src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwNone_GeomW4W4_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              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              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  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     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  96     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  97     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     __m128i          ewitab;
 109     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 110     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 111     real             *ewtab;
 112     __m128           dummy_mask,cutoff_mask;
 113     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 114     __m128           one     = _mm_set1_ps(1.0);
 115     __m128           two     = _mm_set1_ps(2.0);
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = _mm_set1_ps(fr->epsfac);
 128     charge           = mdatoms->chargeA;
 129
 130     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 131     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 132     beta2            = _mm_mul_ps(beta,beta);
 133     beta3            = _mm_mul_ps(beta,beta2);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 141     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 142     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 143
 144     jq1              = _mm_set1_ps(charge[inr+1]);
 145     jq2              = _mm_set1_ps(charge[inr+2]);
 146     jq3              = _mm_set1_ps(charge[inr+3]);
 147     qq11             = _mm_mul_ps(iq1,jq1);
 148     qq12             = _mm_mul_ps(iq1,jq2);
 149     qq13             = _mm_mul_ps(iq1,jq3);
 150     qq21             = _mm_mul_ps(iq2,jq1);
 151     qq22             = _mm_mul_ps(iq2,jq2);
 152     qq23             = _mm_mul_ps(iq2,jq3);
 153     qq31             = _mm_mul_ps(iq3,jq1);
 154     qq32             = _mm_mul_ps(iq3,jq2);
 155     qq33             = _mm_mul_ps(iq3,jq3);
 156
 157     /* Avoid stupid compiler warnings */
 158     jnrA = jnrB = jnrC = jnrD = 0;
 159     j_coord_offsetA = 0;
 160     j_coord_offsetB = 0;
 161     j_coord_offsetC = 0;
 162     j_coord_offsetD = 0;
 163
 164     outeriter        = 0;
 165     inneriter        = 0;
 166
 167     for(iidx=0;iidx<4*DIM;iidx++)
 168     {
 169         scratch[iidx] = 0.0;
 170     }
 171
 172     /* Start outer loop over neighborlists */
 173     for(iidx=0; iidx<nri; iidx++)
 174     {
 175         /* Load shift vector for this list */
 176         i_shift_offset   = DIM*shiftidx[iidx];
 177
 178         /* Load limits for loop over neighbors */
 179         j_index_start    = jindex[iidx];
 180         j_index_end      = jindex[iidx+1];
 181
 182         /* Get outer coordinate index */
 183         inr              = iinr[iidx];
 184         i_coord_offset   = DIM*inr;
 185
 186         /* Load i particle coords and add shift vector */
 187         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 188                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 189
 190         fix1             = _mm_setzero_ps();
 191         fiy1             = _mm_setzero_ps();
 192         fiz1             = _mm_setzero_ps();
 193         fix2             = _mm_setzero_ps();
 194         fiy2             = _mm_setzero_ps();
 195         fiz2             = _mm_setzero_ps();
 196         fix3             = _mm_setzero_ps();
 197         fiy3             = _mm_setzero_ps();
 198         fiz3             = _mm_setzero_ps();
 199
 200         /* Reset potential sums */
 201         velecsum         = _mm_setzero_ps();
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 205         {
 206
 207             /* Get j neighbor index, and coordinate index */
 208             jnrA             = jjnr[jidx];
 209             jnrB             = jjnr[jidx+1];
 210             jnrC             = jjnr[jidx+2];
 211             jnrD             = jjnr[jidx+3];
 212             j_coord_offsetA  = DIM*jnrA;
 213             j_coord_offsetB  = DIM*jnrB;
 214             j_coord_offsetC  = DIM*jnrC;
 215             j_coord_offsetD  = DIM*jnrD;
 216
 217             /* load j atom coordinates */
 218             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 219                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 220                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 221
 222             /* Calculate displacement vector */
 223             dx11             = _mm_sub_ps(ix1,jx1);
 224             dy11             = _mm_sub_ps(iy1,jy1);
 225             dz11             = _mm_sub_ps(iz1,jz1);
 226             dx12             = _mm_sub_ps(ix1,jx2);
 227             dy12             = _mm_sub_ps(iy1,jy2);
 228             dz12             = _mm_sub_ps(iz1,jz2);
 229             dx13             = _mm_sub_ps(ix1,jx3);
 230             dy13             = _mm_sub_ps(iy1,jy3);
 231             dz13             = _mm_sub_ps(iz1,jz3);
 232             dx21             = _mm_sub_ps(ix2,jx1);
 233             dy21             = _mm_sub_ps(iy2,jy1);
 234             dz21             = _mm_sub_ps(iz2,jz1);
 235             dx22             = _mm_sub_ps(ix2,jx2);
 236             dy22             = _mm_sub_ps(iy2,jy2);
 237             dz22             = _mm_sub_ps(iz2,jz2);
 238             dx23             = _mm_sub_ps(ix2,jx3);
 239             dy23             = _mm_sub_ps(iy2,jy3);
 240             dz23             = _mm_sub_ps(iz2,jz3);
 241             dx31             = _mm_sub_ps(ix3,jx1);
 242             dy31             = _mm_sub_ps(iy3,jy1);
 243             dz31             = _mm_sub_ps(iz3,jz1);
 244             dx32             = _mm_sub_ps(ix3,jx2);
 245             dy32             = _mm_sub_ps(iy3,jy2);
 246             dz32             = _mm_sub_ps(iz3,jz2);
 247             dx33             = _mm_sub_ps(ix3,jx3);
 248             dy33             = _mm_sub_ps(iy3,jy3);
 249             dz33             = _mm_sub_ps(iz3,jz3);
 250
 251             /* Calculate squared distance and things based on it */
 252             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 253             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 254             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 255             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 256             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 257             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 258             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 259             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 260             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 261
 262             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 263             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 264             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 265             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 266             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 267             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 268             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 269             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 270             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 271
 272             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 273             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 274             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 275             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 276             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 277             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 278             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 279             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 280             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 281
 282             fjx1             = _mm_setzero_ps();
 283             fjy1             = _mm_setzero_ps();
 284             fjz1             = _mm_setzero_ps();
 285             fjx2             = _mm_setzero_ps();
 286             fjy2             = _mm_setzero_ps();
 287             fjz2             = _mm_setzero_ps();
 288             fjx3             = _mm_setzero_ps();
 289             fjy3             = _mm_setzero_ps();
 290             fjz3             = _mm_setzero_ps();
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             r11              = _mm_mul_ps(rsq11,rinv11);
 297
 298             /* EWALD ELECTROSTATICS */
 299
 300             /* Analytical PME correction */
 301             zeta2            = _mm_mul_ps(beta2,rsq11);
 302             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 303             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 304             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 305             felec            = _mm_mul_ps(qq11,felec);
 306             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 307             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 308             velec            = _mm_mul_ps(qq11,velec);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312
 313             fscal            = felec;
 314
 315              /* Update vectorial force */
 316             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 317             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 318             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 319
 320             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 321             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 322             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             r12              = _mm_mul_ps(rsq12,rinv12);
 329
 330             /* EWALD ELECTROSTATICS */
 331
 332             /* Analytical PME correction */
 333             zeta2            = _mm_mul_ps(beta2,rsq12);
 334             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 335             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 336             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 337             felec            = _mm_mul_ps(qq12,felec);
 338             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 339             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 340             velec            = _mm_mul_ps(qq12,velec);
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velecsum         = _mm_add_ps(velecsum,velec);
 344
 345             fscal            = felec;
 346
 347              /* Update vectorial force */
 348             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 349             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 350             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 351
 352             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 353             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 354             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             r13              = _mm_mul_ps(rsq13,rinv13);
 361
 362             /* EWALD ELECTROSTATICS */
 363
 364             /* Analytical PME correction */
 365             zeta2            = _mm_mul_ps(beta2,rsq13);
 366             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
 367             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 368             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 369             felec            = _mm_mul_ps(qq13,felec);
 370             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 371             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv13);
 372             velec            = _mm_mul_ps(qq13,velec);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_ps(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379              /* Update vectorial force */
 380             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 381             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 382             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 383
 384             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 385             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 386             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             r21              = _mm_mul_ps(rsq21,rinv21);
 393
 394             /* EWALD ELECTROSTATICS */
 395
 396             /* Analytical PME correction */
 397             zeta2            = _mm_mul_ps(beta2,rsq21);
 398             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 399             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 400             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 401             felec            = _mm_mul_ps(qq21,felec);
 402             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 403             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 404             velec            = _mm_mul_ps(qq21,velec);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _mm_add_ps(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411              /* Update vectorial force */
 412             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 413             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 414             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 415
 416             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 417             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 418             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             r22              = _mm_mul_ps(rsq22,rinv22);
 425
 426             /* EWALD ELECTROSTATICS */
 427
 428             /* Analytical PME correction */
 429             zeta2            = _mm_mul_ps(beta2,rsq22);
 430             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 431             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 432             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 433             felec            = _mm_mul_ps(qq22,felec);
 434             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 435             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 436             velec            = _mm_mul_ps(qq22,velec);
 437
 438             /* Update potential sum for this i atom from the interaction with this j atom. */
 439             velecsum         = _mm_add_ps(velecsum,velec);
 440
 441             fscal            = felec;
 442
 443              /* Update vectorial force */
 444             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 445             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 446             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 447
 448             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 449             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 450             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             r23              = _mm_mul_ps(rsq23,rinv23);
 457
 458             /* EWALD ELECTROSTATICS */
 459
 460             /* Analytical PME correction */
 461             zeta2            = _mm_mul_ps(beta2,rsq23);
 462             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
 463             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 464             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 465             felec            = _mm_mul_ps(qq23,felec);
 466             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 467             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv23);
 468             velec            = _mm_mul_ps(qq23,velec);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velecsum         = _mm_add_ps(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475              /* Update vectorial force */
 476             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 477             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 478             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 479
 480             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 481             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 482             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 483
 484             /**************************
 485              * CALCULATE INTERACTIONS *
 486              **************************/
 487
 488             r31              = _mm_mul_ps(rsq31,rinv31);
 489
 490             /* EWALD ELECTROSTATICS */
 491
 492             /* Analytical PME correction */
 493             zeta2            = _mm_mul_ps(beta2,rsq31);
 494             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
 495             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 496             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 497             felec            = _mm_mul_ps(qq31,felec);
 498             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 499             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv31);
 500             velec            = _mm_mul_ps(qq31,velec);
 501
 502             /* Update potential sum for this i atom from the interaction with this j atom. */
 503             velecsum         = _mm_add_ps(velecsum,velec);
 504
 505             fscal            = felec;
 506
 507              /* Update vectorial force */
 508             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 509             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 510             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 511
 512             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 513             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 514             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             r32              = _mm_mul_ps(rsq32,rinv32);
 521
 522             /* EWALD ELECTROSTATICS */
 523
 524             /* Analytical PME correction */
 525             zeta2            = _mm_mul_ps(beta2,rsq32);
 526             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
 527             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 528             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 529             felec            = _mm_mul_ps(qq32,felec);
 530             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 531             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv32);
 532             velec            = _mm_mul_ps(qq32,velec);
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velecsum         = _mm_add_ps(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539              /* Update vectorial force */
 540             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 541             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 542             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 543
 544             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 545             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 546             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 547
 548             /**************************
 549              * CALCULATE INTERACTIONS *
 550              **************************/
 551
 552             r33              = _mm_mul_ps(rsq33,rinv33);
 553
 554             /* EWALD ELECTROSTATICS */
 555
 556             /* Analytical PME correction */
 557             zeta2            = _mm_mul_ps(beta2,rsq33);
 558             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
 559             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 560             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 561             felec            = _mm_mul_ps(qq33,felec);
 562             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 563             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv33);
 564             velec            = _mm_mul_ps(qq33,velec);
 565
 566             /* Update potential sum for this i atom from the interaction with this j atom. */
 567             velecsum         = _mm_add_ps(velecsum,velec);
 568
 569             fscal            = felec;
 570
 571              /* Update vectorial force */
 572             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 573             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 574             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 575
 576             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 577             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 578             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 579
 580             fjptrA             = f+j_coord_offsetA;
 581             fjptrB             = f+j_coord_offsetB;
 582             fjptrC             = f+j_coord_offsetC;
 583             fjptrD             = f+j_coord_offsetD;
 584
 585             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 586                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 587
 588             /* Inner loop uses 261 flops */
 589         }
 590
 591         if(jidx<j_index_end)
 592         {
 593
 594             /* Get j neighbor index, and coordinate index */
 595             jnrlistA         = jjnr[jidx];
 596             jnrlistB         = jjnr[jidx+1];
 597             jnrlistC         = jjnr[jidx+2];
 598             jnrlistD         = jjnr[jidx+3];
 599             /* Sign of each element will be negative for non-real atoms.
 600              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 601              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 602              */
 603             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 604             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 605             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 606             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 607             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 608             j_coord_offsetA  = DIM*jnrA;
 609             j_coord_offsetB  = DIM*jnrB;
 610             j_coord_offsetC  = DIM*jnrC;
 611             j_coord_offsetD  = DIM*jnrD;
 612
 613             /* load j atom coordinates */
 614             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 615                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 616                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 617
 618             /* Calculate displacement vector */
 619             dx11             = _mm_sub_ps(ix1,jx1);
 620             dy11             = _mm_sub_ps(iy1,jy1);
 621             dz11             = _mm_sub_ps(iz1,jz1);
 622             dx12             = _mm_sub_ps(ix1,jx2);
 623             dy12             = _mm_sub_ps(iy1,jy2);
 624             dz12             = _mm_sub_ps(iz1,jz2);
 625             dx13             = _mm_sub_ps(ix1,jx3);
 626             dy13             = _mm_sub_ps(iy1,jy3);
 627             dz13             = _mm_sub_ps(iz1,jz3);
 628             dx21             = _mm_sub_ps(ix2,jx1);
 629             dy21             = _mm_sub_ps(iy2,jy1);
 630             dz21             = _mm_sub_ps(iz2,jz1);
 631             dx22             = _mm_sub_ps(ix2,jx2);
 632             dy22             = _mm_sub_ps(iy2,jy2);
 633             dz22             = _mm_sub_ps(iz2,jz2);
 634             dx23             = _mm_sub_ps(ix2,jx3);
 635             dy23             = _mm_sub_ps(iy2,jy3);
 636             dz23             = _mm_sub_ps(iz2,jz3);
 637             dx31             = _mm_sub_ps(ix3,jx1);
 638             dy31             = _mm_sub_ps(iy3,jy1);
 639             dz31             = _mm_sub_ps(iz3,jz1);
 640             dx32             = _mm_sub_ps(ix3,jx2);
 641             dy32             = _mm_sub_ps(iy3,jy2);
 642             dz32             = _mm_sub_ps(iz3,jz2);
 643             dx33             = _mm_sub_ps(ix3,jx3);
 644             dy33             = _mm_sub_ps(iy3,jy3);
 645             dz33             = _mm_sub_ps(iz3,jz3);
 646
 647             /* Calculate squared distance and things based on it */
 648             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 649             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 650             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 651             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 652             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 653             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 654             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 655             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 656             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 657
 658             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 659             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 660             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 661             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 662             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 663             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 664             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 665             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 666             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 667
 668             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 669             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 670             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 671             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 672             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 673             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 674             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 675             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 676             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 677
 678             fjx1             = _mm_setzero_ps();
 679             fjy1             = _mm_setzero_ps();
 680             fjz1             = _mm_setzero_ps();
 681             fjx2             = _mm_setzero_ps();
 682             fjy2             = _mm_setzero_ps();
 683             fjz2             = _mm_setzero_ps();
 684             fjx3             = _mm_setzero_ps();
 685             fjy3             = _mm_setzero_ps();
 686             fjz3             = _mm_setzero_ps();
 687
 688             /**************************
 689              * CALCULATE INTERACTIONS *
 690              **************************/
 691
 692             r11              = _mm_mul_ps(rsq11,rinv11);
 693             r11              = _mm_andnot_ps(dummy_mask,r11);
 694
 695             /* EWALD ELECTROSTATICS */
 696
 697             /* Analytical PME correction */
 698             zeta2            = _mm_mul_ps(beta2,rsq11);
 699             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 700             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 701             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 702             felec            = _mm_mul_ps(qq11,felec);
 703             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 704             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 705             velec            = _mm_mul_ps(qq11,velec);
 706
 707             /* Update potential sum for this i atom from the interaction with this j atom. */
 708             velec            = _mm_andnot_ps(dummy_mask,velec);
 709             velecsum         = _mm_add_ps(velecsum,velec);
 710
 711             fscal            = felec;
 712
 713             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 714
 715              /* Update vectorial force */
 716             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 717             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 718             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 719
 720             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 721             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 722             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             r12              = _mm_mul_ps(rsq12,rinv12);
 729             r12              = _mm_andnot_ps(dummy_mask,r12);
 730
 731             /* EWALD ELECTROSTATICS */
 732
 733             /* Analytical PME correction */
 734             zeta2            = _mm_mul_ps(beta2,rsq12);
 735             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 736             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 737             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 738             felec            = _mm_mul_ps(qq12,felec);
 739             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 740             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 741             velec            = _mm_mul_ps(qq12,velec);
 742
 743             /* Update potential sum for this i atom from the interaction with this j atom. */
 744             velec            = _mm_andnot_ps(dummy_mask,velec);
 745             velecsum         = _mm_add_ps(velecsum,velec);
 746
 747             fscal            = felec;
 748
 749             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 750
 751              /* Update vectorial force */
 752             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 753             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 754             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 755
 756             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 757             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 758             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 759
 760             /**************************
 761              * CALCULATE INTERACTIONS *
 762              **************************/
 763
 764             r13              = _mm_mul_ps(rsq13,rinv13);
 765             r13              = _mm_andnot_ps(dummy_mask,r13);
 766
 767             /* EWALD ELECTROSTATICS */
 768
 769             /* Analytical PME correction */
 770             zeta2            = _mm_mul_ps(beta2,rsq13);
 771             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
 772             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 773             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 774             felec            = _mm_mul_ps(qq13,felec);
 775             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 776             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv13);
 777             velec            = _mm_mul_ps(qq13,velec);
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _mm_andnot_ps(dummy_mask,velec);
 781             velecsum         = _mm_add_ps(velecsum,velec);
 782
 783             fscal            = felec;
 784
 785             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 786
 787              /* Update vectorial force */
 788             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 789             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 790             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 791
 792             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 793             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 794             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             r21              = _mm_mul_ps(rsq21,rinv21);
 801             r21              = _mm_andnot_ps(dummy_mask,r21);
 802
 803             /* EWALD ELECTROSTATICS */
 804
 805             /* Analytical PME correction */
 806             zeta2            = _mm_mul_ps(beta2,rsq21);
 807             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 808             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 809             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 810             felec            = _mm_mul_ps(qq21,felec);
 811             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 812             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 813             velec            = _mm_mul_ps(qq21,velec);
 814
 815             /* Update potential sum for this i atom from the interaction with this j atom. */
 816             velec            = _mm_andnot_ps(dummy_mask,velec);
 817             velecsum         = _mm_add_ps(velecsum,velec);
 818
 819             fscal            = felec;
 820
 821             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 822
 823              /* Update vectorial force */
 824             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 825             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 826             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 827
 828             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 829             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 830             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 831
 832             /**************************
 833              * CALCULATE INTERACTIONS *
 834              **************************/
 835
 836             r22              = _mm_mul_ps(rsq22,rinv22);
 837             r22              = _mm_andnot_ps(dummy_mask,r22);
 838
 839             /* EWALD ELECTROSTATICS */
 840
 841             /* Analytical PME correction */
 842             zeta2            = _mm_mul_ps(beta2,rsq22);
 843             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 844             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 845             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 846             felec            = _mm_mul_ps(qq22,felec);
 847             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 848             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 849             velec            = _mm_mul_ps(qq22,velec);
 850
 851             /* Update potential sum for this i atom from the interaction with this j atom. */
 852             velec            = _mm_andnot_ps(dummy_mask,velec);
 853             velecsum         = _mm_add_ps(velecsum,velec);
 854
 855             fscal            = felec;
 856
 857             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 858
 859              /* Update vectorial force */
 860             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 861             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 862             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 863
 864             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 865             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 866             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 867
 868             /**************************
 869              * CALCULATE INTERACTIONS *
 870              **************************/
 871
 872             r23              = _mm_mul_ps(rsq23,rinv23);
 873             r23              = _mm_andnot_ps(dummy_mask,r23);
 874
 875             /* EWALD ELECTROSTATICS */
 876
 877             /* Analytical PME correction */
 878             zeta2            = _mm_mul_ps(beta2,rsq23);
 879             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
 880             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 881             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 882             felec            = _mm_mul_ps(qq23,felec);
 883             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 884             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv23);
 885             velec            = _mm_mul_ps(qq23,velec);
 886
 887             /* Update potential sum for this i atom from the interaction with this j atom. */
 888             velec            = _mm_andnot_ps(dummy_mask,velec);
 889             velecsum         = _mm_add_ps(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 894
 895              /* Update vectorial force */
 896             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 897             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 898             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 899
 900             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 901             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 902             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 903
 904             /**************************
 905              * CALCULATE INTERACTIONS *
 906              **************************/
 907
 908             r31              = _mm_mul_ps(rsq31,rinv31);
 909             r31              = _mm_andnot_ps(dummy_mask,r31);
 910
 911             /* EWALD ELECTROSTATICS */
 912
 913             /* Analytical PME correction */
 914             zeta2            = _mm_mul_ps(beta2,rsq31);
 915             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
 916             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 917             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 918             felec            = _mm_mul_ps(qq31,felec);
 919             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 920             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv31);
 921             velec            = _mm_mul_ps(qq31,velec);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_andnot_ps(dummy_mask,velec);
 925             velecsum         = _mm_add_ps(velecsum,velec);
 926
 927             fscal            = felec;
 928
 929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 930
 931              /* Update vectorial force */
 932             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 933             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 934             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 935
 936             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 937             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 938             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 939
 940             /**************************
 941              * CALCULATE INTERACTIONS *
 942              **************************/
 943
 944             r32              = _mm_mul_ps(rsq32,rinv32);
 945             r32              = _mm_andnot_ps(dummy_mask,r32);
 946
 947             /* EWALD ELECTROSTATICS */
 948
 949             /* Analytical PME correction */
 950             zeta2            = _mm_mul_ps(beta2,rsq32);
 951             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
 952             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 953             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 954             felec            = _mm_mul_ps(qq32,felec);
 955             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 956             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv32);
 957             velec            = _mm_mul_ps(qq32,velec);
 958
 959             /* Update potential sum for this i atom from the interaction with this j atom. */
 960             velec            = _mm_andnot_ps(dummy_mask,velec);
 961             velecsum         = _mm_add_ps(velecsum,velec);
 962
 963             fscal            = felec;
 964
 965             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 966
 967              /* Update vectorial force */
 968             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 969             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 970             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 971
 972             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 973             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 974             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 975
 976             /**************************
 977              * CALCULATE INTERACTIONS *
 978              **************************/
 979
 980             r33              = _mm_mul_ps(rsq33,rinv33);
 981             r33              = _mm_andnot_ps(dummy_mask,r33);
 982
 983             /* EWALD ELECTROSTATICS */
 984
 985             /* Analytical PME correction */
 986             zeta2            = _mm_mul_ps(beta2,rsq33);
 987             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
 988             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 989             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 990             felec            = _mm_mul_ps(qq33,felec);
 991             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 992             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv33);
 993             velec            = _mm_mul_ps(qq33,velec);
 994
 995             /* Update potential sum for this i atom from the interaction with this j atom. */
 996             velec            = _mm_andnot_ps(dummy_mask,velec);
 997             velecsum         = _mm_add_ps(velecsum,velec);
 998
 999             fscal            = felec;
1000
1001             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1002
1003              /* Update vectorial force */
1004             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1005             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1006             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1007
1008             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1009             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1010             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1011
1012             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1013             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1014             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1015             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1016
1017             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1018                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1019
1020             /* Inner loop uses 270 flops */
1021         }
1022
1023         /* End of innermost loop */
1024
1025         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1026                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1027
1028         ggid                        = gid[iidx];
1029         /* Update potential energies */
1030         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1031
1032         /* Increment number of inner iterations */
1033         inneriter                  += j_index_end - j_index_start;
1034
1035         /* Outer loop uses 19 flops */
1036     }
1037
1038     /* Increment number of outer iterations */
1039     outeriter        += nri;
1040
1041     /* Update outer/inner flops */
1042
1043     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*270);
1044 }
1045 /*
1046  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_F_avx_128_fma_single
1047  * Electrostatics interaction: Ewald
1048  * VdW interaction:            None
1049  * Geometry:                   Water4-Water4
1050  * Calculate force/pot:        Force
1051  */
1052 void
1053 nb_kernel_ElecEw_VdwNone_GeomW4W4_F_avx_128_fma_single
1054                     (t_nblist                    * gmx_restrict       nlist,
1055                      rvec                        * gmx_restrict          xx,
1056                      rvec                        * gmx_restrict          ff,
1057                      t_forcerec                  * gmx_restrict          fr,
1058                      t_mdatoms                   * gmx_restrict     mdatoms,
1059                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1060                      t_nrnb                      * gmx_restrict        nrnb)
1061 {
1062     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1063      * just 0 for non-waters.
1064      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1065      * jnr indices corresponding to data put in the four positions in the SIMD register.
1066      */
1067     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1068     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1069     int              jnrA,jnrB,jnrC,jnrD;
1070     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1071     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1072     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1073     real             rcutoff_scalar;
1074     real             *shiftvec,*fshift,*x,*f;
1075     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1076     real             scratch[4*DIM];
1077     __m128           fscal,rcutoff,rcutoff2,jidxall;
1078     int              vdwioffset1;
1079     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1080     int              vdwioffset2;
1081     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1082     int              vdwioffset3;
1083     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1084     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1085     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1086     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1087     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1088     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1089     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1090     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1091     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1092     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1093     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1094     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1095     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1096     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1097     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1098     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1099     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1100     real             *charge;
1101     __m128i          ewitab;
1102     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1103     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1104     real             *ewtab;
1105     __m128           dummy_mask,cutoff_mask;
1106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1107     __m128           one     = _mm_set1_ps(1.0);
1108     __m128           two     = _mm_set1_ps(2.0);
1109     x                = xx[0];
1110     f                = ff[0];
1111
1112     nri              = nlist->nri;
1113     iinr             = nlist->iinr;
1114     jindex           = nlist->jindex;
1115     jjnr             = nlist->jjnr;
1116     shiftidx         = nlist->shift;
1117     gid              = nlist->gid;
1118     shiftvec         = fr->shift_vec[0];
1119     fshift           = fr->fshift[0];
1120     facel            = _mm_set1_ps(fr->epsfac);
1121     charge           = mdatoms->chargeA;
1122
1123     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1124     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
1125     beta2            = _mm_mul_ps(beta,beta);
1126     beta3            = _mm_mul_ps(beta,beta2);
1127     ewtab            = fr->ic->tabq_coul_F;
1128     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1129     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1130
1131     /* Setup water-specific parameters */
1132     inr              = nlist->iinr[0];
1133     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1134     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1135     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1136
1137     jq1              = _mm_set1_ps(charge[inr+1]);
1138     jq2              = _mm_set1_ps(charge[inr+2]);
1139     jq3              = _mm_set1_ps(charge[inr+3]);
1140     qq11             = _mm_mul_ps(iq1,jq1);
1141     qq12             = _mm_mul_ps(iq1,jq2);
1142     qq13             = _mm_mul_ps(iq1,jq3);
1143     qq21             = _mm_mul_ps(iq2,jq1);
1144     qq22             = _mm_mul_ps(iq2,jq2);
1145     qq23             = _mm_mul_ps(iq2,jq3);
1146     qq31             = _mm_mul_ps(iq3,jq1);
1147     qq32             = _mm_mul_ps(iq3,jq2);
1148     qq33             = _mm_mul_ps(iq3,jq3);
1149
1150     /* Avoid stupid compiler warnings */
1151     jnrA = jnrB = jnrC = jnrD = 0;
1152     j_coord_offsetA = 0;
1153     j_coord_offsetB = 0;
1154     j_coord_offsetC = 0;
1155     j_coord_offsetD = 0;
1156
1157     outeriter        = 0;
1158     inneriter        = 0;
1159
1160     for(iidx=0;iidx<4*DIM;iidx++)
1161     {
1162         scratch[iidx] = 0.0;
1163     }
1164
1165     /* Start outer loop over neighborlists */
1166     for(iidx=0; iidx<nri; iidx++)
1167     {
1168         /* Load shift vector for this list */
1169         i_shift_offset   = DIM*shiftidx[iidx];
1170
1171         /* Load limits for loop over neighbors */
1172         j_index_start    = jindex[iidx];
1173         j_index_end      = jindex[iidx+1];
1174
1175         /* Get outer coordinate index */
1176         inr              = iinr[iidx];
1177         i_coord_offset   = DIM*inr;
1178
1179         /* Load i particle coords and add shift vector */
1180         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1181                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1182
1183         fix1             = _mm_setzero_ps();
1184         fiy1             = _mm_setzero_ps();
1185         fiz1             = _mm_setzero_ps();
1186         fix2             = _mm_setzero_ps();
1187         fiy2             = _mm_setzero_ps();
1188         fiz2             = _mm_setzero_ps();
1189         fix3             = _mm_setzero_ps();
1190         fiy3             = _mm_setzero_ps();
1191         fiz3             = _mm_setzero_ps();
1192
1193         /* Start inner kernel loop */
1194         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1195         {
1196
1197             /* Get j neighbor index, and coordinate index */
1198             jnrA             = jjnr[jidx];
1199             jnrB             = jjnr[jidx+1];
1200             jnrC             = jjnr[jidx+2];
1201             jnrD             = jjnr[jidx+3];
1202             j_coord_offsetA  = DIM*jnrA;
1203             j_coord_offsetB  = DIM*jnrB;
1204             j_coord_offsetC  = DIM*jnrC;
1205             j_coord_offsetD  = DIM*jnrD;
1206
1207             /* load j atom coordinates */
1208             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1209                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1210                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1211
1212             /* Calculate displacement vector */
1213             dx11             = _mm_sub_ps(ix1,jx1);
1214             dy11             = _mm_sub_ps(iy1,jy1);
1215             dz11             = _mm_sub_ps(iz1,jz1);
1216             dx12             = _mm_sub_ps(ix1,jx2);
1217             dy12             = _mm_sub_ps(iy1,jy2);
1218             dz12             = _mm_sub_ps(iz1,jz2);
1219             dx13             = _mm_sub_ps(ix1,jx3);
1220             dy13             = _mm_sub_ps(iy1,jy3);
1221             dz13             = _mm_sub_ps(iz1,jz3);
1222             dx21             = _mm_sub_ps(ix2,jx1);
1223             dy21             = _mm_sub_ps(iy2,jy1);
1224             dz21             = _mm_sub_ps(iz2,jz1);
1225             dx22             = _mm_sub_ps(ix2,jx2);
1226             dy22             = _mm_sub_ps(iy2,jy2);
1227             dz22             = _mm_sub_ps(iz2,jz2);
1228             dx23             = _mm_sub_ps(ix2,jx3);
1229             dy23             = _mm_sub_ps(iy2,jy3);
1230             dz23             = _mm_sub_ps(iz2,jz3);
1231             dx31             = _mm_sub_ps(ix3,jx1);
1232             dy31             = _mm_sub_ps(iy3,jy1);
1233             dz31             = _mm_sub_ps(iz3,jz1);
1234             dx32             = _mm_sub_ps(ix3,jx2);
1235             dy32             = _mm_sub_ps(iy3,jy2);
1236             dz32             = _mm_sub_ps(iz3,jz2);
1237             dx33             = _mm_sub_ps(ix3,jx3);
1238             dy33             = _mm_sub_ps(iy3,jy3);
1239             dz33             = _mm_sub_ps(iz3,jz3);
1240
1241             /* Calculate squared distance and things based on it */
1242             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1243             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1244             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1245             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1246             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1247             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1248             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1249             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1250             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1251
1252             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1253             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1254             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1255             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1256             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1257             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1258             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1259             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1260             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1261
1262             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1263             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1264             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1265             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1266             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1267             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1268             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1269             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1270             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1271
1272             fjx1             = _mm_setzero_ps();
1273             fjy1             = _mm_setzero_ps();
1274             fjz1             = _mm_setzero_ps();
1275             fjx2             = _mm_setzero_ps();
1276             fjy2             = _mm_setzero_ps();
1277             fjz2             = _mm_setzero_ps();
1278             fjx3             = _mm_setzero_ps();
1279             fjy3             = _mm_setzero_ps();
1280             fjz3             = _mm_setzero_ps();
1281
1282             /**************************
1283              * CALCULATE INTERACTIONS *
1284              **************************/
1285
1286             r11              = _mm_mul_ps(rsq11,rinv11);
1287
1288             /* EWALD ELECTROSTATICS */
1289
1290             /* Analytical PME correction */
1291             zeta2            = _mm_mul_ps(beta2,rsq11);
1292             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1293             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1294             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1295             felec            = _mm_mul_ps(qq11,felec);
1296
1297             fscal            = felec;
1298
1299              /* Update vectorial force */
1300             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1301             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1302             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1303
1304             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1305             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1306             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1307
1308             /**************************
1309              * CALCULATE INTERACTIONS *
1310              **************************/
1311
1312             r12              = _mm_mul_ps(rsq12,rinv12);
1313
1314             /* EWALD ELECTROSTATICS */
1315
1316             /* Analytical PME correction */
1317             zeta2            = _mm_mul_ps(beta2,rsq12);
1318             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1319             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1320             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1321             felec            = _mm_mul_ps(qq12,felec);
1322
1323             fscal            = felec;
1324
1325              /* Update vectorial force */
1326             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1327             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1328             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1329
1330             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1331             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1332             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1333
1334             /**************************
1335              * CALCULATE INTERACTIONS *
1336              **************************/
1337
1338             r13              = _mm_mul_ps(rsq13,rinv13);
1339
1340             /* EWALD ELECTROSTATICS */
1341
1342             /* Analytical PME correction */
1343             zeta2            = _mm_mul_ps(beta2,rsq13);
1344             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
1345             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1346             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1347             felec            = _mm_mul_ps(qq13,felec);
1348
1349             fscal            = felec;
1350
1351              /* Update vectorial force */
1352             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1353             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1354             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1355
1356             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1357             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1358             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1359
1360             /**************************
1361              * CALCULATE INTERACTIONS *
1362              **************************/
1363
1364             r21              = _mm_mul_ps(rsq21,rinv21);
1365
1366             /* EWALD ELECTROSTATICS */
1367
1368             /* Analytical PME correction */
1369             zeta2            = _mm_mul_ps(beta2,rsq21);
1370             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1371             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1372             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1373             felec            = _mm_mul_ps(qq21,felec);
1374
1375             fscal            = felec;
1376
1377              /* Update vectorial force */
1378             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1379             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1380             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1381
1382             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1383             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1384             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             r22              = _mm_mul_ps(rsq22,rinv22);
1391
1392             /* EWALD ELECTROSTATICS */
1393
1394             /* Analytical PME correction */
1395             zeta2            = _mm_mul_ps(beta2,rsq22);
1396             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1397             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1398             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1399             felec            = _mm_mul_ps(qq22,felec);
1400
1401             fscal            = felec;
1402
1403              /* Update vectorial force */
1404             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1405             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1406             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1407
1408             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1409             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1410             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             r23              = _mm_mul_ps(rsq23,rinv23);
1417
1418             /* EWALD ELECTROSTATICS */
1419
1420             /* Analytical PME correction */
1421             zeta2            = _mm_mul_ps(beta2,rsq23);
1422             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
1423             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1424             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1425             felec            = _mm_mul_ps(qq23,felec);
1426
1427             fscal            = felec;
1428
1429              /* Update vectorial force */
1430             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1431             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1432             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1433
1434             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1435             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1436             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             r31              = _mm_mul_ps(rsq31,rinv31);
1443
1444             /* EWALD ELECTROSTATICS */
1445
1446             /* Analytical PME correction */
1447             zeta2            = _mm_mul_ps(beta2,rsq31);
1448             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
1449             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1450             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1451             felec            = _mm_mul_ps(qq31,felec);
1452
1453             fscal            = felec;
1454
1455              /* Update vectorial force */
1456             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1457             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1458             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1459
1460             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1461             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1462             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             r32              = _mm_mul_ps(rsq32,rinv32);
1469
1470             /* EWALD ELECTROSTATICS */
1471
1472             /* Analytical PME correction */
1473             zeta2            = _mm_mul_ps(beta2,rsq32);
1474             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
1475             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1476             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1477             felec            = _mm_mul_ps(qq32,felec);
1478
1479             fscal            = felec;
1480
1481              /* Update vectorial force */
1482             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1483             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1484             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1485
1486             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1487             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1488             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1489
1490             /**************************
1491              * CALCULATE INTERACTIONS *
1492              **************************/
1493
1494             r33              = _mm_mul_ps(rsq33,rinv33);
1495
1496             /* EWALD ELECTROSTATICS */
1497
1498             /* Analytical PME correction */
1499             zeta2            = _mm_mul_ps(beta2,rsq33);
1500             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
1501             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1502             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1503             felec            = _mm_mul_ps(qq33,felec);
1504
1505             fscal            = felec;
1506
1507              /* Update vectorial force */
1508             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1509             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1510             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1511
1512             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1513             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1514             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1515
1516             fjptrA             = f+j_coord_offsetA;
1517             fjptrB             = f+j_coord_offsetB;
1518             fjptrC             = f+j_coord_offsetC;
1519             fjptrD             = f+j_coord_offsetD;
1520
1521             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1522                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1523
1524             /* Inner loop uses 252 flops */
1525         }
1526
1527         if(jidx<j_index_end)
1528         {
1529
1530             /* Get j neighbor index, and coordinate index */
1531             jnrlistA         = jjnr[jidx];
1532             jnrlistB         = jjnr[jidx+1];
1533             jnrlistC         = jjnr[jidx+2];
1534             jnrlistD         = jjnr[jidx+3];
1535             /* Sign of each element will be negative for non-real atoms.
1536              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1537              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1538              */
1539             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1540             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1541             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1542             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1543             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1544             j_coord_offsetA  = DIM*jnrA;
1545             j_coord_offsetB  = DIM*jnrB;
1546             j_coord_offsetC  = DIM*jnrC;
1547             j_coord_offsetD  = DIM*jnrD;
1548
1549             /* load j atom coordinates */
1550             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1551                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1552                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1553
1554             /* Calculate displacement vector */
1555             dx11             = _mm_sub_ps(ix1,jx1);
1556             dy11             = _mm_sub_ps(iy1,jy1);
1557             dz11             = _mm_sub_ps(iz1,jz1);
1558             dx12             = _mm_sub_ps(ix1,jx2);
1559             dy12             = _mm_sub_ps(iy1,jy2);
1560             dz12             = _mm_sub_ps(iz1,jz2);
1561             dx13             = _mm_sub_ps(ix1,jx3);
1562             dy13             = _mm_sub_ps(iy1,jy3);
1563             dz13             = _mm_sub_ps(iz1,jz3);
1564             dx21             = _mm_sub_ps(ix2,jx1);
1565             dy21             = _mm_sub_ps(iy2,jy1);
1566             dz21             = _mm_sub_ps(iz2,jz1);
1567             dx22             = _mm_sub_ps(ix2,jx2);
1568             dy22             = _mm_sub_ps(iy2,jy2);
1569             dz22             = _mm_sub_ps(iz2,jz2);
1570             dx23             = _mm_sub_ps(ix2,jx3);
1571             dy23             = _mm_sub_ps(iy2,jy3);
1572             dz23             = _mm_sub_ps(iz2,jz3);
1573             dx31             = _mm_sub_ps(ix3,jx1);
1574             dy31             = _mm_sub_ps(iy3,jy1);
1575             dz31             = _mm_sub_ps(iz3,jz1);
1576             dx32             = _mm_sub_ps(ix3,jx2);
1577             dy32             = _mm_sub_ps(iy3,jy2);
1578             dz32             = _mm_sub_ps(iz3,jz2);
1579             dx33             = _mm_sub_ps(ix3,jx3);
1580             dy33             = _mm_sub_ps(iy3,jy3);
1581             dz33             = _mm_sub_ps(iz3,jz3);
1582
1583             /* Calculate squared distance and things based on it */
1584             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1585             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1586             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1587             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1588             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1589             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1590             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1591             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1592             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1593
1594             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1595             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1596             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1597             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1598             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1599             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1600             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1601             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1602             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1603
1604             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1605             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1606             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1607             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1608             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1609             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1610             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1611             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1612             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1613
1614             fjx1             = _mm_setzero_ps();
1615             fjy1             = _mm_setzero_ps();
1616             fjz1             = _mm_setzero_ps();
1617             fjx2             = _mm_setzero_ps();
1618             fjy2             = _mm_setzero_ps();
1619             fjz2             = _mm_setzero_ps();
1620             fjx3             = _mm_setzero_ps();
1621             fjy3             = _mm_setzero_ps();
1622             fjz3             = _mm_setzero_ps();
1623
1624             /**************************
1625              * CALCULATE INTERACTIONS *
1626              **************************/
1627
1628             r11              = _mm_mul_ps(rsq11,rinv11);
1629             r11              = _mm_andnot_ps(dummy_mask,r11);
1630
1631             /* EWALD ELECTROSTATICS */
1632
1633             /* Analytical PME correction */
1634             zeta2            = _mm_mul_ps(beta2,rsq11);
1635             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1636             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1637             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1638             felec            = _mm_mul_ps(qq11,felec);
1639
1640             fscal            = felec;
1641
1642             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1643
1644              /* Update vectorial force */
1645             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1646             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1647             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1648
1649             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1650             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1651             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             r12              = _mm_mul_ps(rsq12,rinv12);
1658             r12              = _mm_andnot_ps(dummy_mask,r12);
1659
1660             /* EWALD ELECTROSTATICS */
1661
1662             /* Analytical PME correction */
1663             zeta2            = _mm_mul_ps(beta2,rsq12);
1664             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1665             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1666             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1667             felec            = _mm_mul_ps(qq12,felec);
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1672
1673              /* Update vectorial force */
1674             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1675             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1676             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1677
1678             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1679             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1680             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1681
1682             /**************************
1683              * CALCULATE INTERACTIONS *
1684              **************************/
1685
1686             r13              = _mm_mul_ps(rsq13,rinv13);
1687             r13              = _mm_andnot_ps(dummy_mask,r13);
1688
1689             /* EWALD ELECTROSTATICS */
1690
1691             /* Analytical PME correction */
1692             zeta2            = _mm_mul_ps(beta2,rsq13);
1693             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
1694             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1695             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1696             felec            = _mm_mul_ps(qq13,felec);
1697
1698             fscal            = felec;
1699
1700             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1701
1702              /* Update vectorial force */
1703             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1704             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1705             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1706
1707             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1708             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1709             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1710
1711             /**************************
1712              * CALCULATE INTERACTIONS *
1713              **************************/
1714
1715             r21              = _mm_mul_ps(rsq21,rinv21);
1716             r21              = _mm_andnot_ps(dummy_mask,r21);
1717
1718             /* EWALD ELECTROSTATICS */
1719
1720             /* Analytical PME correction */
1721             zeta2            = _mm_mul_ps(beta2,rsq21);
1722             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1723             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1724             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1725             felec            = _mm_mul_ps(qq21,felec);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1730
1731              /* Update vectorial force */
1732             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1733             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1734             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1735
1736             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1737             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1738             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1739
1740             /**************************
1741              * CALCULATE INTERACTIONS *
1742              **************************/
1743
1744             r22              = _mm_mul_ps(rsq22,rinv22);
1745             r22              = _mm_andnot_ps(dummy_mask,r22);
1746
1747             /* EWALD ELECTROSTATICS */
1748
1749             /* Analytical PME correction */
1750             zeta2            = _mm_mul_ps(beta2,rsq22);
1751             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1752             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1753             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1754             felec            = _mm_mul_ps(qq22,felec);
1755
1756             fscal            = felec;
1757
1758             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1759
1760              /* Update vectorial force */
1761             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1762             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1763             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1764
1765             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1766             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1767             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             r23              = _mm_mul_ps(rsq23,rinv23);
1774             r23              = _mm_andnot_ps(dummy_mask,r23);
1775
1776             /* EWALD ELECTROSTATICS */
1777
1778             /* Analytical PME correction */
1779             zeta2            = _mm_mul_ps(beta2,rsq23);
1780             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
1781             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1782             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1783             felec            = _mm_mul_ps(qq23,felec);
1784
1785             fscal            = felec;
1786
1787             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1788
1789              /* Update vectorial force */
1790             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1791             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1792             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1793
1794             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1795             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1796             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1797
1798             /**************************
1799              * CALCULATE INTERACTIONS *
1800              **************************/
1801
1802             r31              = _mm_mul_ps(rsq31,rinv31);
1803             r31              = _mm_andnot_ps(dummy_mask,r31);
1804
1805             /* EWALD ELECTROSTATICS */
1806
1807             /* Analytical PME correction */
1808             zeta2            = _mm_mul_ps(beta2,rsq31);
1809             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
1810             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1811             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1812             felec            = _mm_mul_ps(qq31,felec);
1813
1814             fscal            = felec;
1815
1816             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1817
1818              /* Update vectorial force */
1819             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1820             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1821             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1822
1823             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1824             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1825             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1826
1827             /**************************
1828              * CALCULATE INTERACTIONS *
1829              **************************/
1830
1831             r32              = _mm_mul_ps(rsq32,rinv32);
1832             r32              = _mm_andnot_ps(dummy_mask,r32);
1833
1834             /* EWALD ELECTROSTATICS */
1835
1836             /* Analytical PME correction */
1837             zeta2            = _mm_mul_ps(beta2,rsq32);
1838             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
1839             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1840             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1841             felec            = _mm_mul_ps(qq32,felec);
1842
1843             fscal            = felec;
1844
1845             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1846
1847              /* Update vectorial force */
1848             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1849             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1850             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1851
1852             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1853             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1854             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1855
1856             /**************************
1857              * CALCULATE INTERACTIONS *
1858              **************************/
1859
1860             r33              = _mm_mul_ps(rsq33,rinv33);
1861             r33              = _mm_andnot_ps(dummy_mask,r33);
1862
1863             /* EWALD ELECTROSTATICS */
1864
1865             /* Analytical PME correction */
1866             zeta2            = _mm_mul_ps(beta2,rsq33);
1867             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
1868             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1869             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1870             felec            = _mm_mul_ps(qq33,felec);
1871
1872             fscal            = felec;
1873
1874             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1875
1876              /* Update vectorial force */
1877             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1878             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1879             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1880
1881             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1882             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1883             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1884
1885             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1886             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1887             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1888             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1889
1890             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1891                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1892
1893             /* Inner loop uses 261 flops */
1894         }
1895
1896         /* End of innermost loop */
1897
1898         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1899                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1900
1901         /* Increment number of inner iterations */
1902         inneriter                  += j_index_end - j_index_start;
1903
1904         /* Outer loop uses 18 flops */
1905     }
1906
1907     /* Increment number of outer iterations */
1908     outeriter        += nri;
1909
1910     /* Update outer/inner flops */
1911
1912     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*261);
1913 }