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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEw_VdwNone_GeomW4W4_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset1;
  84     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwioffset3;
  88     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  89     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  90     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  92     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  94     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  95     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  98     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 101     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 102     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 103     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     __m128i          ewitab;
 107     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 108     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 109     real             *ewtab;
 110     __m128           dummy_mask,cutoff_mask;
 111     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 112     __m128           one     = _mm_set1_ps(1.0);
 113     __m128           two     = _mm_set1_ps(2.0);
 114     x                = xx[0];
 115     f                = ff[0];
 116
 117     nri              = nlist->nri;
 118     iinr             = nlist->iinr;
 119     jindex           = nlist->jindex;
 120     jjnr             = nlist->jjnr;
 121     shiftidx         = nlist->shift;
 122     gid              = nlist->gid;
 123     shiftvec         = fr->shift_vec[0];
 124     fshift           = fr->fshift[0];
 125     facel            = _mm_set1_ps(fr->epsfac);
 126     charge           = mdatoms->chargeA;
 127
 128     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 129     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 130     beta2            = _mm_mul_ps(beta,beta);
 131     beta3            = _mm_mul_ps(beta,beta2);
 132     ewtab            = fr->ic->tabq_coul_FDV0;
 133     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 134     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 135
 136     /* Setup water-specific parameters */
 137     inr              = nlist->iinr[0];
 138     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 139     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 140     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 141
 142     jq1              = _mm_set1_ps(charge[inr+1]);
 143     jq2              = _mm_set1_ps(charge[inr+2]);
 144     jq3              = _mm_set1_ps(charge[inr+3]);
 145     qq11             = _mm_mul_ps(iq1,jq1);
 146     qq12             = _mm_mul_ps(iq1,jq2);
 147     qq13             = _mm_mul_ps(iq1,jq3);
 148     qq21             = _mm_mul_ps(iq2,jq1);
 149     qq22             = _mm_mul_ps(iq2,jq2);
 150     qq23             = _mm_mul_ps(iq2,jq3);
 151     qq31             = _mm_mul_ps(iq3,jq1);
 152     qq32             = _mm_mul_ps(iq3,jq2);
 153     qq33             = _mm_mul_ps(iq3,jq3);
 154
 155     /* Avoid stupid compiler warnings */
 156     jnrA = jnrB = jnrC = jnrD = 0;
 157     j_coord_offsetA = 0;
 158     j_coord_offsetB = 0;
 159     j_coord_offsetC = 0;
 160     j_coord_offsetD = 0;
 161
 162     outeriter        = 0;
 163     inneriter        = 0;
 164
 165     for(iidx=0;iidx<4*DIM;iidx++)
 166     {
 167         scratch[iidx] = 0.0;
 168     }
 169
 170     /* Start outer loop over neighborlists */
 171     for(iidx=0; iidx<nri; iidx++)
 172     {
 173         /* Load shift vector for this list */
 174         i_shift_offset   = DIM*shiftidx[iidx];
 175
 176         /* Load limits for loop over neighbors */
 177         j_index_start    = jindex[iidx];
 178         j_index_end      = jindex[iidx+1];
 179
 180         /* Get outer coordinate index */
 181         inr              = iinr[iidx];
 182         i_coord_offset   = DIM*inr;
 183
 184         /* Load i particle coords and add shift vector */
 185         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 186                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 187
 188         fix1             = _mm_setzero_ps();
 189         fiy1             = _mm_setzero_ps();
 190         fiz1             = _mm_setzero_ps();
 191         fix2             = _mm_setzero_ps();
 192         fiy2             = _mm_setzero_ps();
 193         fiz2             = _mm_setzero_ps();
 194         fix3             = _mm_setzero_ps();
 195         fiy3             = _mm_setzero_ps();
 196         fiz3             = _mm_setzero_ps();
 197
 198         /* Reset potential sums */
 199         velecsum         = _mm_setzero_ps();
 200
 201         /* Start inner kernel loop */
 202         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 203         {
 204
 205             /* Get j neighbor index, and coordinate index */
 206             jnrA             = jjnr[jidx];
 207             jnrB             = jjnr[jidx+1];
 208             jnrC             = jjnr[jidx+2];
 209             jnrD             = jjnr[jidx+3];
 210             j_coord_offsetA  = DIM*jnrA;
 211             j_coord_offsetB  = DIM*jnrB;
 212             j_coord_offsetC  = DIM*jnrC;
 213             j_coord_offsetD  = DIM*jnrD;
 214
 215             /* load j atom coordinates */
 216             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 217                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 218                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 219
 220             /* Calculate displacement vector */
 221             dx11             = _mm_sub_ps(ix1,jx1);
 222             dy11             = _mm_sub_ps(iy1,jy1);
 223             dz11             = _mm_sub_ps(iz1,jz1);
 224             dx12             = _mm_sub_ps(ix1,jx2);
 225             dy12             = _mm_sub_ps(iy1,jy2);
 226             dz12             = _mm_sub_ps(iz1,jz2);
 227             dx13             = _mm_sub_ps(ix1,jx3);
 228             dy13             = _mm_sub_ps(iy1,jy3);
 229             dz13             = _mm_sub_ps(iz1,jz3);
 230             dx21             = _mm_sub_ps(ix2,jx1);
 231             dy21             = _mm_sub_ps(iy2,jy1);
 232             dz21             = _mm_sub_ps(iz2,jz1);
 233             dx22             = _mm_sub_ps(ix2,jx2);
 234             dy22             = _mm_sub_ps(iy2,jy2);
 235             dz22             = _mm_sub_ps(iz2,jz2);
 236             dx23             = _mm_sub_ps(ix2,jx3);
 237             dy23             = _mm_sub_ps(iy2,jy3);
 238             dz23             = _mm_sub_ps(iz2,jz3);
 239             dx31             = _mm_sub_ps(ix3,jx1);
 240             dy31             = _mm_sub_ps(iy3,jy1);
 241             dz31             = _mm_sub_ps(iz3,jz1);
 242             dx32             = _mm_sub_ps(ix3,jx2);
 243             dy32             = _mm_sub_ps(iy3,jy2);
 244             dz32             = _mm_sub_ps(iz3,jz2);
 245             dx33             = _mm_sub_ps(ix3,jx3);
 246             dy33             = _mm_sub_ps(iy3,jy3);
 247             dz33             = _mm_sub_ps(iz3,jz3);
 248
 249             /* Calculate squared distance and things based on it */
 250             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 251             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 252             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 253             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 254             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 255             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 256             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 257             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 258             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 259
 260             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 261             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 262             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 263             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 264             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 265             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 266             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 267             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 268             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 269
 270             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 271             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 272             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 273             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 274             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 275             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 276             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 277             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 278             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 279
 280             fjx1             = _mm_setzero_ps();
 281             fjy1             = _mm_setzero_ps();
 282             fjz1             = _mm_setzero_ps();
 283             fjx2             = _mm_setzero_ps();
 284             fjy2             = _mm_setzero_ps();
 285             fjz2             = _mm_setzero_ps();
 286             fjx3             = _mm_setzero_ps();
 287             fjy3             = _mm_setzero_ps();
 288             fjz3             = _mm_setzero_ps();
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             r11              = _mm_mul_ps(rsq11,rinv11);
 295
 296             /* EWALD ELECTROSTATICS */
 297
 298             /* Analytical PME correction */
 299             zeta2            = _mm_mul_ps(beta2,rsq11);
 300             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 301             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 302             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 303             felec            = _mm_mul_ps(qq11,felec);
 304             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 305             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 306             velec            = _mm_mul_ps(qq11,velec);
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             velecsum         = _mm_add_ps(velecsum,velec);
 310
 311             fscal            = felec;
 312
 313              /* Update vectorial force */
 314             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 315             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 316             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 317
 318             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 319             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 320             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             r12              = _mm_mul_ps(rsq12,rinv12);
 327
 328             /* EWALD ELECTROSTATICS */
 329
 330             /* Analytical PME correction */
 331             zeta2            = _mm_mul_ps(beta2,rsq12);
 332             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 333             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 334             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 335             felec            = _mm_mul_ps(qq12,felec);
 336             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 337             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 338             velec            = _mm_mul_ps(qq12,velec);
 339
 340             /* Update potential sum for this i atom from the interaction with this j atom. */
 341             velecsum         = _mm_add_ps(velecsum,velec);
 342
 343             fscal            = felec;
 344
 345              /* Update vectorial force */
 346             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 347             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 348             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 349
 350             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 351             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 352             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             r13              = _mm_mul_ps(rsq13,rinv13);
 359
 360             /* EWALD ELECTROSTATICS */
 361
 362             /* Analytical PME correction */
 363             zeta2            = _mm_mul_ps(beta2,rsq13);
 364             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
 365             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 366             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 367             felec            = _mm_mul_ps(qq13,felec);
 368             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 369             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv13);
 370             velec            = _mm_mul_ps(qq13,velec);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velecsum         = _mm_add_ps(velecsum,velec);
 374
 375             fscal            = felec;
 376
 377              /* Update vectorial force */
 378             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 379             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 380             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 381
 382             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 383             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 384             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             r21              = _mm_mul_ps(rsq21,rinv21);
 391
 392             /* EWALD ELECTROSTATICS */
 393
 394             /* Analytical PME correction */
 395             zeta2            = _mm_mul_ps(beta2,rsq21);
 396             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 397             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 398             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 399             felec            = _mm_mul_ps(qq21,felec);
 400             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 401             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 402             velec            = _mm_mul_ps(qq21,velec);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velecsum         = _mm_add_ps(velecsum,velec);
 406
 407             fscal            = felec;
 408
 409              /* Update vectorial force */
 410             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 411             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 412             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 413
 414             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 415             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 416             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 417
 418             /**************************
 419              * CALCULATE INTERACTIONS *
 420              **************************/
 421
 422             r22              = _mm_mul_ps(rsq22,rinv22);
 423
 424             /* EWALD ELECTROSTATICS */
 425
 426             /* Analytical PME correction */
 427             zeta2            = _mm_mul_ps(beta2,rsq22);
 428             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 429             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 430             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 431             felec            = _mm_mul_ps(qq22,felec);
 432             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 433             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 434             velec            = _mm_mul_ps(qq22,velec);
 435
 436             /* Update potential sum for this i atom from the interaction with this j atom. */
 437             velecsum         = _mm_add_ps(velecsum,velec);
 438
 439             fscal            = felec;
 440
 441              /* Update vectorial force */
 442             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 443             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 444             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 445
 446             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 447             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 448             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 449
 450             /**************************
 451              * CALCULATE INTERACTIONS *
 452              **************************/
 453
 454             r23              = _mm_mul_ps(rsq23,rinv23);
 455
 456             /* EWALD ELECTROSTATICS */
 457
 458             /* Analytical PME correction */
 459             zeta2            = _mm_mul_ps(beta2,rsq23);
 460             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
 461             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 462             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 463             felec            = _mm_mul_ps(qq23,felec);
 464             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 465             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv23);
 466             velec            = _mm_mul_ps(qq23,velec);
 467
 468             /* Update potential sum for this i atom from the interaction with this j atom. */
 469             velecsum         = _mm_add_ps(velecsum,velec);
 470
 471             fscal            = felec;
 472
 473              /* Update vectorial force */
 474             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 475             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 476             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 477
 478             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 479             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 480             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             r31              = _mm_mul_ps(rsq31,rinv31);
 487
 488             /* EWALD ELECTROSTATICS */
 489
 490             /* Analytical PME correction */
 491             zeta2            = _mm_mul_ps(beta2,rsq31);
 492             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
 493             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 494             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 495             felec            = _mm_mul_ps(qq31,felec);
 496             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 497             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv31);
 498             velec            = _mm_mul_ps(qq31,velec);
 499
 500             /* Update potential sum for this i atom from the interaction with this j atom. */
 501             velecsum         = _mm_add_ps(velecsum,velec);
 502
 503             fscal            = felec;
 504
 505              /* Update vectorial force */
 506             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 507             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 508             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 509
 510             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 511             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 512             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 513
 514             /**************************
 515              * CALCULATE INTERACTIONS *
 516              **************************/
 517
 518             r32              = _mm_mul_ps(rsq32,rinv32);
 519
 520             /* EWALD ELECTROSTATICS */
 521
 522             /* Analytical PME correction */
 523             zeta2            = _mm_mul_ps(beta2,rsq32);
 524             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
 525             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 526             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 527             felec            = _mm_mul_ps(qq32,felec);
 528             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 529             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv32);
 530             velec            = _mm_mul_ps(qq32,velec);
 531
 532             /* Update potential sum for this i atom from the interaction with this j atom. */
 533             velecsum         = _mm_add_ps(velecsum,velec);
 534
 535             fscal            = felec;
 536
 537              /* Update vectorial force */
 538             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 539             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 540             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 541
 542             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 543             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 544             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 545
 546             /**************************
 547              * CALCULATE INTERACTIONS *
 548              **************************/
 549
 550             r33              = _mm_mul_ps(rsq33,rinv33);
 551
 552             /* EWALD ELECTROSTATICS */
 553
 554             /* Analytical PME correction */
 555             zeta2            = _mm_mul_ps(beta2,rsq33);
 556             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
 557             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 558             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 559             felec            = _mm_mul_ps(qq33,felec);
 560             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 561             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv33);
 562             velec            = _mm_mul_ps(qq33,velec);
 563
 564             /* Update potential sum for this i atom from the interaction with this j atom. */
 565             velecsum         = _mm_add_ps(velecsum,velec);
 566
 567             fscal            = felec;
 568
 569              /* Update vectorial force */
 570             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 571             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 572             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 573
 574             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 575             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 576             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 577
 578             fjptrA             = f+j_coord_offsetA;
 579             fjptrB             = f+j_coord_offsetB;
 580             fjptrC             = f+j_coord_offsetC;
 581             fjptrD             = f+j_coord_offsetD;
 582
 583             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 584                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 585
 586             /* Inner loop uses 261 flops */
 587         }
 588
 589         if(jidx<j_index_end)
 590         {
 591
 592             /* Get j neighbor index, and coordinate index */
 593             jnrlistA         = jjnr[jidx];
 594             jnrlistB         = jjnr[jidx+1];
 595             jnrlistC         = jjnr[jidx+2];
 596             jnrlistD         = jjnr[jidx+3];
 597             /* Sign of each element will be negative for non-real atoms.
 598              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 599              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 600              */
 601             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 602             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 603             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 604             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 605             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 606             j_coord_offsetA  = DIM*jnrA;
 607             j_coord_offsetB  = DIM*jnrB;
 608             j_coord_offsetC  = DIM*jnrC;
 609             j_coord_offsetD  = DIM*jnrD;
 610
 611             /* load j atom coordinates */
 612             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 613                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 614                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 615
 616             /* Calculate displacement vector */
 617             dx11             = _mm_sub_ps(ix1,jx1);
 618             dy11             = _mm_sub_ps(iy1,jy1);
 619             dz11             = _mm_sub_ps(iz1,jz1);
 620             dx12             = _mm_sub_ps(ix1,jx2);
 621             dy12             = _mm_sub_ps(iy1,jy2);
 622             dz12             = _mm_sub_ps(iz1,jz2);
 623             dx13             = _mm_sub_ps(ix1,jx3);
 624             dy13             = _mm_sub_ps(iy1,jy3);
 625             dz13             = _mm_sub_ps(iz1,jz3);
 626             dx21             = _mm_sub_ps(ix2,jx1);
 627             dy21             = _mm_sub_ps(iy2,jy1);
 628             dz21             = _mm_sub_ps(iz2,jz1);
 629             dx22             = _mm_sub_ps(ix2,jx2);
 630             dy22             = _mm_sub_ps(iy2,jy2);
 631             dz22             = _mm_sub_ps(iz2,jz2);
 632             dx23             = _mm_sub_ps(ix2,jx3);
 633             dy23             = _mm_sub_ps(iy2,jy3);
 634             dz23             = _mm_sub_ps(iz2,jz3);
 635             dx31             = _mm_sub_ps(ix3,jx1);
 636             dy31             = _mm_sub_ps(iy3,jy1);
 637             dz31             = _mm_sub_ps(iz3,jz1);
 638             dx32             = _mm_sub_ps(ix3,jx2);
 639             dy32             = _mm_sub_ps(iy3,jy2);
 640             dz32             = _mm_sub_ps(iz3,jz2);
 641             dx33             = _mm_sub_ps(ix3,jx3);
 642             dy33             = _mm_sub_ps(iy3,jy3);
 643             dz33             = _mm_sub_ps(iz3,jz3);
 644
 645             /* Calculate squared distance and things based on it */
 646             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 647             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 648             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 649             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 650             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 651             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 652             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 653             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 654             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 655
 656             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 657             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 658             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 659             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 660             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 661             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 662             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 663             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 664             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 665
 666             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 667             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 668             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 669             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 670             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 671             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 672             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 673             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 674             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 675
 676             fjx1             = _mm_setzero_ps();
 677             fjy1             = _mm_setzero_ps();
 678             fjz1             = _mm_setzero_ps();
 679             fjx2             = _mm_setzero_ps();
 680             fjy2             = _mm_setzero_ps();
 681             fjz2             = _mm_setzero_ps();
 682             fjx3             = _mm_setzero_ps();
 683             fjy3             = _mm_setzero_ps();
 684             fjz3             = _mm_setzero_ps();
 685
 686             /**************************
 687              * CALCULATE INTERACTIONS *
 688              **************************/
 689
 690             r11              = _mm_mul_ps(rsq11,rinv11);
 691             r11              = _mm_andnot_ps(dummy_mask,r11);
 692
 693             /* EWALD ELECTROSTATICS */
 694
 695             /* Analytical PME correction */
 696             zeta2            = _mm_mul_ps(beta2,rsq11);
 697             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 698             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 699             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 700             felec            = _mm_mul_ps(qq11,felec);
 701             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 702             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv11);
 703             velec            = _mm_mul_ps(qq11,velec);
 704
 705             /* Update potential sum for this i atom from the interaction with this j atom. */
 706             velec            = _mm_andnot_ps(dummy_mask,velec);
 707             velecsum         = _mm_add_ps(velecsum,velec);
 708
 709             fscal            = felec;
 710
 711             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 712
 713              /* Update vectorial force */
 714             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 715             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 716             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 717
 718             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 719             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 720             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 721
 722             /**************************
 723              * CALCULATE INTERACTIONS *
 724              **************************/
 725
 726             r12              = _mm_mul_ps(rsq12,rinv12);
 727             r12              = _mm_andnot_ps(dummy_mask,r12);
 728
 729             /* EWALD ELECTROSTATICS */
 730
 731             /* Analytical PME correction */
 732             zeta2            = _mm_mul_ps(beta2,rsq12);
 733             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 734             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 735             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 736             felec            = _mm_mul_ps(qq12,felec);
 737             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 738             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv12);
 739             velec            = _mm_mul_ps(qq12,velec);
 740
 741             /* Update potential sum for this i atom from the interaction with this j atom. */
 742             velec            = _mm_andnot_ps(dummy_mask,velec);
 743             velecsum         = _mm_add_ps(velecsum,velec);
 744
 745             fscal            = felec;
 746
 747             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 748
 749              /* Update vectorial force */
 750             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 751             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 752             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 753
 754             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 755             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 756             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 757
 758             /**************************
 759              * CALCULATE INTERACTIONS *
 760              **************************/
 761
 762             r13              = _mm_mul_ps(rsq13,rinv13);
 763             r13              = _mm_andnot_ps(dummy_mask,r13);
 764
 765             /* EWALD ELECTROSTATICS */
 766
 767             /* Analytical PME correction */
 768             zeta2            = _mm_mul_ps(beta2,rsq13);
 769             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
 770             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 771             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 772             felec            = _mm_mul_ps(qq13,felec);
 773             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 774             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv13);
 775             velec            = _mm_mul_ps(qq13,velec);
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_andnot_ps(dummy_mask,velec);
 779             velecsum         = _mm_add_ps(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 784
 785              /* Update vectorial force */
 786             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 787             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 788             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 789
 790             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 791             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 792             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             r21              = _mm_mul_ps(rsq21,rinv21);
 799             r21              = _mm_andnot_ps(dummy_mask,r21);
 800
 801             /* EWALD ELECTROSTATICS */
 802
 803             /* Analytical PME correction */
 804             zeta2            = _mm_mul_ps(beta2,rsq21);
 805             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 806             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 807             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 808             felec            = _mm_mul_ps(qq21,felec);
 809             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 810             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv21);
 811             velec            = _mm_mul_ps(qq21,velec);
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_andnot_ps(dummy_mask,velec);
 815             velecsum         = _mm_add_ps(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 820
 821              /* Update vectorial force */
 822             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 823             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 824             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 825
 826             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 827             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 828             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             r22              = _mm_mul_ps(rsq22,rinv22);
 835             r22              = _mm_andnot_ps(dummy_mask,r22);
 836
 837             /* EWALD ELECTROSTATICS */
 838
 839             /* Analytical PME correction */
 840             zeta2            = _mm_mul_ps(beta2,rsq22);
 841             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 842             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 843             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 844             felec            = _mm_mul_ps(qq22,felec);
 845             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 846             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv22);
 847             velec            = _mm_mul_ps(qq22,velec);
 848
 849             /* Update potential sum for this i atom from the interaction with this j atom. */
 850             velec            = _mm_andnot_ps(dummy_mask,velec);
 851             velecsum         = _mm_add_ps(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 856
 857              /* Update vectorial force */
 858             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 859             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 860             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 861
 862             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 863             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 864             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 865
 866             /**************************
 867              * CALCULATE INTERACTIONS *
 868              **************************/
 869
 870             r23              = _mm_mul_ps(rsq23,rinv23);
 871             r23              = _mm_andnot_ps(dummy_mask,r23);
 872
 873             /* EWALD ELECTROSTATICS */
 874
 875             /* Analytical PME correction */
 876             zeta2            = _mm_mul_ps(beta2,rsq23);
 877             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
 878             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 879             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 880             felec            = _mm_mul_ps(qq23,felec);
 881             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 882             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv23);
 883             velec            = _mm_mul_ps(qq23,velec);
 884
 885             /* Update potential sum for this i atom from the interaction with this j atom. */
 886             velec            = _mm_andnot_ps(dummy_mask,velec);
 887             velecsum         = _mm_add_ps(velecsum,velec);
 888
 889             fscal            = felec;
 890
 891             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 892
 893              /* Update vectorial force */
 894             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 895             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 896             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 897
 898             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 899             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 900             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 901
 902             /**************************
 903              * CALCULATE INTERACTIONS *
 904              **************************/
 905
 906             r31              = _mm_mul_ps(rsq31,rinv31);
 907             r31              = _mm_andnot_ps(dummy_mask,r31);
 908
 909             /* EWALD ELECTROSTATICS */
 910
 911             /* Analytical PME correction */
 912             zeta2            = _mm_mul_ps(beta2,rsq31);
 913             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
 914             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 915             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 916             felec            = _mm_mul_ps(qq31,felec);
 917             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 918             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv31);
 919             velec            = _mm_mul_ps(qq31,velec);
 920
 921             /* Update potential sum for this i atom from the interaction with this j atom. */
 922             velec            = _mm_andnot_ps(dummy_mask,velec);
 923             velecsum         = _mm_add_ps(velecsum,velec);
 924
 925             fscal            = felec;
 926
 927             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 928
 929              /* Update vectorial force */
 930             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 931             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 932             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 933
 934             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 935             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 936             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 937
 938             /**************************
 939              * CALCULATE INTERACTIONS *
 940              **************************/
 941
 942             r32              = _mm_mul_ps(rsq32,rinv32);
 943             r32              = _mm_andnot_ps(dummy_mask,r32);
 944
 945             /* EWALD ELECTROSTATICS */
 946
 947             /* Analytical PME correction */
 948             zeta2            = _mm_mul_ps(beta2,rsq32);
 949             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
 950             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 951             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 952             felec            = _mm_mul_ps(qq32,felec);
 953             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 954             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv32);
 955             velec            = _mm_mul_ps(qq32,velec);
 956
 957             /* Update potential sum for this i atom from the interaction with this j atom. */
 958             velec            = _mm_andnot_ps(dummy_mask,velec);
 959             velecsum         = _mm_add_ps(velecsum,velec);
 960
 961             fscal            = felec;
 962
 963             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 964
 965              /* Update vectorial force */
 966             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 967             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 968             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 969
 970             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 971             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 972             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 973
 974             /**************************
 975              * CALCULATE INTERACTIONS *
 976              **************************/
 977
 978             r33              = _mm_mul_ps(rsq33,rinv33);
 979             r33              = _mm_andnot_ps(dummy_mask,r33);
 980
 981             /* EWALD ELECTROSTATICS */
 982
 983             /* Analytical PME correction */
 984             zeta2            = _mm_mul_ps(beta2,rsq33);
 985             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
 986             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 987             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 988             felec            = _mm_mul_ps(qq33,felec);
 989             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 990             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv33);
 991             velec            = _mm_mul_ps(qq33,velec);
 992
 993             /* Update potential sum for this i atom from the interaction with this j atom. */
 994             velec            = _mm_andnot_ps(dummy_mask,velec);
 995             velecsum         = _mm_add_ps(velecsum,velec);
 996
 997             fscal            = felec;
 998
 999             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1000
1001              /* Update vectorial force */
1002             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1003             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1004             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1005
1006             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1007             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1008             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1009
1010             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1011             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1012             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1013             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1014
1015             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1016                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1017
1018             /* Inner loop uses 270 flops */
1019         }
1020
1021         /* End of innermost loop */
1022
1023         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1024                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1025
1026         ggid                        = gid[iidx];
1027         /* Update potential energies */
1028         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1029
1030         /* Increment number of inner iterations */
1031         inneriter                  += j_index_end - j_index_start;
1032
1033         /* Outer loop uses 19 flops */
1034     }
1035
1036     /* Increment number of outer iterations */
1037     outeriter        += nri;
1038
1039     /* Update outer/inner flops */
1040
1041     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*270);
1042 }
1043 /*
1044  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwNone_GeomW4W4_F_avx_128_fma_single
1045  * Electrostatics interaction: Ewald
1046  * VdW interaction:            None
1047  * Geometry:                   Water4-Water4
1048  * Calculate force/pot:        Force
1049  */
1050 void
1051 nb_kernel_ElecEw_VdwNone_GeomW4W4_F_avx_128_fma_single
1052                     (t_nblist                    * gmx_restrict       nlist,
1053                      rvec                        * gmx_restrict          xx,
1054                      rvec                        * gmx_restrict          ff,
1055                      t_forcerec                  * gmx_restrict          fr,
1056                      t_mdatoms                   * gmx_restrict     mdatoms,
1057                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1058                      t_nrnb                      * gmx_restrict        nrnb)
1059 {
1060     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1061      * just 0 for non-waters.
1062      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1063      * jnr indices corresponding to data put in the four positions in the SIMD register.
1064      */
1065     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1066     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1067     int              jnrA,jnrB,jnrC,jnrD;
1068     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1069     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1070     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1071     real             rcutoff_scalar;
1072     real             *shiftvec,*fshift,*x,*f;
1073     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1074     real             scratch[4*DIM];
1075     __m128           fscal,rcutoff,rcutoff2,jidxall;
1076     int              vdwioffset1;
1077     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1078     int              vdwioffset2;
1079     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1080     int              vdwioffset3;
1081     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1082     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1083     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1084     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1085     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1086     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1087     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1088     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1089     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1090     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1091     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1092     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1093     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1094     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1095     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1096     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1097     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1098     real             *charge;
1099     __m128i          ewitab;
1100     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1101     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1102     real             *ewtab;
1103     __m128           dummy_mask,cutoff_mask;
1104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1105     __m128           one     = _mm_set1_ps(1.0);
1106     __m128           two     = _mm_set1_ps(2.0);
1107     x                = xx[0];
1108     f                = ff[0];
1109
1110     nri              = nlist->nri;
1111     iinr             = nlist->iinr;
1112     jindex           = nlist->jindex;
1113     jjnr             = nlist->jjnr;
1114     shiftidx         = nlist->shift;
1115     gid              = nlist->gid;
1116     shiftvec         = fr->shift_vec[0];
1117     fshift           = fr->fshift[0];
1118     facel            = _mm_set1_ps(fr->epsfac);
1119     charge           = mdatoms->chargeA;
1120
1121     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1122     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1123     beta2            = _mm_mul_ps(beta,beta);
1124     beta3            = _mm_mul_ps(beta,beta2);
1125     ewtab            = fr->ic->tabq_coul_F;
1126     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1127     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1128
1129     /* Setup water-specific parameters */
1130     inr              = nlist->iinr[0];
1131     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1132     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1133     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1134
1135     jq1              = _mm_set1_ps(charge[inr+1]);
1136     jq2              = _mm_set1_ps(charge[inr+2]);
1137     jq3              = _mm_set1_ps(charge[inr+3]);
1138     qq11             = _mm_mul_ps(iq1,jq1);
1139     qq12             = _mm_mul_ps(iq1,jq2);
1140     qq13             = _mm_mul_ps(iq1,jq3);
1141     qq21             = _mm_mul_ps(iq2,jq1);
1142     qq22             = _mm_mul_ps(iq2,jq2);
1143     qq23             = _mm_mul_ps(iq2,jq3);
1144     qq31             = _mm_mul_ps(iq3,jq1);
1145     qq32             = _mm_mul_ps(iq3,jq2);
1146     qq33             = _mm_mul_ps(iq3,jq3);
1147
1148     /* Avoid stupid compiler warnings */
1149     jnrA = jnrB = jnrC = jnrD = 0;
1150     j_coord_offsetA = 0;
1151     j_coord_offsetB = 0;
1152     j_coord_offsetC = 0;
1153     j_coord_offsetD = 0;
1154
1155     outeriter        = 0;
1156     inneriter        = 0;
1157
1158     for(iidx=0;iidx<4*DIM;iidx++)
1159     {
1160         scratch[iidx] = 0.0;
1161     }
1162
1163     /* Start outer loop over neighborlists */
1164     for(iidx=0; iidx<nri; iidx++)
1165     {
1166         /* Load shift vector for this list */
1167         i_shift_offset   = DIM*shiftidx[iidx];
1168
1169         /* Load limits for loop over neighbors */
1170         j_index_start    = jindex[iidx];
1171         j_index_end      = jindex[iidx+1];
1172
1173         /* Get outer coordinate index */
1174         inr              = iinr[iidx];
1175         i_coord_offset   = DIM*inr;
1176
1177         /* Load i particle coords and add shift vector */
1178         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1179                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1180
1181         fix1             = _mm_setzero_ps();
1182         fiy1             = _mm_setzero_ps();
1183         fiz1             = _mm_setzero_ps();
1184         fix2             = _mm_setzero_ps();
1185         fiy2             = _mm_setzero_ps();
1186         fiz2             = _mm_setzero_ps();
1187         fix3             = _mm_setzero_ps();
1188         fiy3             = _mm_setzero_ps();
1189         fiz3             = _mm_setzero_ps();
1190
1191         /* Start inner kernel loop */
1192         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1193         {
1194
1195             /* Get j neighbor index, and coordinate index */
1196             jnrA             = jjnr[jidx];
1197             jnrB             = jjnr[jidx+1];
1198             jnrC             = jjnr[jidx+2];
1199             jnrD             = jjnr[jidx+3];
1200             j_coord_offsetA  = DIM*jnrA;
1201             j_coord_offsetB  = DIM*jnrB;
1202             j_coord_offsetC  = DIM*jnrC;
1203             j_coord_offsetD  = DIM*jnrD;
1204
1205             /* load j atom coordinates */
1206             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1207                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1208                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1209
1210             /* Calculate displacement vector */
1211             dx11             = _mm_sub_ps(ix1,jx1);
1212             dy11             = _mm_sub_ps(iy1,jy1);
1213             dz11             = _mm_sub_ps(iz1,jz1);
1214             dx12             = _mm_sub_ps(ix1,jx2);
1215             dy12             = _mm_sub_ps(iy1,jy2);
1216             dz12             = _mm_sub_ps(iz1,jz2);
1217             dx13             = _mm_sub_ps(ix1,jx3);
1218             dy13             = _mm_sub_ps(iy1,jy3);
1219             dz13             = _mm_sub_ps(iz1,jz3);
1220             dx21             = _mm_sub_ps(ix2,jx1);
1221             dy21             = _mm_sub_ps(iy2,jy1);
1222             dz21             = _mm_sub_ps(iz2,jz1);
1223             dx22             = _mm_sub_ps(ix2,jx2);
1224             dy22             = _mm_sub_ps(iy2,jy2);
1225             dz22             = _mm_sub_ps(iz2,jz2);
1226             dx23             = _mm_sub_ps(ix2,jx3);
1227             dy23             = _mm_sub_ps(iy2,jy3);
1228             dz23             = _mm_sub_ps(iz2,jz3);
1229             dx31             = _mm_sub_ps(ix3,jx1);
1230             dy31             = _mm_sub_ps(iy3,jy1);
1231             dz31             = _mm_sub_ps(iz3,jz1);
1232             dx32             = _mm_sub_ps(ix3,jx2);
1233             dy32             = _mm_sub_ps(iy3,jy2);
1234             dz32             = _mm_sub_ps(iz3,jz2);
1235             dx33             = _mm_sub_ps(ix3,jx3);
1236             dy33             = _mm_sub_ps(iy3,jy3);
1237             dz33             = _mm_sub_ps(iz3,jz3);
1238
1239             /* Calculate squared distance and things based on it */
1240             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1241             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1242             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1243             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1244             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1245             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1246             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1247             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1248             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1249
1250             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1251             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1252             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1253             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1254             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1255             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1256             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1257             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1258             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1259
1260             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1261             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1262             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1263             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1264             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1265             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1266             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1267             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1268             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1269
1270             fjx1             = _mm_setzero_ps();
1271             fjy1             = _mm_setzero_ps();
1272             fjz1             = _mm_setzero_ps();
1273             fjx2             = _mm_setzero_ps();
1274             fjy2             = _mm_setzero_ps();
1275             fjz2             = _mm_setzero_ps();
1276             fjx3             = _mm_setzero_ps();
1277             fjy3             = _mm_setzero_ps();
1278             fjz3             = _mm_setzero_ps();
1279
1280             /**************************
1281              * CALCULATE INTERACTIONS *
1282              **************************/
1283
1284             r11              = _mm_mul_ps(rsq11,rinv11);
1285
1286             /* EWALD ELECTROSTATICS */
1287
1288             /* Analytical PME correction */
1289             zeta2            = _mm_mul_ps(beta2,rsq11);
1290             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1291             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1292             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1293             felec            = _mm_mul_ps(qq11,felec);
1294
1295             fscal            = felec;
1296
1297              /* Update vectorial force */
1298             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1299             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1300             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1301
1302             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1303             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1304             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             r12              = _mm_mul_ps(rsq12,rinv12);
1311
1312             /* EWALD ELECTROSTATICS */
1313
1314             /* Analytical PME correction */
1315             zeta2            = _mm_mul_ps(beta2,rsq12);
1316             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1317             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1318             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1319             felec            = _mm_mul_ps(qq12,felec);
1320
1321             fscal            = felec;
1322
1323              /* Update vectorial force */
1324             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1325             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1326             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1327
1328             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1329             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1330             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             r13              = _mm_mul_ps(rsq13,rinv13);
1337
1338             /* EWALD ELECTROSTATICS */
1339
1340             /* Analytical PME correction */
1341             zeta2            = _mm_mul_ps(beta2,rsq13);
1342             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
1343             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1344             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1345             felec            = _mm_mul_ps(qq13,felec);
1346
1347             fscal            = felec;
1348
1349              /* Update vectorial force */
1350             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1351             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1352             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1353
1354             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1355             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1356             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             r21              = _mm_mul_ps(rsq21,rinv21);
1363
1364             /* EWALD ELECTROSTATICS */
1365
1366             /* Analytical PME correction */
1367             zeta2            = _mm_mul_ps(beta2,rsq21);
1368             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1369             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1370             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1371             felec            = _mm_mul_ps(qq21,felec);
1372
1373             fscal            = felec;
1374
1375              /* Update vectorial force */
1376             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1377             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1378             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1379
1380             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1381             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1382             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             r22              = _mm_mul_ps(rsq22,rinv22);
1389
1390             /* EWALD ELECTROSTATICS */
1391
1392             /* Analytical PME correction */
1393             zeta2            = _mm_mul_ps(beta2,rsq22);
1394             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1395             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1396             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1397             felec            = _mm_mul_ps(qq22,felec);
1398
1399             fscal            = felec;
1400
1401              /* Update vectorial force */
1402             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1403             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1404             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1405
1406             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1407             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1408             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1409
1410             /**************************
1411              * CALCULATE INTERACTIONS *
1412              **************************/
1413
1414             r23              = _mm_mul_ps(rsq23,rinv23);
1415
1416             /* EWALD ELECTROSTATICS */
1417
1418             /* Analytical PME correction */
1419             zeta2            = _mm_mul_ps(beta2,rsq23);
1420             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
1421             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1422             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1423             felec            = _mm_mul_ps(qq23,felec);
1424
1425             fscal            = felec;
1426
1427              /* Update vectorial force */
1428             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1429             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1430             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1431
1432             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1433             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1434             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1435
1436             /**************************
1437              * CALCULATE INTERACTIONS *
1438              **************************/
1439
1440             r31              = _mm_mul_ps(rsq31,rinv31);
1441
1442             /* EWALD ELECTROSTATICS */
1443
1444             /* Analytical PME correction */
1445             zeta2            = _mm_mul_ps(beta2,rsq31);
1446             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
1447             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1448             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1449             felec            = _mm_mul_ps(qq31,felec);
1450
1451             fscal            = felec;
1452
1453              /* Update vectorial force */
1454             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1455             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1456             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1457
1458             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1459             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1460             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             r32              = _mm_mul_ps(rsq32,rinv32);
1467
1468             /* EWALD ELECTROSTATICS */
1469
1470             /* Analytical PME correction */
1471             zeta2            = _mm_mul_ps(beta2,rsq32);
1472             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
1473             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1474             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1475             felec            = _mm_mul_ps(qq32,felec);
1476
1477             fscal            = felec;
1478
1479              /* Update vectorial force */
1480             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1481             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1482             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1483
1484             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1485             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1486             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1487
1488             /**************************
1489              * CALCULATE INTERACTIONS *
1490              **************************/
1491
1492             r33              = _mm_mul_ps(rsq33,rinv33);
1493
1494             /* EWALD ELECTROSTATICS */
1495
1496             /* Analytical PME correction */
1497             zeta2            = _mm_mul_ps(beta2,rsq33);
1498             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
1499             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1500             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1501             felec            = _mm_mul_ps(qq33,felec);
1502
1503             fscal            = felec;
1504
1505              /* Update vectorial force */
1506             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1507             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1508             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1509
1510             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1511             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1512             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1513
1514             fjptrA             = f+j_coord_offsetA;
1515             fjptrB             = f+j_coord_offsetB;
1516             fjptrC             = f+j_coord_offsetC;
1517             fjptrD             = f+j_coord_offsetD;
1518
1519             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1520                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1521
1522             /* Inner loop uses 252 flops */
1523         }
1524
1525         if(jidx<j_index_end)
1526         {
1527
1528             /* Get j neighbor index, and coordinate index */
1529             jnrlistA         = jjnr[jidx];
1530             jnrlistB         = jjnr[jidx+1];
1531             jnrlistC         = jjnr[jidx+2];
1532             jnrlistD         = jjnr[jidx+3];
1533             /* Sign of each element will be negative for non-real atoms.
1534              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1535              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1536              */
1537             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1538             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1539             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1540             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1541             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1542             j_coord_offsetA  = DIM*jnrA;
1543             j_coord_offsetB  = DIM*jnrB;
1544             j_coord_offsetC  = DIM*jnrC;
1545             j_coord_offsetD  = DIM*jnrD;
1546
1547             /* load j atom coordinates */
1548             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1549                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1550                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1551
1552             /* Calculate displacement vector */
1553             dx11             = _mm_sub_ps(ix1,jx1);
1554             dy11             = _mm_sub_ps(iy1,jy1);
1555             dz11             = _mm_sub_ps(iz1,jz1);
1556             dx12             = _mm_sub_ps(ix1,jx2);
1557             dy12             = _mm_sub_ps(iy1,jy2);
1558             dz12             = _mm_sub_ps(iz1,jz2);
1559             dx13             = _mm_sub_ps(ix1,jx3);
1560             dy13             = _mm_sub_ps(iy1,jy3);
1561             dz13             = _mm_sub_ps(iz1,jz3);
1562             dx21             = _mm_sub_ps(ix2,jx1);
1563             dy21             = _mm_sub_ps(iy2,jy1);
1564             dz21             = _mm_sub_ps(iz2,jz1);
1565             dx22             = _mm_sub_ps(ix2,jx2);
1566             dy22             = _mm_sub_ps(iy2,jy2);
1567             dz22             = _mm_sub_ps(iz2,jz2);
1568             dx23             = _mm_sub_ps(ix2,jx3);
1569             dy23             = _mm_sub_ps(iy2,jy3);
1570             dz23             = _mm_sub_ps(iz2,jz3);
1571             dx31             = _mm_sub_ps(ix3,jx1);
1572             dy31             = _mm_sub_ps(iy3,jy1);
1573             dz31             = _mm_sub_ps(iz3,jz1);
1574             dx32             = _mm_sub_ps(ix3,jx2);
1575             dy32             = _mm_sub_ps(iy3,jy2);
1576             dz32             = _mm_sub_ps(iz3,jz2);
1577             dx33             = _mm_sub_ps(ix3,jx3);
1578             dy33             = _mm_sub_ps(iy3,jy3);
1579             dz33             = _mm_sub_ps(iz3,jz3);
1580
1581             /* Calculate squared distance and things based on it */
1582             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1583             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1584             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1585             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1586             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1587             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1588             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1589             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1590             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1591
1592             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1593             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1594             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1595             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1596             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1597             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1598             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1599             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1600             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1601
1602             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1603             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1604             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1605             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1606             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1607             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1608             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1609             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1610             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1611
1612             fjx1             = _mm_setzero_ps();
1613             fjy1             = _mm_setzero_ps();
1614             fjz1             = _mm_setzero_ps();
1615             fjx2             = _mm_setzero_ps();
1616             fjy2             = _mm_setzero_ps();
1617             fjz2             = _mm_setzero_ps();
1618             fjx3             = _mm_setzero_ps();
1619             fjy3             = _mm_setzero_ps();
1620             fjz3             = _mm_setzero_ps();
1621
1622             /**************************
1623              * CALCULATE INTERACTIONS *
1624              **************************/
1625
1626             r11              = _mm_mul_ps(rsq11,rinv11);
1627             r11              = _mm_andnot_ps(dummy_mask,r11);
1628
1629             /* EWALD ELECTROSTATICS */
1630
1631             /* Analytical PME correction */
1632             zeta2            = _mm_mul_ps(beta2,rsq11);
1633             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1634             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1635             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1636             felec            = _mm_mul_ps(qq11,felec);
1637
1638             fscal            = felec;
1639
1640             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1641
1642              /* Update vectorial force */
1643             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1644             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1645             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1646
1647             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1648             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1649             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1650
1651             /**************************
1652              * CALCULATE INTERACTIONS *
1653              **************************/
1654
1655             r12              = _mm_mul_ps(rsq12,rinv12);
1656             r12              = _mm_andnot_ps(dummy_mask,r12);
1657
1658             /* EWALD ELECTROSTATICS */
1659
1660             /* Analytical PME correction */
1661             zeta2            = _mm_mul_ps(beta2,rsq12);
1662             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1663             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1664             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1665             felec            = _mm_mul_ps(qq12,felec);
1666
1667             fscal            = felec;
1668
1669             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1670
1671              /* Update vectorial force */
1672             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1673             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1674             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1675
1676             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1677             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1678             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             r13              = _mm_mul_ps(rsq13,rinv13);
1685             r13              = _mm_andnot_ps(dummy_mask,r13);
1686
1687             /* EWALD ELECTROSTATICS */
1688
1689             /* Analytical PME correction */
1690             zeta2            = _mm_mul_ps(beta2,rsq13);
1691             rinv3            = _mm_mul_ps(rinvsq13,rinv13);
1692             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1693             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1694             felec            = _mm_mul_ps(qq13,felec);
1695
1696             fscal            = felec;
1697
1698             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1699
1700              /* Update vectorial force */
1701             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1702             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1703             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1704
1705             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1706             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1707             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1708
1709             /**************************
1710              * CALCULATE INTERACTIONS *
1711              **************************/
1712
1713             r21              = _mm_mul_ps(rsq21,rinv21);
1714             r21              = _mm_andnot_ps(dummy_mask,r21);
1715
1716             /* EWALD ELECTROSTATICS */
1717
1718             /* Analytical PME correction */
1719             zeta2            = _mm_mul_ps(beta2,rsq21);
1720             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1721             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1722             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1723             felec            = _mm_mul_ps(qq21,felec);
1724
1725             fscal            = felec;
1726
1727             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1728
1729              /* Update vectorial force */
1730             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1731             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1732             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1733
1734             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1735             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1736             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1737
1738             /**************************
1739              * CALCULATE INTERACTIONS *
1740              **************************/
1741
1742             r22              = _mm_mul_ps(rsq22,rinv22);
1743             r22              = _mm_andnot_ps(dummy_mask,r22);
1744
1745             /* EWALD ELECTROSTATICS */
1746
1747             /* Analytical PME correction */
1748             zeta2            = _mm_mul_ps(beta2,rsq22);
1749             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1750             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1751             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1752             felec            = _mm_mul_ps(qq22,felec);
1753
1754             fscal            = felec;
1755
1756             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1757
1758              /* Update vectorial force */
1759             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1760             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1761             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1762
1763             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1764             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1765             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1766
1767             /**************************
1768              * CALCULATE INTERACTIONS *
1769              **************************/
1770
1771             r23              = _mm_mul_ps(rsq23,rinv23);
1772             r23              = _mm_andnot_ps(dummy_mask,r23);
1773
1774             /* EWALD ELECTROSTATICS */
1775
1776             /* Analytical PME correction */
1777             zeta2            = _mm_mul_ps(beta2,rsq23);
1778             rinv3            = _mm_mul_ps(rinvsq23,rinv23);
1779             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1780             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1781             felec            = _mm_mul_ps(qq23,felec);
1782
1783             fscal            = felec;
1784
1785             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1786
1787              /* Update vectorial force */
1788             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1789             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1790             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1791
1792             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1793             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1794             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1795
1796             /**************************
1797              * CALCULATE INTERACTIONS *
1798              **************************/
1799
1800             r31              = _mm_mul_ps(rsq31,rinv31);
1801             r31              = _mm_andnot_ps(dummy_mask,r31);
1802
1803             /* EWALD ELECTROSTATICS */
1804
1805             /* Analytical PME correction */
1806             zeta2            = _mm_mul_ps(beta2,rsq31);
1807             rinv3            = _mm_mul_ps(rinvsq31,rinv31);
1808             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1809             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1810             felec            = _mm_mul_ps(qq31,felec);
1811
1812             fscal            = felec;
1813
1814             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1815
1816              /* Update vectorial force */
1817             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1818             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1819             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1820
1821             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1822             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1823             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1824
1825             /**************************
1826              * CALCULATE INTERACTIONS *
1827              **************************/
1828
1829             r32              = _mm_mul_ps(rsq32,rinv32);
1830             r32              = _mm_andnot_ps(dummy_mask,r32);
1831
1832             /* EWALD ELECTROSTATICS */
1833
1834             /* Analytical PME correction */
1835             zeta2            = _mm_mul_ps(beta2,rsq32);
1836             rinv3            = _mm_mul_ps(rinvsq32,rinv32);
1837             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1838             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1839             felec            = _mm_mul_ps(qq32,felec);
1840
1841             fscal            = felec;
1842
1843             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1844
1845              /* Update vectorial force */
1846             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1847             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1848             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1849
1850             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1851             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1852             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1853
1854             /**************************
1855              * CALCULATE INTERACTIONS *
1856              **************************/
1857
1858             r33              = _mm_mul_ps(rsq33,rinv33);
1859             r33              = _mm_andnot_ps(dummy_mask,r33);
1860
1861             /* EWALD ELECTROSTATICS */
1862
1863             /* Analytical PME correction */
1864             zeta2            = _mm_mul_ps(beta2,rsq33);
1865             rinv3            = _mm_mul_ps(rinvsq33,rinv33);
1866             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1867             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1868             felec            = _mm_mul_ps(qq33,felec);
1869
1870             fscal            = felec;
1871
1872             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1873
1874              /* Update vectorial force */
1875             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1876             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1877             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1878
1879             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1880             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1881             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1882
1883             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1884             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1885             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1886             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1887
1888             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1889                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1890
1891             /* Inner loop uses 261 flops */
1892         }
1893
1894         /* End of innermost loop */
1895
1896         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1897                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1898
1899         /* Increment number of inner iterations */
1900         inneriter                  += j_index_end - j_index_start;
1901
1902         /* Outer loop uses 18 flops */
1903     }
1904
1905     /* Increment number of outer iterations */
1906     outeriter        += nri;
1907
1908     /* Update outer/inner flops */
1909
1910     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*261);
1911 }