src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRF_VdwNone_GeomW4W4_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset1;
  83     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  89     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  91     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  93     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  94     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  95     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  96     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  97     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  98     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  99     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 100     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 101     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 102     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_ps(fr->ic->k_rf);
 123     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_ps(fr->ic->c_rf);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 129     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 130     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 131
 132     jq1              = _mm_set1_ps(charge[inr+1]);
 133     jq2              = _mm_set1_ps(charge[inr+2]);
 134     jq3              = _mm_set1_ps(charge[inr+3]);
 135     qq11             = _mm_mul_ps(iq1,jq1);
 136     qq12             = _mm_mul_ps(iq1,jq2);
 137     qq13             = _mm_mul_ps(iq1,jq3);
 138     qq21             = _mm_mul_ps(iq2,jq1);
 139     qq22             = _mm_mul_ps(iq2,jq2);
 140     qq23             = _mm_mul_ps(iq2,jq3);
 141     qq31             = _mm_mul_ps(iq3,jq1);
 142     qq32             = _mm_mul_ps(iq3,jq2);
 143     qq33             = _mm_mul_ps(iq3,jq3);
 144
 145     /* Avoid stupid compiler warnings */
 146     jnrA = jnrB = jnrC = jnrD = 0;
 147     j_coord_offsetA = 0;
 148     j_coord_offsetB = 0;
 149     j_coord_offsetC = 0;
 150     j_coord_offsetD = 0;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     for(iidx=0;iidx<4*DIM;iidx++)
 156     {
 157         scratch[iidx] = 0.0;
 158     }
 159
 160     /* Start outer loop over neighborlists */
 161     for(iidx=0; iidx<nri; iidx++)
 162     {
 163         /* Load shift vector for this list */
 164         i_shift_offset   = DIM*shiftidx[iidx];
 165
 166         /* Load limits for loop over neighbors */
 167         j_index_start    = jindex[iidx];
 168         j_index_end      = jindex[iidx+1];
 169
 170         /* Get outer coordinate index */
 171         inr              = iinr[iidx];
 172         i_coord_offset   = DIM*inr;
 173
 174         /* Load i particle coords and add shift vector */
 175         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 176                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 177
 178         fix1             = _mm_setzero_ps();
 179         fiy1             = _mm_setzero_ps();
 180         fiz1             = _mm_setzero_ps();
 181         fix2             = _mm_setzero_ps();
 182         fiy2             = _mm_setzero_ps();
 183         fiz2             = _mm_setzero_ps();
 184         fix3             = _mm_setzero_ps();
 185         fiy3             = _mm_setzero_ps();
 186         fiz3             = _mm_setzero_ps();
 187
 188         /* Reset potential sums */
 189         velecsum         = _mm_setzero_ps();
 190
 191         /* Start inner kernel loop */
 192         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 193         {
 194
 195             /* Get j neighbor index, and coordinate index */
 196             jnrA             = jjnr[jidx];
 197             jnrB             = jjnr[jidx+1];
 198             jnrC             = jjnr[jidx+2];
 199             jnrD             = jjnr[jidx+3];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202             j_coord_offsetC  = DIM*jnrC;
 203             j_coord_offsetD  = DIM*jnrD;
 204
 205             /* load j atom coordinates */
 206             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 207                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 208                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 209
 210             /* Calculate displacement vector */
 211             dx11             = _mm_sub_ps(ix1,jx1);
 212             dy11             = _mm_sub_ps(iy1,jy1);
 213             dz11             = _mm_sub_ps(iz1,jz1);
 214             dx12             = _mm_sub_ps(ix1,jx2);
 215             dy12             = _mm_sub_ps(iy1,jy2);
 216             dz12             = _mm_sub_ps(iz1,jz2);
 217             dx13             = _mm_sub_ps(ix1,jx3);
 218             dy13             = _mm_sub_ps(iy1,jy3);
 219             dz13             = _mm_sub_ps(iz1,jz3);
 220             dx21             = _mm_sub_ps(ix2,jx1);
 221             dy21             = _mm_sub_ps(iy2,jy1);
 222             dz21             = _mm_sub_ps(iz2,jz1);
 223             dx22             = _mm_sub_ps(ix2,jx2);
 224             dy22             = _mm_sub_ps(iy2,jy2);
 225             dz22             = _mm_sub_ps(iz2,jz2);
 226             dx23             = _mm_sub_ps(ix2,jx3);
 227             dy23             = _mm_sub_ps(iy2,jy3);
 228             dz23             = _mm_sub_ps(iz2,jz3);
 229             dx31             = _mm_sub_ps(ix3,jx1);
 230             dy31             = _mm_sub_ps(iy3,jy1);
 231             dz31             = _mm_sub_ps(iz3,jz1);
 232             dx32             = _mm_sub_ps(ix3,jx2);
 233             dy32             = _mm_sub_ps(iy3,jy2);
 234             dz32             = _mm_sub_ps(iz3,jz2);
 235             dx33             = _mm_sub_ps(ix3,jx3);
 236             dy33             = _mm_sub_ps(iy3,jy3);
 237             dz33             = _mm_sub_ps(iz3,jz3);
 238
 239             /* Calculate squared distance and things based on it */
 240             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 241             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 242             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 243             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 244             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 245             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 246             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 247             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 248             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 249
 250             rinv11           = avx128fma_invsqrt_f(rsq11);
 251             rinv12           = avx128fma_invsqrt_f(rsq12);
 252             rinv13           = avx128fma_invsqrt_f(rsq13);
 253             rinv21           = avx128fma_invsqrt_f(rsq21);
 254             rinv22           = avx128fma_invsqrt_f(rsq22);
 255             rinv23           = avx128fma_invsqrt_f(rsq23);
 256             rinv31           = avx128fma_invsqrt_f(rsq31);
 257             rinv32           = avx128fma_invsqrt_f(rsq32);
 258             rinv33           = avx128fma_invsqrt_f(rsq33);
 259
 260             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 261             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 262             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 263             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 264             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 265             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 266             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 267             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 268             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 269
 270             fjx1             = _mm_setzero_ps();
 271             fjy1             = _mm_setzero_ps();
 272             fjz1             = _mm_setzero_ps();
 273             fjx2             = _mm_setzero_ps();
 274             fjy2             = _mm_setzero_ps();
 275             fjz2             = _mm_setzero_ps();
 276             fjx3             = _mm_setzero_ps();
 277             fjy3             = _mm_setzero_ps();
 278             fjz3             = _mm_setzero_ps();
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             /* REACTION-FIELD ELECTROSTATICS */
 285             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 286             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 287
 288             /* Update potential sum for this i atom from the interaction with this j atom. */
 289             velecsum         = _mm_add_ps(velecsum,velec);
 290
 291             fscal            = felec;
 292
 293              /* Update vectorial force */
 294             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 295             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 296             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 297
 298             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 299             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 300             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             /* REACTION-FIELD ELECTROSTATICS */
 307             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 308             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312
 313             fscal            = felec;
 314
 315              /* Update vectorial force */
 316             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 317             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 318             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 319
 320             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 321             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 322             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             /* REACTION-FIELD ELECTROSTATICS */
 329             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 330             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 331
 332             /* Update potential sum for this i atom from the interaction with this j atom. */
 333             velecsum         = _mm_add_ps(velecsum,velec);
 334
 335             fscal            = felec;
 336
 337              /* Update vectorial force */
 338             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 339             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 340             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 341
 342             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 343             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 344             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 345
 346             /**************************
 347              * CALCULATE INTERACTIONS *
 348              **************************/
 349
 350             /* REACTION-FIELD ELECTROSTATICS */
 351             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 352             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 353
 354             /* Update potential sum for this i atom from the interaction with this j atom. */
 355             velecsum         = _mm_add_ps(velecsum,velec);
 356
 357             fscal            = felec;
 358
 359              /* Update vectorial force */
 360             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 361             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 362             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 363
 364             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 365             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 366             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 374             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velecsum         = _mm_add_ps(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381              /* Update vectorial force */
 382             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 383             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 384             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 385
 386             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 387             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 388             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 389
 390             /**************************
 391              * CALCULATE INTERACTIONS *
 392              **************************/
 393
 394             /* REACTION-FIELD ELECTROSTATICS */
 395             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 396             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 397
 398             /* Update potential sum for this i atom from the interaction with this j atom. */
 399             velecsum         = _mm_add_ps(velecsum,velec);
 400
 401             fscal            = felec;
 402
 403              /* Update vectorial force */
 404             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 405             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 406             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 407
 408             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 409             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 410             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 411
 412             /**************************
 413              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             /* REACTION-FIELD ELECTROSTATICS */
 417             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 418             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velecsum         = _mm_add_ps(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425              /* Update vectorial force */
 426             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 427             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 428             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 429
 430             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 431             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 432             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 440             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velecsum         = _mm_add_ps(velecsum,velec);
 444
 445             fscal            = felec;
 446
 447              /* Update vectorial force */
 448             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 449             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 450             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 451
 452             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 453             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 454             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             /* REACTION-FIELD ELECTROSTATICS */
 461             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 462             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 463
 464             /* Update potential sum for this i atom from the interaction with this j atom. */
 465             velecsum         = _mm_add_ps(velecsum,velec);
 466
 467             fscal            = felec;
 468
 469              /* Update vectorial force */
 470             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 471             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 472             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 473
 474             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 475             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 476             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 477
 478             fjptrA             = f+j_coord_offsetA;
 479             fjptrB             = f+j_coord_offsetB;
 480             fjptrC             = f+j_coord_offsetC;
 481             fjptrD             = f+j_coord_offsetD;
 482
 483             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 484                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 485
 486             /* Inner loop uses 315 flops */
 487         }
 488
 489         if(jidx<j_index_end)
 490         {
 491
 492             /* Get j neighbor index, and coordinate index */
 493             jnrlistA         = jjnr[jidx];
 494             jnrlistB         = jjnr[jidx+1];
 495             jnrlistC         = jjnr[jidx+2];
 496             jnrlistD         = jjnr[jidx+3];
 497             /* Sign of each element will be negative for non-real atoms.
 498              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 499              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 500              */
 501             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 502             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 503             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 504             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 505             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 506             j_coord_offsetA  = DIM*jnrA;
 507             j_coord_offsetB  = DIM*jnrB;
 508             j_coord_offsetC  = DIM*jnrC;
 509             j_coord_offsetD  = DIM*jnrD;
 510
 511             /* load j atom coordinates */
 512             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 513                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 514                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 515
 516             /* Calculate displacement vector */
 517             dx11             = _mm_sub_ps(ix1,jx1);
 518             dy11             = _mm_sub_ps(iy1,jy1);
 519             dz11             = _mm_sub_ps(iz1,jz1);
 520             dx12             = _mm_sub_ps(ix1,jx2);
 521             dy12             = _mm_sub_ps(iy1,jy2);
 522             dz12             = _mm_sub_ps(iz1,jz2);
 523             dx13             = _mm_sub_ps(ix1,jx3);
 524             dy13             = _mm_sub_ps(iy1,jy3);
 525             dz13             = _mm_sub_ps(iz1,jz3);
 526             dx21             = _mm_sub_ps(ix2,jx1);
 527             dy21             = _mm_sub_ps(iy2,jy1);
 528             dz21             = _mm_sub_ps(iz2,jz1);
 529             dx22             = _mm_sub_ps(ix2,jx2);
 530             dy22             = _mm_sub_ps(iy2,jy2);
 531             dz22             = _mm_sub_ps(iz2,jz2);
 532             dx23             = _mm_sub_ps(ix2,jx3);
 533             dy23             = _mm_sub_ps(iy2,jy3);
 534             dz23             = _mm_sub_ps(iz2,jz3);
 535             dx31             = _mm_sub_ps(ix3,jx1);
 536             dy31             = _mm_sub_ps(iy3,jy1);
 537             dz31             = _mm_sub_ps(iz3,jz1);
 538             dx32             = _mm_sub_ps(ix3,jx2);
 539             dy32             = _mm_sub_ps(iy3,jy2);
 540             dz32             = _mm_sub_ps(iz3,jz2);
 541             dx33             = _mm_sub_ps(ix3,jx3);
 542             dy33             = _mm_sub_ps(iy3,jy3);
 543             dz33             = _mm_sub_ps(iz3,jz3);
 544
 545             /* Calculate squared distance and things based on it */
 546             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 547             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 548             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 549             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 550             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 551             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 552             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 553             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 554             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 555
 556             rinv11           = avx128fma_invsqrt_f(rsq11);
 557             rinv12           = avx128fma_invsqrt_f(rsq12);
 558             rinv13           = avx128fma_invsqrt_f(rsq13);
 559             rinv21           = avx128fma_invsqrt_f(rsq21);
 560             rinv22           = avx128fma_invsqrt_f(rsq22);
 561             rinv23           = avx128fma_invsqrt_f(rsq23);
 562             rinv31           = avx128fma_invsqrt_f(rsq31);
 563             rinv32           = avx128fma_invsqrt_f(rsq32);
 564             rinv33           = avx128fma_invsqrt_f(rsq33);
 565
 566             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 567             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 568             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 569             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 570             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 571             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 572             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 573             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 574             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 575
 576             fjx1             = _mm_setzero_ps();
 577             fjy1             = _mm_setzero_ps();
 578             fjz1             = _mm_setzero_ps();
 579             fjx2             = _mm_setzero_ps();
 580             fjy2             = _mm_setzero_ps();
 581             fjz2             = _mm_setzero_ps();
 582             fjx3             = _mm_setzero_ps();
 583             fjy3             = _mm_setzero_ps();
 584             fjz3             = _mm_setzero_ps();
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             /* REACTION-FIELD ELECTROSTATICS */
 591             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 592             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 593
 594             /* Update potential sum for this i atom from the interaction with this j atom. */
 595             velec            = _mm_andnot_ps(dummy_mask,velec);
 596             velecsum         = _mm_add_ps(velecsum,velec);
 597
 598             fscal            = felec;
 599
 600             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 601
 602              /* Update vectorial force */
 603             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 604             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 605             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 606
 607             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 608             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 609             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 610
 611             /**************************
 612              * CALCULATE INTERACTIONS *
 613              **************************/
 614
 615             /* REACTION-FIELD ELECTROSTATICS */
 616             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 617             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 618
 619             /* Update potential sum for this i atom from the interaction with this j atom. */
 620             velec            = _mm_andnot_ps(dummy_mask,velec);
 621             velecsum         = _mm_add_ps(velecsum,velec);
 622
 623             fscal            = felec;
 624
 625             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 626
 627              /* Update vectorial force */
 628             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 629             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 630             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 631
 632             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 633             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 634             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 635
 636             /**************************
 637              * CALCULATE INTERACTIONS *
 638              **************************/
 639
 640             /* REACTION-FIELD ELECTROSTATICS */
 641             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 642             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 643
 644             /* Update potential sum for this i atom from the interaction with this j atom. */
 645             velec            = _mm_andnot_ps(dummy_mask,velec);
 646             velecsum         = _mm_add_ps(velecsum,velec);
 647
 648             fscal            = felec;
 649
 650             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 651
 652              /* Update vectorial force */
 653             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 654             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 655             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 656
 657             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 658             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 659             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 660
 661             /**************************
 662              * CALCULATE INTERACTIONS *
 663              **************************/
 664
 665             /* REACTION-FIELD ELECTROSTATICS */
 666             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 667             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 668
 669             /* Update potential sum for this i atom from the interaction with this j atom. */
 670             velec            = _mm_andnot_ps(dummy_mask,velec);
 671             velecsum         = _mm_add_ps(velecsum,velec);
 672
 673             fscal            = felec;
 674
 675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 676
 677              /* Update vectorial force */
 678             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 679             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 680             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 681
 682             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 683             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 684             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 685
 686             /**************************
 687              * CALCULATE INTERACTIONS *
 688              **************************/
 689
 690             /* REACTION-FIELD ELECTROSTATICS */
 691             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 692             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 693
 694             /* Update potential sum for this i atom from the interaction with this j atom. */
 695             velec            = _mm_andnot_ps(dummy_mask,velec);
 696             velecsum         = _mm_add_ps(velecsum,velec);
 697
 698             fscal            = felec;
 699
 700             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 701
 702              /* Update vectorial force */
 703             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 704             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 705             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 706
 707             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 708             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 709             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             /* REACTION-FIELD ELECTROSTATICS */
 716             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 717             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 718
 719             /* Update potential sum for this i atom from the interaction with this j atom. */
 720             velec            = _mm_andnot_ps(dummy_mask,velec);
 721             velecsum         = _mm_add_ps(velecsum,velec);
 722
 723             fscal            = felec;
 724
 725             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 726
 727              /* Update vectorial force */
 728             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 729             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 730             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 731
 732             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 733             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 734             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 735
 736             /**************************
 737              * CALCULATE INTERACTIONS *
 738              **************************/
 739
 740             /* REACTION-FIELD ELECTROSTATICS */
 741             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 742             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 743
 744             /* Update potential sum for this i atom from the interaction with this j atom. */
 745             velec            = _mm_andnot_ps(dummy_mask,velec);
 746             velecsum         = _mm_add_ps(velecsum,velec);
 747
 748             fscal            = felec;
 749
 750             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 751
 752              /* Update vectorial force */
 753             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 754             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 755             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 756
 757             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 758             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 759             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 760
 761             /**************************
 762              * CALCULATE INTERACTIONS *
 763              **************************/
 764
 765             /* REACTION-FIELD ELECTROSTATICS */
 766             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 767             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 768
 769             /* Update potential sum for this i atom from the interaction with this j atom. */
 770             velec            = _mm_andnot_ps(dummy_mask,velec);
 771             velecsum         = _mm_add_ps(velecsum,velec);
 772
 773             fscal            = felec;
 774
 775             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 776
 777              /* Update vectorial force */
 778             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 779             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 780             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 781
 782             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 783             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 784             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 785
 786             /**************************
 787              * CALCULATE INTERACTIONS *
 788              **************************/
 789
 790             /* REACTION-FIELD ELECTROSTATICS */
 791             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 792             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 793
 794             /* Update potential sum for this i atom from the interaction with this j atom. */
 795             velec            = _mm_andnot_ps(dummy_mask,velec);
 796             velecsum         = _mm_add_ps(velecsum,velec);
 797
 798             fscal            = felec;
 799
 800             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 801
 802              /* Update vectorial force */
 803             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 804             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 805             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 806
 807             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 808             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 809             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 810
 811             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 812             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 813             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 814             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 815
 816             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 817                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 818
 819             /* Inner loop uses 315 flops */
 820         }
 821
 822         /* End of innermost loop */
 823
 824         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 825                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 826
 827         ggid                        = gid[iidx];
 828         /* Update potential energies */
 829         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 830
 831         /* Increment number of inner iterations */
 832         inneriter                  += j_index_end - j_index_start;
 833
 834         /* Outer loop uses 19 flops */
 835     }
 836
 837     /* Increment number of outer iterations */
 838     outeriter        += nri;
 839
 840     /* Update outer/inner flops */
 841
 842     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*315);
 843 }
 844 /*
 845  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_avx_128_fma_single
 846  * Electrostatics interaction: ReactionField
 847  * VdW interaction:            None
 848  * Geometry:                   Water4-Water4
 849  * Calculate force/pot:        Force
 850  */
 851 void
 852 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_avx_128_fma_single
 853                     (t_nblist                    * gmx_restrict       nlist,
 854                      rvec                        * gmx_restrict          xx,
 855                      rvec                        * gmx_restrict          ff,
 856                      struct t_forcerec           * gmx_restrict          fr,
 857                      t_mdatoms                   * gmx_restrict     mdatoms,
 858                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 859                      t_nrnb                      * gmx_restrict        nrnb)
 860 {
 861     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 862      * just 0 for non-waters.
 863      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 864      * jnr indices corresponding to data put in the four positions in the SIMD register.
 865      */
 866     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 867     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 868     int              jnrA,jnrB,jnrC,jnrD;
 869     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 870     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 871     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 872     real             rcutoff_scalar;
 873     real             *shiftvec,*fshift,*x,*f;
 874     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 875     real             scratch[4*DIM];
 876     __m128           fscal,rcutoff,rcutoff2,jidxall;
 877     int              vdwioffset1;
 878     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 879     int              vdwioffset2;
 880     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 881     int              vdwioffset3;
 882     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 883     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 884     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 885     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 886     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 887     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 888     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 889     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 890     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 891     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 892     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 893     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 894     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 895     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 896     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 897     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 898     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 899     real             *charge;
 900     __m128           dummy_mask,cutoff_mask;
 901     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 902     __m128           one     = _mm_set1_ps(1.0);
 903     __m128           two     = _mm_set1_ps(2.0);
 904     x                = xx[0];
 905     f                = ff[0];
 906
 907     nri              = nlist->nri;
 908     iinr             = nlist->iinr;
 909     jindex           = nlist->jindex;
 910     jjnr             = nlist->jjnr;
 911     shiftidx         = nlist->shift;
 912     gid              = nlist->gid;
 913     shiftvec         = fr->shift_vec[0];
 914     fshift           = fr->fshift[0];
 915     facel            = _mm_set1_ps(fr->ic->epsfac);
 916     charge           = mdatoms->chargeA;
 917     krf              = _mm_set1_ps(fr->ic->k_rf);
 918     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 919     crf              = _mm_set1_ps(fr->ic->c_rf);
 920
 921     /* Setup water-specific parameters */
 922     inr              = nlist->iinr[0];
 923     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 924     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 925     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 926
 927     jq1              = _mm_set1_ps(charge[inr+1]);
 928     jq2              = _mm_set1_ps(charge[inr+2]);
 929     jq3              = _mm_set1_ps(charge[inr+3]);
 930     qq11             = _mm_mul_ps(iq1,jq1);
 931     qq12             = _mm_mul_ps(iq1,jq2);
 932     qq13             = _mm_mul_ps(iq1,jq3);
 933     qq21             = _mm_mul_ps(iq2,jq1);
 934     qq22             = _mm_mul_ps(iq2,jq2);
 935     qq23             = _mm_mul_ps(iq2,jq3);
 936     qq31             = _mm_mul_ps(iq3,jq1);
 937     qq32             = _mm_mul_ps(iq3,jq2);
 938     qq33             = _mm_mul_ps(iq3,jq3);
 939
 940     /* Avoid stupid compiler warnings */
 941     jnrA = jnrB = jnrC = jnrD = 0;
 942     j_coord_offsetA = 0;
 943     j_coord_offsetB = 0;
 944     j_coord_offsetC = 0;
 945     j_coord_offsetD = 0;
 946
 947     outeriter        = 0;
 948     inneriter        = 0;
 949
 950     for(iidx=0;iidx<4*DIM;iidx++)
 951     {
 952         scratch[iidx] = 0.0;
 953     }
 954
 955     /* Start outer loop over neighborlists */
 956     for(iidx=0; iidx<nri; iidx++)
 957     {
 958         /* Load shift vector for this list */
 959         i_shift_offset   = DIM*shiftidx[iidx];
 960
 961         /* Load limits for loop over neighbors */
 962         j_index_start    = jindex[iidx];
 963         j_index_end      = jindex[iidx+1];
 964
 965         /* Get outer coordinate index */
 966         inr              = iinr[iidx];
 967         i_coord_offset   = DIM*inr;
 968
 969         /* Load i particle coords and add shift vector */
 970         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 971                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 972
 973         fix1             = _mm_setzero_ps();
 974         fiy1             = _mm_setzero_ps();
 975         fiz1             = _mm_setzero_ps();
 976         fix2             = _mm_setzero_ps();
 977         fiy2             = _mm_setzero_ps();
 978         fiz2             = _mm_setzero_ps();
 979         fix3             = _mm_setzero_ps();
 980         fiy3             = _mm_setzero_ps();
 981         fiz3             = _mm_setzero_ps();
 982
 983         /* Start inner kernel loop */
 984         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 985         {
 986
 987             /* Get j neighbor index, and coordinate index */
 988             jnrA             = jjnr[jidx];
 989             jnrB             = jjnr[jidx+1];
 990             jnrC             = jjnr[jidx+2];
 991             jnrD             = jjnr[jidx+3];
 992             j_coord_offsetA  = DIM*jnrA;
 993             j_coord_offsetB  = DIM*jnrB;
 994             j_coord_offsetC  = DIM*jnrC;
 995             j_coord_offsetD  = DIM*jnrD;
 996
 997             /* load j atom coordinates */
 998             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 999                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1000                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1001
1002             /* Calculate displacement vector */
1003             dx11             = _mm_sub_ps(ix1,jx1);
1004             dy11             = _mm_sub_ps(iy1,jy1);
1005             dz11             = _mm_sub_ps(iz1,jz1);
1006             dx12             = _mm_sub_ps(ix1,jx2);
1007             dy12             = _mm_sub_ps(iy1,jy2);
1008             dz12             = _mm_sub_ps(iz1,jz2);
1009             dx13             = _mm_sub_ps(ix1,jx3);
1010             dy13             = _mm_sub_ps(iy1,jy3);
1011             dz13             = _mm_sub_ps(iz1,jz3);
1012             dx21             = _mm_sub_ps(ix2,jx1);
1013             dy21             = _mm_sub_ps(iy2,jy1);
1014             dz21             = _mm_sub_ps(iz2,jz1);
1015             dx22             = _mm_sub_ps(ix2,jx2);
1016             dy22             = _mm_sub_ps(iy2,jy2);
1017             dz22             = _mm_sub_ps(iz2,jz2);
1018             dx23             = _mm_sub_ps(ix2,jx3);
1019             dy23             = _mm_sub_ps(iy2,jy3);
1020             dz23             = _mm_sub_ps(iz2,jz3);
1021             dx31             = _mm_sub_ps(ix3,jx1);
1022             dy31             = _mm_sub_ps(iy3,jy1);
1023             dz31             = _mm_sub_ps(iz3,jz1);
1024             dx32             = _mm_sub_ps(ix3,jx2);
1025             dy32             = _mm_sub_ps(iy3,jy2);
1026             dz32             = _mm_sub_ps(iz3,jz2);
1027             dx33             = _mm_sub_ps(ix3,jx3);
1028             dy33             = _mm_sub_ps(iy3,jy3);
1029             dz33             = _mm_sub_ps(iz3,jz3);
1030
1031             /* Calculate squared distance and things based on it */
1032             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1033             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1034             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1035             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1036             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1037             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1038             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1039             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1040             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1041
1042             rinv11           = avx128fma_invsqrt_f(rsq11);
1043             rinv12           = avx128fma_invsqrt_f(rsq12);
1044             rinv13           = avx128fma_invsqrt_f(rsq13);
1045             rinv21           = avx128fma_invsqrt_f(rsq21);
1046             rinv22           = avx128fma_invsqrt_f(rsq22);
1047             rinv23           = avx128fma_invsqrt_f(rsq23);
1048             rinv31           = avx128fma_invsqrt_f(rsq31);
1049             rinv32           = avx128fma_invsqrt_f(rsq32);
1050             rinv33           = avx128fma_invsqrt_f(rsq33);
1051
1052             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1053             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1054             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1055             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1056             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1057             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1058             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1059             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1060             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1061
1062             fjx1             = _mm_setzero_ps();
1063             fjy1             = _mm_setzero_ps();
1064             fjz1             = _mm_setzero_ps();
1065             fjx2             = _mm_setzero_ps();
1066             fjy2             = _mm_setzero_ps();
1067             fjz2             = _mm_setzero_ps();
1068             fjx3             = _mm_setzero_ps();
1069             fjy3             = _mm_setzero_ps();
1070             fjz3             = _mm_setzero_ps();
1071
1072             /**************************
1073              * CALCULATE INTERACTIONS *
1074              **************************/
1075
1076             /* REACTION-FIELD ELECTROSTATICS */
1077             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1078
1079             fscal            = felec;
1080
1081              /* Update vectorial force */
1082             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1083             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1084             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1085
1086             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1087             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1088             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1089
1090             /**************************
1091              * CALCULATE INTERACTIONS *
1092              **************************/
1093
1094             /* REACTION-FIELD ELECTROSTATICS */
1095             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1096
1097             fscal            = felec;
1098
1099              /* Update vectorial force */
1100             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1101             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1102             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1103
1104             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1105             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1106             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1107
1108             /**************************
1109              * CALCULATE INTERACTIONS *
1110              **************************/
1111
1112             /* REACTION-FIELD ELECTROSTATICS */
1113             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1114
1115             fscal            = felec;
1116
1117              /* Update vectorial force */
1118             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1119             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1120             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1121
1122             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1123             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1124             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1125
1126             /**************************
1127              * CALCULATE INTERACTIONS *
1128              **************************/
1129
1130             /* REACTION-FIELD ELECTROSTATICS */
1131             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1132
1133             fscal            = felec;
1134
1135              /* Update vectorial force */
1136             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1137             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1138             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1139
1140             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1141             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1142             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1143
1144             /**************************
1145              * CALCULATE INTERACTIONS *
1146              **************************/
1147
1148             /* REACTION-FIELD ELECTROSTATICS */
1149             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1150
1151             fscal            = felec;
1152
1153              /* Update vectorial force */
1154             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1155             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1156             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1157
1158             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1159             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1160             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1161
1162             /**************************
1163              * CALCULATE INTERACTIONS *
1164              **************************/
1165
1166             /* REACTION-FIELD ELECTROSTATICS */
1167             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1168
1169             fscal            = felec;
1170
1171              /* Update vectorial force */
1172             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1173             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1174             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1175
1176             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1177             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1178             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1179
1180             /**************************
1181              * CALCULATE INTERACTIONS *
1182              **************************/
1183
1184             /* REACTION-FIELD ELECTROSTATICS */
1185             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1186
1187             fscal            = felec;
1188
1189              /* Update vectorial force */
1190             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1191             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1192             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1193
1194             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1195             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1196             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1197
1198             /**************************
1199              * CALCULATE INTERACTIONS *
1200              **************************/
1201
1202             /* REACTION-FIELD ELECTROSTATICS */
1203             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1204
1205             fscal            = felec;
1206
1207              /* Update vectorial force */
1208             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1209             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1210             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1211
1212             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1213             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1214             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1215
1216             /**************************
1217              * CALCULATE INTERACTIONS *
1218              **************************/
1219
1220             /* REACTION-FIELD ELECTROSTATICS */
1221             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1222
1223             fscal            = felec;
1224
1225              /* Update vectorial force */
1226             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1227             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1228             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1229
1230             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1231             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1232             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1233
1234             fjptrA             = f+j_coord_offsetA;
1235             fjptrB             = f+j_coord_offsetB;
1236             fjptrC             = f+j_coord_offsetC;
1237             fjptrD             = f+j_coord_offsetD;
1238
1239             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1240                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1241
1242             /* Inner loop uses 270 flops */
1243         }
1244
1245         if(jidx<j_index_end)
1246         {
1247
1248             /* Get j neighbor index, and coordinate index */
1249             jnrlistA         = jjnr[jidx];
1250             jnrlistB         = jjnr[jidx+1];
1251             jnrlistC         = jjnr[jidx+2];
1252             jnrlistD         = jjnr[jidx+3];
1253             /* Sign of each element will be negative for non-real atoms.
1254              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1255              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1256              */
1257             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1258             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1259             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1260             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1261             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1262             j_coord_offsetA  = DIM*jnrA;
1263             j_coord_offsetB  = DIM*jnrB;
1264             j_coord_offsetC  = DIM*jnrC;
1265             j_coord_offsetD  = DIM*jnrD;
1266
1267             /* load j atom coordinates */
1268             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1269                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1270                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1271
1272             /* Calculate displacement vector */
1273             dx11             = _mm_sub_ps(ix1,jx1);
1274             dy11             = _mm_sub_ps(iy1,jy1);
1275             dz11             = _mm_sub_ps(iz1,jz1);
1276             dx12             = _mm_sub_ps(ix1,jx2);
1277             dy12             = _mm_sub_ps(iy1,jy2);
1278             dz12             = _mm_sub_ps(iz1,jz2);
1279             dx13             = _mm_sub_ps(ix1,jx3);
1280             dy13             = _mm_sub_ps(iy1,jy3);
1281             dz13             = _mm_sub_ps(iz1,jz3);
1282             dx21             = _mm_sub_ps(ix2,jx1);
1283             dy21             = _mm_sub_ps(iy2,jy1);
1284             dz21             = _mm_sub_ps(iz2,jz1);
1285             dx22             = _mm_sub_ps(ix2,jx2);
1286             dy22             = _mm_sub_ps(iy2,jy2);
1287             dz22             = _mm_sub_ps(iz2,jz2);
1288             dx23             = _mm_sub_ps(ix2,jx3);
1289             dy23             = _mm_sub_ps(iy2,jy3);
1290             dz23             = _mm_sub_ps(iz2,jz3);
1291             dx31             = _mm_sub_ps(ix3,jx1);
1292             dy31             = _mm_sub_ps(iy3,jy1);
1293             dz31             = _mm_sub_ps(iz3,jz1);
1294             dx32             = _mm_sub_ps(ix3,jx2);
1295             dy32             = _mm_sub_ps(iy3,jy2);
1296             dz32             = _mm_sub_ps(iz3,jz2);
1297             dx33             = _mm_sub_ps(ix3,jx3);
1298             dy33             = _mm_sub_ps(iy3,jy3);
1299             dz33             = _mm_sub_ps(iz3,jz3);
1300
1301             /* Calculate squared distance and things based on it */
1302             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1303             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1304             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1305             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1306             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1307             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1308             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1309             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1310             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1311
1312             rinv11           = avx128fma_invsqrt_f(rsq11);
1313             rinv12           = avx128fma_invsqrt_f(rsq12);
1314             rinv13           = avx128fma_invsqrt_f(rsq13);
1315             rinv21           = avx128fma_invsqrt_f(rsq21);
1316             rinv22           = avx128fma_invsqrt_f(rsq22);
1317             rinv23           = avx128fma_invsqrt_f(rsq23);
1318             rinv31           = avx128fma_invsqrt_f(rsq31);
1319             rinv32           = avx128fma_invsqrt_f(rsq32);
1320             rinv33           = avx128fma_invsqrt_f(rsq33);
1321
1322             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1323             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1324             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1325             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1326             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1327             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1328             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1329             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1330             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1331
1332             fjx1             = _mm_setzero_ps();
1333             fjy1             = _mm_setzero_ps();
1334             fjz1             = _mm_setzero_ps();
1335             fjx2             = _mm_setzero_ps();
1336             fjy2             = _mm_setzero_ps();
1337             fjz2             = _mm_setzero_ps();
1338             fjx3             = _mm_setzero_ps();
1339             fjy3             = _mm_setzero_ps();
1340             fjz3             = _mm_setzero_ps();
1341
1342             /**************************
1343              * CALCULATE INTERACTIONS *
1344              **************************/
1345
1346             /* REACTION-FIELD ELECTROSTATICS */
1347             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1348
1349             fscal            = felec;
1350
1351             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1352
1353              /* Update vectorial force */
1354             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1355             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1356             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1357
1358             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1359             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1360             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1361
1362             /**************************
1363              * CALCULATE INTERACTIONS *
1364              **************************/
1365
1366             /* REACTION-FIELD ELECTROSTATICS */
1367             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1368
1369             fscal            = felec;
1370
1371             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1372
1373              /* Update vectorial force */
1374             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1375             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1376             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1377
1378             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1379             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1380             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1381
1382             /**************************
1383              * CALCULATE INTERACTIONS *
1384              **************************/
1385
1386             /* REACTION-FIELD ELECTROSTATICS */
1387             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1388
1389             fscal            = felec;
1390
1391             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1392
1393              /* Update vectorial force */
1394             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1395             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1396             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1397
1398             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1399             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1400             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1401
1402             /**************************
1403              * CALCULATE INTERACTIONS *
1404              **************************/
1405
1406             /* REACTION-FIELD ELECTROSTATICS */
1407             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1408
1409             fscal            = felec;
1410
1411             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1412
1413              /* Update vectorial force */
1414             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1415             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1416             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1417
1418             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1419             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1420             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             /* REACTION-FIELD ELECTROSTATICS */
1427             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1428
1429             fscal            = felec;
1430
1431             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1432
1433              /* Update vectorial force */
1434             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1435             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1436             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1437
1438             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1439             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1440             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1441
1442             /**************************
1443              * CALCULATE INTERACTIONS *
1444              **************************/
1445
1446             /* REACTION-FIELD ELECTROSTATICS */
1447             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1448
1449             fscal            = felec;
1450
1451             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1452
1453              /* Update vectorial force */
1454             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1455             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1456             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1457
1458             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1459             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1460             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             /* REACTION-FIELD ELECTROSTATICS */
1467             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1468
1469             fscal            = felec;
1470
1471             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1472
1473              /* Update vectorial force */
1474             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1475             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1476             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1477
1478             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1479             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1480             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1481
1482             /**************************
1483              * CALCULATE INTERACTIONS *
1484              **************************/
1485
1486             /* REACTION-FIELD ELECTROSTATICS */
1487             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1488
1489             fscal            = felec;
1490
1491             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1492
1493              /* Update vectorial force */
1494             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1495             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1496             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1497
1498             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1499             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1500             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1501
1502             /**************************
1503              * CALCULATE INTERACTIONS *
1504              **************************/
1505
1506             /* REACTION-FIELD ELECTROSTATICS */
1507             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1508
1509             fscal            = felec;
1510
1511             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1512
1513              /* Update vectorial force */
1514             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1515             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1516             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1517
1518             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1519             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1520             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1521
1522             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1523             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1524             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1525             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1526
1527             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1528                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1529
1530             /* Inner loop uses 270 flops */
1531         }
1532
1533         /* End of innermost loop */
1534
1535         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1536                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1537
1538         /* Increment number of inner iterations */
1539         inneriter                  += j_index_end - j_index_start;
1540
1541         /* Outer loop uses 18 flops */
1542     }
1543
1544     /* Increment number of outer iterations */
1545     outeriter        += nri;
1546
1547     /* Update outer/inner flops */
1548
1549     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1550 }