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

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset1;
  70     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  71     int              vdwioffset2;
  72     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  73     int              vdwioffset3;
  74     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  75     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  76     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  77     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  78     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  79     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  80     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  81     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  82     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  83     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  84     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  85     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  86     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  87     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  88     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  89     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128           dummy_mask,cutoff_mask;
  93     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  94     __m128           one     = _mm_set1_ps(1.0);
  95     __m128           two     = _mm_set1_ps(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm_set1_ps(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109     krf              = _mm_set1_ps(fr->ic->k_rf);
 110     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 111     crf              = _mm_set1_ps(fr->ic->c_rf);
 112
 113     /* Setup water-specific parameters */
 114     inr              = nlist->iinr[0];
 115     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 116     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 117     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 118
 119     jq1              = _mm_set1_ps(charge[inr+1]);
 120     jq2              = _mm_set1_ps(charge[inr+2]);
 121     jq3              = _mm_set1_ps(charge[inr+3]);
 122     qq11             = _mm_mul_ps(iq1,jq1);
 123     qq12             = _mm_mul_ps(iq1,jq2);
 124     qq13             = _mm_mul_ps(iq1,jq3);
 125     qq21             = _mm_mul_ps(iq2,jq1);
 126     qq22             = _mm_mul_ps(iq2,jq2);
 127     qq23             = _mm_mul_ps(iq2,jq3);
 128     qq31             = _mm_mul_ps(iq3,jq1);
 129     qq32             = _mm_mul_ps(iq3,jq2);
 130     qq33             = _mm_mul_ps(iq3,jq3);
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = jnrC = jnrD = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136     j_coord_offsetC = 0;
 137     j_coord_offsetD = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 163                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 164
 165         fix1             = _mm_setzero_ps();
 166         fiy1             = _mm_setzero_ps();
 167         fiz1             = _mm_setzero_ps();
 168         fix2             = _mm_setzero_ps();
 169         fiy2             = _mm_setzero_ps();
 170         fiz2             = _mm_setzero_ps();
 171         fix3             = _mm_setzero_ps();
 172         fiy3             = _mm_setzero_ps();
 173         fiz3             = _mm_setzero_ps();
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_ps();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189             j_coord_offsetC  = DIM*jnrC;
 190             j_coord_offsetD  = DIM*jnrD;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 194                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 195                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 196
 197             /* Calculate displacement vector */
 198             dx11             = _mm_sub_ps(ix1,jx1);
 199             dy11             = _mm_sub_ps(iy1,jy1);
 200             dz11             = _mm_sub_ps(iz1,jz1);
 201             dx12             = _mm_sub_ps(ix1,jx2);
 202             dy12             = _mm_sub_ps(iy1,jy2);
 203             dz12             = _mm_sub_ps(iz1,jz2);
 204             dx13             = _mm_sub_ps(ix1,jx3);
 205             dy13             = _mm_sub_ps(iy1,jy3);
 206             dz13             = _mm_sub_ps(iz1,jz3);
 207             dx21             = _mm_sub_ps(ix2,jx1);
 208             dy21             = _mm_sub_ps(iy2,jy1);
 209             dz21             = _mm_sub_ps(iz2,jz1);
 210             dx22             = _mm_sub_ps(ix2,jx2);
 211             dy22             = _mm_sub_ps(iy2,jy2);
 212             dz22             = _mm_sub_ps(iz2,jz2);
 213             dx23             = _mm_sub_ps(ix2,jx3);
 214             dy23             = _mm_sub_ps(iy2,jy3);
 215             dz23             = _mm_sub_ps(iz2,jz3);
 216             dx31             = _mm_sub_ps(ix3,jx1);
 217             dy31             = _mm_sub_ps(iy3,jy1);
 218             dz31             = _mm_sub_ps(iz3,jz1);
 219             dx32             = _mm_sub_ps(ix3,jx2);
 220             dy32             = _mm_sub_ps(iy3,jy2);
 221             dz32             = _mm_sub_ps(iz3,jz2);
 222             dx33             = _mm_sub_ps(ix3,jx3);
 223             dy33             = _mm_sub_ps(iy3,jy3);
 224             dz33             = _mm_sub_ps(iz3,jz3);
 225
 226             /* Calculate squared distance and things based on it */
 227             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 228             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 229             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 230             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 231             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 232             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 233             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 234             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 235             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 236
 237             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 238             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 239             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 240             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 241             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 242             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 243             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 244             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 245             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 246
 247             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 248             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 249             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 250             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 251             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 252             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 253             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 254             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 255             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 256
 257             fjx1             = _mm_setzero_ps();
 258             fjy1             = _mm_setzero_ps();
 259             fjz1             = _mm_setzero_ps();
 260             fjx2             = _mm_setzero_ps();
 261             fjy2             = _mm_setzero_ps();
 262             fjz2             = _mm_setzero_ps();
 263             fjx3             = _mm_setzero_ps();
 264             fjy3             = _mm_setzero_ps();
 265             fjz3             = _mm_setzero_ps();
 266
 267             /**************************
 268              * CALCULATE INTERACTIONS *
 269              **************************/
 270
 271             /* REACTION-FIELD ELECTROSTATICS */
 272             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 273             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 274
 275             /* Update potential sum for this i atom from the interaction with this j atom. */
 276             velecsum         = _mm_add_ps(velecsum,velec);
 277
 278             fscal            = felec;
 279
 280              /* Update vectorial force */
 281             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 282             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 283             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 284
 285             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 286             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 287             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 288
 289             /**************************
 290              * CALCULATE INTERACTIONS *
 291              **************************/
 292
 293             /* REACTION-FIELD ELECTROSTATICS */
 294             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 295             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _mm_add_ps(velecsum,velec);
 299
 300             fscal            = felec;
 301
 302              /* Update vectorial force */
 303             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 304             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 305             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 306
 307             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 308             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 309             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* REACTION-FIELD ELECTROSTATICS */
 316             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 317             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velecsum         = _mm_add_ps(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324              /* Update vectorial force */
 325             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 326             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 327             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 328
 329             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 330             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 331             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             /* REACTION-FIELD ELECTROSTATICS */
 338             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 339             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 340
 341             /* Update potential sum for this i atom from the interaction with this j atom. */
 342             velecsum         = _mm_add_ps(velecsum,velec);
 343
 344             fscal            = felec;
 345
 346              /* Update vectorial force */
 347             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 348             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 349             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 350
 351             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 352             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 353             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             /* REACTION-FIELD ELECTROSTATICS */
 360             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 361             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368              /* Update vectorial force */
 369             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 370             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 371             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 372
 373             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 374             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 375             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 376
 377             /**************************
 378              * CALCULATE INTERACTIONS *
 379              **************************/
 380
 381             /* REACTION-FIELD ELECTROSTATICS */
 382             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 383             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velecsum         = _mm_add_ps(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390              /* Update vectorial force */
 391             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 392             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 393             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 394
 395             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 396             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 397             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 405             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velecsum         = _mm_add_ps(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412              /* Update vectorial force */
 413             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 414             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 415             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 416
 417             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 418             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 419             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* REACTION-FIELD ELECTROSTATICS */
 426             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 427             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velecsum         = _mm_add_ps(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434              /* Update vectorial force */
 435             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 436             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 437             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 438
 439             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 440             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 441             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             /* REACTION-FIELD ELECTROSTATICS */
 448             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 449             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 450
 451             /* Update potential sum for this i atom from the interaction with this j atom. */
 452             velecsum         = _mm_add_ps(velecsum,velec);
 453
 454             fscal            = felec;
 455
 456              /* Update vectorial force */
 457             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 458             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 459             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 460
 461             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 462             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 463             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 464
 465             fjptrA             = f+j_coord_offsetA;
 466             fjptrB             = f+j_coord_offsetB;
 467             fjptrC             = f+j_coord_offsetC;
 468             fjptrD             = f+j_coord_offsetD;
 469
 470             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 471                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 472
 473             /* Inner loop uses 315 flops */
 474         }
 475
 476         if(jidx<j_index_end)
 477         {
 478
 479             /* Get j neighbor index, and coordinate index */
 480             jnrlistA         = jjnr[jidx];
 481             jnrlistB         = jjnr[jidx+1];
 482             jnrlistC         = jjnr[jidx+2];
 483             jnrlistD         = jjnr[jidx+3];
 484             /* Sign of each element will be negative for non-real atoms.
 485              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 486              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 487              */
 488             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 489             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 490             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 491             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 492             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 493             j_coord_offsetA  = DIM*jnrA;
 494             j_coord_offsetB  = DIM*jnrB;
 495             j_coord_offsetC  = DIM*jnrC;
 496             j_coord_offsetD  = DIM*jnrD;
 497
 498             /* load j atom coordinates */
 499             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 500                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 501                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 502
 503             /* Calculate displacement vector */
 504             dx11             = _mm_sub_ps(ix1,jx1);
 505             dy11             = _mm_sub_ps(iy1,jy1);
 506             dz11             = _mm_sub_ps(iz1,jz1);
 507             dx12             = _mm_sub_ps(ix1,jx2);
 508             dy12             = _mm_sub_ps(iy1,jy2);
 509             dz12             = _mm_sub_ps(iz1,jz2);
 510             dx13             = _mm_sub_ps(ix1,jx3);
 511             dy13             = _mm_sub_ps(iy1,jy3);
 512             dz13             = _mm_sub_ps(iz1,jz3);
 513             dx21             = _mm_sub_ps(ix2,jx1);
 514             dy21             = _mm_sub_ps(iy2,jy1);
 515             dz21             = _mm_sub_ps(iz2,jz1);
 516             dx22             = _mm_sub_ps(ix2,jx2);
 517             dy22             = _mm_sub_ps(iy2,jy2);
 518             dz22             = _mm_sub_ps(iz2,jz2);
 519             dx23             = _mm_sub_ps(ix2,jx3);
 520             dy23             = _mm_sub_ps(iy2,jy3);
 521             dz23             = _mm_sub_ps(iz2,jz3);
 522             dx31             = _mm_sub_ps(ix3,jx1);
 523             dy31             = _mm_sub_ps(iy3,jy1);
 524             dz31             = _mm_sub_ps(iz3,jz1);
 525             dx32             = _mm_sub_ps(ix3,jx2);
 526             dy32             = _mm_sub_ps(iy3,jy2);
 527             dz32             = _mm_sub_ps(iz3,jz2);
 528             dx33             = _mm_sub_ps(ix3,jx3);
 529             dy33             = _mm_sub_ps(iy3,jy3);
 530             dz33             = _mm_sub_ps(iz3,jz3);
 531
 532             /* Calculate squared distance and things based on it */
 533             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 534             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 535             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 536             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 537             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 538             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 539             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 540             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 541             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 542
 543             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 544             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 545             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 546             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 547             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 548             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 549             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 550             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 551             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 552
 553             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 554             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 555             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 556             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 557             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 558             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 559             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 560             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 561             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 562
 563             fjx1             = _mm_setzero_ps();
 564             fjy1             = _mm_setzero_ps();
 565             fjz1             = _mm_setzero_ps();
 566             fjx2             = _mm_setzero_ps();
 567             fjy2             = _mm_setzero_ps();
 568             fjz2             = _mm_setzero_ps();
 569             fjx3             = _mm_setzero_ps();
 570             fjy3             = _mm_setzero_ps();
 571             fjz3             = _mm_setzero_ps();
 572
 573             /**************************
 574              * CALCULATE INTERACTIONS *
 575              **************************/
 576
 577             /* REACTION-FIELD ELECTROSTATICS */
 578             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 579             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 580
 581             /* Update potential sum for this i atom from the interaction with this j atom. */
 582             velec            = _mm_andnot_ps(dummy_mask,velec);
 583             velecsum         = _mm_add_ps(velecsum,velec);
 584
 585             fscal            = felec;
 586
 587             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 588
 589              /* Update vectorial force */
 590             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 591             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 592             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 593
 594             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 595             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 596             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             /* REACTION-FIELD ELECTROSTATICS */
 603             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 604             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 605
 606             /* Update potential sum for this i atom from the interaction with this j atom. */
 607             velec            = _mm_andnot_ps(dummy_mask,velec);
 608             velecsum         = _mm_add_ps(velecsum,velec);
 609
 610             fscal            = felec;
 611
 612             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 613
 614              /* Update vectorial force */
 615             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 616             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 617             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 618
 619             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 620             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 621             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             /* REACTION-FIELD ELECTROSTATICS */
 628             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 629             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 630
 631             /* Update potential sum for this i atom from the interaction with this j atom. */
 632             velec            = _mm_andnot_ps(dummy_mask,velec);
 633             velecsum         = _mm_add_ps(velecsum,velec);
 634
 635             fscal            = felec;
 636
 637             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 638
 639              /* Update vectorial force */
 640             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 641             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 642             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 643
 644             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 645             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 646             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 647
 648             /**************************
 649              * CALCULATE INTERACTIONS *
 650              **************************/
 651
 652             /* REACTION-FIELD ELECTROSTATICS */
 653             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 654             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 655
 656             /* Update potential sum for this i atom from the interaction with this j atom. */
 657             velec            = _mm_andnot_ps(dummy_mask,velec);
 658             velecsum         = _mm_add_ps(velecsum,velec);
 659
 660             fscal            = felec;
 661
 662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 663
 664              /* Update vectorial force */
 665             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 666             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 667             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 668
 669             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 670             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 671             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 672
 673             /**************************
 674              * CALCULATE INTERACTIONS *
 675              **************************/
 676
 677             /* REACTION-FIELD ELECTROSTATICS */
 678             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 679             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 680
 681             /* Update potential sum for this i atom from the interaction with this j atom. */
 682             velec            = _mm_andnot_ps(dummy_mask,velec);
 683             velecsum         = _mm_add_ps(velecsum,velec);
 684
 685             fscal            = felec;
 686
 687             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 688
 689              /* Update vectorial force */
 690             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 691             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 692             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 693
 694             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 695             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 696             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             /* REACTION-FIELD ELECTROSTATICS */
 703             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 704             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 705
 706             /* Update potential sum for this i atom from the interaction with this j atom. */
 707             velec            = _mm_andnot_ps(dummy_mask,velec);
 708             velecsum         = _mm_add_ps(velecsum,velec);
 709
 710             fscal            = felec;
 711
 712             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 713
 714              /* Update vectorial force */
 715             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 716             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 717             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 718
 719             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 720             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 721             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             /* REACTION-FIELD ELECTROSTATICS */
 728             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 729             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 730
 731             /* Update potential sum for this i atom from the interaction with this j atom. */
 732             velec            = _mm_andnot_ps(dummy_mask,velec);
 733             velecsum         = _mm_add_ps(velecsum,velec);
 734
 735             fscal            = felec;
 736
 737             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 738
 739              /* Update vectorial force */
 740             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 741             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 742             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 743
 744             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 745             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 746             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             /* REACTION-FIELD ELECTROSTATICS */
 753             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 754             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 755
 756             /* Update potential sum for this i atom from the interaction with this j atom. */
 757             velec            = _mm_andnot_ps(dummy_mask,velec);
 758             velecsum         = _mm_add_ps(velecsum,velec);
 759
 760             fscal            = felec;
 761
 762             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 763
 764              /* Update vectorial force */
 765             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 766             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 767             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 768
 769             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 770             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 771             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             /* REACTION-FIELD ELECTROSTATICS */
 778             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 779             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 780
 781             /* Update potential sum for this i atom from the interaction with this j atom. */
 782             velec            = _mm_andnot_ps(dummy_mask,velec);
 783             velecsum         = _mm_add_ps(velecsum,velec);
 784
 785             fscal            = felec;
 786
 787             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 788
 789              /* Update vectorial force */
 790             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 791             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 792             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 793
 794             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 795             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 796             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 797
 798             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 799             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 800             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 801             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 802
 803             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 804                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 805
 806             /* Inner loop uses 315 flops */
 807         }
 808
 809         /* End of innermost loop */
 810
 811         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 812                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 813
 814         ggid                        = gid[iidx];
 815         /* Update potential energies */
 816         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 817
 818         /* Increment number of inner iterations */
 819         inneriter                  += j_index_end - j_index_start;
 820
 821         /* Outer loop uses 19 flops */
 822     }
 823
 824     /* Increment number of outer iterations */
 825     outeriter        += nri;
 826
 827     /* Update outer/inner flops */
 828
 829     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*315);
 830 }
 831 /*
 832  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_avx_128_fma_single
 833  * Electrostatics interaction: ReactionField
 834  * VdW interaction:            None
 835  * Geometry:                   Water4-Water4
 836  * Calculate force/pot:        Force
 837  */
 838 void
 839 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_avx_128_fma_single
 840                     (t_nblist * gmx_restrict                nlist,
 841                      rvec * gmx_restrict                    xx,
 842                      rvec * gmx_restrict                    ff,
 843                      t_forcerec * gmx_restrict              fr,
 844                      t_mdatoms * gmx_restrict               mdatoms,
 845                      nb_kernel_data_t * gmx_restrict        kernel_data,
 846                      t_nrnb * gmx_restrict                  nrnb)
 847 {
 848     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 849      * just 0 for non-waters.
 850      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 851      * jnr indices corresponding to data put in the four positions in the SIMD register.
 852      */
 853     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 854     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 855     int              jnrA,jnrB,jnrC,jnrD;
 856     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 857     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 858     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 859     real             rcutoff_scalar;
 860     real             *shiftvec,*fshift,*x,*f;
 861     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 862     real             scratch[4*DIM];
 863     __m128           fscal,rcutoff,rcutoff2,jidxall;
 864     int              vdwioffset1;
 865     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 866     int              vdwioffset2;
 867     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 868     int              vdwioffset3;
 869     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 870     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 871     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 872     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 873     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 874     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 875     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 876     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 877     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 878     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 879     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 880     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 881     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 882     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 883     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 884     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 885     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 886     real             *charge;
 887     __m128           dummy_mask,cutoff_mask;
 888     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 889     __m128           one     = _mm_set1_ps(1.0);
 890     __m128           two     = _mm_set1_ps(2.0);
 891     x                = xx[0];
 892     f                = ff[0];
 893
 894     nri              = nlist->nri;
 895     iinr             = nlist->iinr;
 896     jindex           = nlist->jindex;
 897     jjnr             = nlist->jjnr;
 898     shiftidx         = nlist->shift;
 899     gid              = nlist->gid;
 900     shiftvec         = fr->shift_vec[0];
 901     fshift           = fr->fshift[0];
 902     facel            = _mm_set1_ps(fr->epsfac);
 903     charge           = mdatoms->chargeA;
 904     krf              = _mm_set1_ps(fr->ic->k_rf);
 905     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 906     crf              = _mm_set1_ps(fr->ic->c_rf);
 907
 908     /* Setup water-specific parameters */
 909     inr              = nlist->iinr[0];
 910     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 911     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 912     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 913
 914     jq1              = _mm_set1_ps(charge[inr+1]);
 915     jq2              = _mm_set1_ps(charge[inr+2]);
 916     jq3              = _mm_set1_ps(charge[inr+3]);
 917     qq11             = _mm_mul_ps(iq1,jq1);
 918     qq12             = _mm_mul_ps(iq1,jq2);
 919     qq13             = _mm_mul_ps(iq1,jq3);
 920     qq21             = _mm_mul_ps(iq2,jq1);
 921     qq22             = _mm_mul_ps(iq2,jq2);
 922     qq23             = _mm_mul_ps(iq2,jq3);
 923     qq31             = _mm_mul_ps(iq3,jq1);
 924     qq32             = _mm_mul_ps(iq3,jq2);
 925     qq33             = _mm_mul_ps(iq3,jq3);
 926
 927     /* Avoid stupid compiler warnings */
 928     jnrA = jnrB = jnrC = jnrD = 0;
 929     j_coord_offsetA = 0;
 930     j_coord_offsetB = 0;
 931     j_coord_offsetC = 0;
 932     j_coord_offsetD = 0;
 933
 934     outeriter        = 0;
 935     inneriter        = 0;
 936
 937     for(iidx=0;iidx<4*DIM;iidx++)
 938     {
 939         scratch[iidx] = 0.0;
 940     }
 941
 942     /* Start outer loop over neighborlists */
 943     for(iidx=0; iidx<nri; iidx++)
 944     {
 945         /* Load shift vector for this list */
 946         i_shift_offset   = DIM*shiftidx[iidx];
 947
 948         /* Load limits for loop over neighbors */
 949         j_index_start    = jindex[iidx];
 950         j_index_end      = jindex[iidx+1];
 951
 952         /* Get outer coordinate index */
 953         inr              = iinr[iidx];
 954         i_coord_offset   = DIM*inr;
 955
 956         /* Load i particle coords and add shift vector */
 957         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 958                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 959
 960         fix1             = _mm_setzero_ps();
 961         fiy1             = _mm_setzero_ps();
 962         fiz1             = _mm_setzero_ps();
 963         fix2             = _mm_setzero_ps();
 964         fiy2             = _mm_setzero_ps();
 965         fiz2             = _mm_setzero_ps();
 966         fix3             = _mm_setzero_ps();
 967         fiy3             = _mm_setzero_ps();
 968         fiz3             = _mm_setzero_ps();
 969
 970         /* Start inner kernel loop */
 971         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 972         {
 973
 974             /* Get j neighbor index, and coordinate index */
 975             jnrA             = jjnr[jidx];
 976             jnrB             = jjnr[jidx+1];
 977             jnrC             = jjnr[jidx+2];
 978             jnrD             = jjnr[jidx+3];
 979             j_coord_offsetA  = DIM*jnrA;
 980             j_coord_offsetB  = DIM*jnrB;
 981             j_coord_offsetC  = DIM*jnrC;
 982             j_coord_offsetD  = DIM*jnrD;
 983
 984             /* load j atom coordinates */
 985             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 986                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 987                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 988
 989             /* Calculate displacement vector */
 990             dx11             = _mm_sub_ps(ix1,jx1);
 991             dy11             = _mm_sub_ps(iy1,jy1);
 992             dz11             = _mm_sub_ps(iz1,jz1);
 993             dx12             = _mm_sub_ps(ix1,jx2);
 994             dy12             = _mm_sub_ps(iy1,jy2);
 995             dz12             = _mm_sub_ps(iz1,jz2);
 996             dx13             = _mm_sub_ps(ix1,jx3);
 997             dy13             = _mm_sub_ps(iy1,jy3);
 998             dz13             = _mm_sub_ps(iz1,jz3);
 999             dx21             = _mm_sub_ps(ix2,jx1);
1000             dy21             = _mm_sub_ps(iy2,jy1);
1001             dz21             = _mm_sub_ps(iz2,jz1);
1002             dx22             = _mm_sub_ps(ix2,jx2);
1003             dy22             = _mm_sub_ps(iy2,jy2);
1004             dz22             = _mm_sub_ps(iz2,jz2);
1005             dx23             = _mm_sub_ps(ix2,jx3);
1006             dy23             = _mm_sub_ps(iy2,jy3);
1007             dz23             = _mm_sub_ps(iz2,jz3);
1008             dx31             = _mm_sub_ps(ix3,jx1);
1009             dy31             = _mm_sub_ps(iy3,jy1);
1010             dz31             = _mm_sub_ps(iz3,jz1);
1011             dx32             = _mm_sub_ps(ix3,jx2);
1012             dy32             = _mm_sub_ps(iy3,jy2);
1013             dz32             = _mm_sub_ps(iz3,jz2);
1014             dx33             = _mm_sub_ps(ix3,jx3);
1015             dy33             = _mm_sub_ps(iy3,jy3);
1016             dz33             = _mm_sub_ps(iz3,jz3);
1017
1018             /* Calculate squared distance and things based on it */
1019             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1020             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1021             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1022             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1023             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1024             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1025             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1026             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1027             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1028
1029             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1030             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1031             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1032             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1033             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1034             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1035             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1036             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1037             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1038
1039             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1040             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1041             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1042             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1043             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1044             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1045             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1046             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1047             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1048
1049             fjx1             = _mm_setzero_ps();
1050             fjy1             = _mm_setzero_ps();
1051             fjz1             = _mm_setzero_ps();
1052             fjx2             = _mm_setzero_ps();
1053             fjy2             = _mm_setzero_ps();
1054             fjz2             = _mm_setzero_ps();
1055             fjx3             = _mm_setzero_ps();
1056             fjy3             = _mm_setzero_ps();
1057             fjz3             = _mm_setzero_ps();
1058
1059             /**************************
1060              * CALCULATE INTERACTIONS *
1061              **************************/
1062
1063             /* REACTION-FIELD ELECTROSTATICS */
1064             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1065
1066             fscal            = felec;
1067
1068              /* Update vectorial force */
1069             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1070             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1071             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1072
1073             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1074             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1075             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1076
1077             /**************************
1078              * CALCULATE INTERACTIONS *
1079              **************************/
1080
1081             /* REACTION-FIELD ELECTROSTATICS */
1082             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1083
1084             fscal            = felec;
1085
1086              /* Update vectorial force */
1087             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1088             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1089             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1090
1091             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1092             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1093             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1094
1095             /**************************
1096              * CALCULATE INTERACTIONS *
1097              **************************/
1098
1099             /* REACTION-FIELD ELECTROSTATICS */
1100             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1101
1102             fscal            = felec;
1103
1104              /* Update vectorial force */
1105             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1106             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1107             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1108
1109             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1110             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1111             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1112
1113             /**************************
1114              * CALCULATE INTERACTIONS *
1115              **************************/
1116
1117             /* REACTION-FIELD ELECTROSTATICS */
1118             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1119
1120             fscal            = felec;
1121
1122              /* Update vectorial force */
1123             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1124             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1125             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1126
1127             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1128             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1129             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1130
1131             /**************************
1132              * CALCULATE INTERACTIONS *
1133              **************************/
1134
1135             /* REACTION-FIELD ELECTROSTATICS */
1136             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1137
1138             fscal            = felec;
1139
1140              /* Update vectorial force */
1141             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1142             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1143             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1144
1145             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1146             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1147             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1148
1149             /**************************
1150              * CALCULATE INTERACTIONS *
1151              **************************/
1152
1153             /* REACTION-FIELD ELECTROSTATICS */
1154             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1155
1156             fscal            = felec;
1157
1158              /* Update vectorial force */
1159             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1160             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1161             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1162
1163             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1164             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1165             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1166
1167             /**************************
1168              * CALCULATE INTERACTIONS *
1169              **************************/
1170
1171             /* REACTION-FIELD ELECTROSTATICS */
1172             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1173
1174             fscal            = felec;
1175
1176              /* Update vectorial force */
1177             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1178             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1179             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1180
1181             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1182             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1183             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1184
1185             /**************************
1186              * CALCULATE INTERACTIONS *
1187              **************************/
1188
1189             /* REACTION-FIELD ELECTROSTATICS */
1190             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1191
1192             fscal            = felec;
1193
1194              /* Update vectorial force */
1195             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1196             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1197             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1198
1199             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1200             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1201             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1202
1203             /**************************
1204              * CALCULATE INTERACTIONS *
1205              **************************/
1206
1207             /* REACTION-FIELD ELECTROSTATICS */
1208             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1209
1210             fscal            = felec;
1211
1212              /* Update vectorial force */
1213             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1214             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1215             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1216
1217             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1218             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1219             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1220
1221             fjptrA             = f+j_coord_offsetA;
1222             fjptrB             = f+j_coord_offsetB;
1223             fjptrC             = f+j_coord_offsetC;
1224             fjptrD             = f+j_coord_offsetD;
1225
1226             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1227                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1228
1229             /* Inner loop uses 270 flops */
1230         }
1231
1232         if(jidx<j_index_end)
1233         {
1234
1235             /* Get j neighbor index, and coordinate index */
1236             jnrlistA         = jjnr[jidx];
1237             jnrlistB         = jjnr[jidx+1];
1238             jnrlistC         = jjnr[jidx+2];
1239             jnrlistD         = jjnr[jidx+3];
1240             /* Sign of each element will be negative for non-real atoms.
1241              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1242              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1243              */
1244             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1245             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1246             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1247             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1248             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1249             j_coord_offsetA  = DIM*jnrA;
1250             j_coord_offsetB  = DIM*jnrB;
1251             j_coord_offsetC  = DIM*jnrC;
1252             j_coord_offsetD  = DIM*jnrD;
1253
1254             /* load j atom coordinates */
1255             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1256                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1257                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1258
1259             /* Calculate displacement vector */
1260             dx11             = _mm_sub_ps(ix1,jx1);
1261             dy11             = _mm_sub_ps(iy1,jy1);
1262             dz11             = _mm_sub_ps(iz1,jz1);
1263             dx12             = _mm_sub_ps(ix1,jx2);
1264             dy12             = _mm_sub_ps(iy1,jy2);
1265             dz12             = _mm_sub_ps(iz1,jz2);
1266             dx13             = _mm_sub_ps(ix1,jx3);
1267             dy13             = _mm_sub_ps(iy1,jy3);
1268             dz13             = _mm_sub_ps(iz1,jz3);
1269             dx21             = _mm_sub_ps(ix2,jx1);
1270             dy21             = _mm_sub_ps(iy2,jy1);
1271             dz21             = _mm_sub_ps(iz2,jz1);
1272             dx22             = _mm_sub_ps(ix2,jx2);
1273             dy22             = _mm_sub_ps(iy2,jy2);
1274             dz22             = _mm_sub_ps(iz2,jz2);
1275             dx23             = _mm_sub_ps(ix2,jx3);
1276             dy23             = _mm_sub_ps(iy2,jy3);
1277             dz23             = _mm_sub_ps(iz2,jz3);
1278             dx31             = _mm_sub_ps(ix3,jx1);
1279             dy31             = _mm_sub_ps(iy3,jy1);
1280             dz31             = _mm_sub_ps(iz3,jz1);
1281             dx32             = _mm_sub_ps(ix3,jx2);
1282             dy32             = _mm_sub_ps(iy3,jy2);
1283             dz32             = _mm_sub_ps(iz3,jz2);
1284             dx33             = _mm_sub_ps(ix3,jx3);
1285             dy33             = _mm_sub_ps(iy3,jy3);
1286             dz33             = _mm_sub_ps(iz3,jz3);
1287
1288             /* Calculate squared distance and things based on it */
1289             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1290             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1291             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1292             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1293             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1294             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1295             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1296             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1297             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1298
1299             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1300             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1301             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1302             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1303             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1304             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1305             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1306             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1307             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1308
1309             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1310             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1311             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1312             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1313             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1314             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1315             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1316             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1317             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1318
1319             fjx1             = _mm_setzero_ps();
1320             fjy1             = _mm_setzero_ps();
1321             fjz1             = _mm_setzero_ps();
1322             fjx2             = _mm_setzero_ps();
1323             fjy2             = _mm_setzero_ps();
1324             fjz2             = _mm_setzero_ps();
1325             fjx3             = _mm_setzero_ps();
1326             fjy3             = _mm_setzero_ps();
1327             fjz3             = _mm_setzero_ps();
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             /* REACTION-FIELD ELECTROSTATICS */
1334             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1335
1336             fscal            = felec;
1337
1338             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1339
1340              /* Update vectorial force */
1341             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1342             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1343             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1344
1345             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1346             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1347             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1348
1349             /**************************
1350              * CALCULATE INTERACTIONS *
1351              **************************/
1352
1353             /* REACTION-FIELD ELECTROSTATICS */
1354             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1355
1356             fscal            = felec;
1357
1358             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1359
1360              /* Update vectorial force */
1361             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1362             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1363             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1364
1365             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1366             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1367             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1368
1369             /**************************
1370              * CALCULATE INTERACTIONS *
1371              **************************/
1372
1373             /* REACTION-FIELD ELECTROSTATICS */
1374             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1375
1376             fscal            = felec;
1377
1378             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1379
1380              /* Update vectorial force */
1381             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1382             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1383             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1384
1385             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1386             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1387             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1388
1389             /**************************
1390              * CALCULATE INTERACTIONS *
1391              **************************/
1392
1393             /* REACTION-FIELD ELECTROSTATICS */
1394             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1395
1396             fscal            = felec;
1397
1398             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1399
1400              /* Update vectorial force */
1401             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1402             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1403             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1404
1405             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1406             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1407             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1408
1409             /**************************
1410              * CALCULATE INTERACTIONS *
1411              **************************/
1412
1413             /* REACTION-FIELD ELECTROSTATICS */
1414             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1415
1416             fscal            = felec;
1417
1418             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1419
1420              /* Update vectorial force */
1421             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1422             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1423             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1424
1425             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1426             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1427             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1428
1429             /**************************
1430              * CALCULATE INTERACTIONS *
1431              **************************/
1432
1433             /* REACTION-FIELD ELECTROSTATICS */
1434             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1435
1436             fscal            = felec;
1437
1438             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1439
1440              /* Update vectorial force */
1441             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1442             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1443             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1444
1445             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1446             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1447             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1448
1449             /**************************
1450              * CALCULATE INTERACTIONS *
1451              **************************/
1452
1453             /* REACTION-FIELD ELECTROSTATICS */
1454             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1455
1456             fscal            = felec;
1457
1458             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1459
1460              /* Update vectorial force */
1461             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1462             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1463             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1464
1465             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1466             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1467             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1468
1469             /**************************
1470              * CALCULATE INTERACTIONS *
1471              **************************/
1472
1473             /* REACTION-FIELD ELECTROSTATICS */
1474             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1475
1476             fscal            = felec;
1477
1478             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1479
1480              /* Update vectorial force */
1481             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1482             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1483             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1484
1485             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1486             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1487             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1488
1489             /**************************
1490              * CALCULATE INTERACTIONS *
1491              **************************/
1492
1493             /* REACTION-FIELD ELECTROSTATICS */
1494             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1495
1496             fscal            = felec;
1497
1498             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1499
1500              /* Update vectorial force */
1501             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1502             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1503             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1504
1505             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1506             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1507             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1508
1509             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1510             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1511             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1512             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1513
1514             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1515                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1516
1517             /* Inner loop uses 270 flops */
1518         }
1519
1520         /* End of innermost loop */
1521
1522         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1523                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1524
1525         /* Increment number of inner iterations */
1526         inneriter                  += j_index_end - j_index_start;
1527
1528         /* Outer loop uses 18 flops */
1529     }
1530
1531     /* Increment number of outer iterations */
1532     outeriter        += nri;
1533
1534     /* Update outer/inner flops */
1535
1536     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1537 }