             fjptrA             = f+j_coord_offsetA;
1567             fjptrB             = f+j_coord_offsetB;
1568             fjptrC             = f+j_coord_offsetC;
1569             fjptrD             = f+j_coord_offsetD;
1570
1571             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1572                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1573
1574             /* Inner loop uses 259 flops */
1575         }
1576
1577         if(jidx<j_index_end)
1578         {
1579
1580             /* Get j neighbor index, and coordinate index */
1581             jnrlistA         = jjnr[jidx];
1582             jnrlistB         = jjnr[jidx+1];
1583             jnrlistC         = jjnr[jidx+2];
1584             jnrlistD         = jjnr[jidx+3];
1585             /* Sign of each element will be negative for non-real atoms.
1586              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1587              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1588              */
1589             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1590             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1591             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1592             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1593             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1594             j_coord_offsetA  = DIM*jnrA;
1595             j_coord_offsetB  = DIM*jnrB;
1596             j_coord_offsetC  = DIM*jnrC;
1597             j_coord_offsetD  = DIM*jnrD;
1598
1599             /* load j atom coordinates */
1600             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1601                                               x+j_coord_offsetC,x+j_coord_offsetD,
1602                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1603
1604             /* Calculate displacement vector */
1605             dx00             = _mm_sub_ps(ix0,jx0);
1606             dy00             = _mm_sub_ps(iy0,jy0);
1607             dz00             = _mm_sub_ps(iz0,jz0);
1608             dx01             = _mm_sub_ps(ix0,jx1);
1609             dy01             = _mm_sub_ps(iy0,jy1);
1610             dz01             = _mm_sub_ps(iz0,jz1);
1611             dx02             = _mm_sub_ps(ix0,jx2);
1612             dy02             = _mm_sub_ps(iy0,jy2);
1613             dz02             = _mm_sub_ps(iz0,jz2);
1614             dx10             = _mm_sub_ps(ix1,jx0);
1615             dy10             = _mm_sub_ps(iy1,jy0);
1616             dz10             = _mm_sub_ps(iz1,jz0);
1617             dx11             = _mm_sub_ps(ix1,jx1);
1618             dy11             = _mm_sub_ps(iy1,jy1);
1619             dz11             = _mm_sub_ps(iz1,jz1);
1620             dx12             = _mm_sub_ps(ix1,jx2);
1621             dy12             = _mm_sub_ps(iy1,jy2);
1622             dz12             = _mm_sub_ps(iz1,jz2);
1623             dx20             = _mm_sub_ps(ix2,jx0);
1624             dy20             = _mm_sub_ps(iy2,jy0);
1625             dz20             = _mm_sub_ps(iz2,jz0);
1626             dx21             = _mm_sub_ps(ix2,jx1);
1627             dy21             = _mm_sub_ps(iy2,jy1);
1628             dz21             = _mm_sub_ps(iz2,jz1);
1629             dx22             = _mm_sub_ps(ix2,jx2);
1630             dy22             = _mm_sub_ps(iy2,jy2);
1631             dz22             = _mm_sub_ps(iz2,jz2);
1632
1633             /* Calculate squared distance and things based on it */
1634             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1635             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1636             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1637             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1638             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1639             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1640             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1641             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1642             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1643
1644             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1645             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1646             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1647             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1648             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1649             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1650             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1651             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1652             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1653
1654             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1655             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1656             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1657             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1658             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1659             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1660             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1661             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1662             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1663
1664             fjx0             = _mm_setzero_ps();
1665             fjy0             = _mm_setzero_ps();
1666             fjz0             = _mm_setzero_ps();
1667             fjx1             = _mm_setzero_ps();
1668             fjy1             = _mm_setzero_ps();
1669             fjz1             = _mm_setzero_ps();
1670             fjx2             = _mm_setzero_ps();
1671             fjy2             = _mm_setzero_ps();
1672             fjz2             = _mm_setzero_ps();
1673
1674             /**************************
1675              * CALCULATE INTERACTIONS *
1676              **************************/
1677
1678             r00              = _mm_mul_ps(rsq00,rinv00);
1679             r00              = _mm_andnot_ps(dummy_mask,r00);
1680
1681             /* EWALD ELECTROSTATICS */
1682
1683             /* Analytical PME correction */
1684             zeta2            = _mm_mul_ps(beta2,rsq00);
1685             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1686             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1687             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1688             felec            = _mm_mul_ps(qq00,felec);
1689
1690             /* LENNARD-JONES DISPERSION/REPULSION */
1691
1692             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1693             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1694
1695             fscal            = _mm_add_ps(felec,fvdw);
1696
1697             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1698
1699              /* Update vectorial force */
1700             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1701             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1702             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1703
1704             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1705             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1706             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1707
1708             /**************************
1709              * CALCULATE INTERACTIONS *
1710              **************************/
1711
1712             r01              = _mm_mul_ps(rsq01,rinv01);
1713             r01              = _mm_andnot_ps(dummy_mask,r01);
1714
1715             /* EWALD ELECTROSTATICS */
1716
1717             /* Analytical PME correction */
1718             zeta2            = _mm_mul_ps(beta2,rsq01);
1719             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1720             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1721             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1722             felec            = _mm_mul_ps(qq01,felec);
1723
1724             fscal            = felec;
1725
1726             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1727
1728              /* Update vectorial force */
1729             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1730             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1731             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1732
1733             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1734             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1735             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1736
1737             /**************************
1738              * CALCULATE INTERACTIONS *
1739              **************************/
1740
1741             r02              = _mm_mul_ps(rsq02,rinv02);
1742             r02              = _mm_andnot_ps(dummy_mask,r02);
1743
1744             /* EWALD ELECTROSTATICS */
1745
1746             /* Analytical PME correction */
1747             zeta2            = _mm_mul_ps(beta2,rsq02);
1748             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1749             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1750             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1751             felec            = _mm_mul_ps(qq02,felec);
1752
1753             fscal            = felec;
1754
1755             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1756
1757              /* Update vectorial force */
1758             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1759             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1760             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1761
1762             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1763             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1764             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1765
1766             /**************************
1767              * CALCULATE INTERACTIONS *
1768              **************************/
1769
1770             r10              = _mm_mul_ps(rsq10,rinv10);
1771             r10              = _mm_andnot_ps(dummy_mask,r10);
1772
1773             /* EWALD ELECTROSTATICS */
1774
1775             /* Analytical PME correction */
1776             zeta2            = _mm_mul_ps(beta2,rsq10);
1777             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1778             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1779             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1780             felec            = _mm_mul_ps(qq10,felec);
1781
1782             fscal            = felec;
1783
1784             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1785
1786              /* Update vectorial force */
1787             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1788             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1789             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1790
1791             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1792             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1793             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1794
1795             /**************************
1796              * CALCULATE INTERACTIONS *
1797              **************************/
1798
1799             r11              = _mm_mul_ps(rsq11,rinv11);
1800             r11              = _mm_andnot_ps(dummy_mask,r11);
1801
1802             /* EWALD ELECTROSTATICS */
1803
1804             /* Analytical PME correction */
1805             zeta2            = _mm_mul_ps(beta2,rsq11);
1806             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1807             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1808             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1809             felec            = _mm_mul_ps(qq11,felec);
1810
1811             fscal            = felec;
1812
1813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1814
1815              /* Update vectorial force */
1816             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1817             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1818             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1819
1820             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1821             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1822             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1823
1824             /**************************
1825              * CALCULATE INTERACTIONS *
1826              **************************/
1827
1828             r12              = _mm_mul_ps(rsq12,rinv12);
1829             r12              = _mm_andnot_ps(dummy_mask,r12);
1830
1831             /* EWALD ELECTROSTATICS */
1832
1833             /* Analytical PME correction */
1834             zeta2            = _mm_mul_ps(beta2,rsq12);
1835             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1836             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1837             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1838             felec            = _mm_mul_ps(qq12,felec);
1839
1840             fscal            = felec;
1841
1842             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1843
1844              /* Update vectorial force */
1845             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1846             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1847             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1848
1849             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1850             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1851             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1852
1853             /**************************
1854              * CALCULATE INTERACTIONS *
1855              **************************/
1856
1857             r20              = _mm_mul_ps(rsq20,rinv20);
1858             r20              = _mm_andnot_ps(dummy_mask,r20);
1859
1860             /* EWALD ELECTROSTATICS */
1861
1862             /* Analytical PME correction */
1863             zeta2            = _mm_mul_ps(beta2,rsq20);
1864             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1865             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1866             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1867             felec            = _mm_mul_ps(qq20,felec);
1868
1869             fscal            = felec;
1870
1871             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1872
1873              /* Update vectorial force */
1874             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1875             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1876             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1877
1878             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1879             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1880             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1881
1882             /**************************
1883              * CALCULATE INTERACTIONS *
1884              **************************/
1885
1886             r21              = _mm_mul_ps(rsq21,rinv21);
1887             r21              = _mm_andnot_ps(dummy_mask,r21);
1888
1889             /* EWALD ELECTROSTATICS */
1890
1891             /* Analytical PME correction */
1892             zeta2            = _mm_mul_ps(beta2,rsq21);
1893             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1894             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1895             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1896             felec            = _mm_mul_ps(qq21,felec);
1897
1898             fscal            = felec;
1899
1900             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1901
1902              /* Update vectorial force */
1903             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1904             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1905             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1906
1907             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1908             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1909             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1910
1911             /**************************
1912              * CALCULATE INTERACTIONS *
1913              **************************/
1914
1915             r22              = _mm_mul_ps(rsq22,rinv22);
1916             r22              = _mm_andnot_ps(dummy_mask,r22);
1917
1918             /* EWALD ELECTROSTATICS */
1919
1920             /* Analytical PME correction */
1921             zeta2            = _mm_mul_ps(beta2,rsq22);
1922             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1923             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1924             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1925             felec            = _mm_mul_ps(qq22,felec);
1926
1927             fscal            = felec;
1928
1929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1930
1931              /* Update vectorial force */
1932             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1933             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1934             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1935
1936             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1937             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1938             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1939
1940             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1941             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1942             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1943             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1944
1945             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1946                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1947
1948             /* Inner loop uses 268 flops */
1949         }
1950
1951         /* End of innermost loop */
1952
1953         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1954                                               f+i_coord_offset,fshift+i_shift_offset);
1955
1956         /* Increment number of inner iterations */
1957         inneriter                  += j_index_end - j_index_start;
1958
1959         /* Outer loop uses 18 flops */
1960     }
1961
1962     /* Increment number of outer iterations */
1963     outeriter        += nri;
1964
1965     /* Update outer/inner flops */
1966
1967     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*268);
1968 }