src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwNone_GeomW4W4_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_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_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse2_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse2_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 SSE, 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              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset1;
  67     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  68     int              vdwioffset2;
  69     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  70     int              vdwioffset3;
  71     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  72     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  73     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  74     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  75     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  76     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  77     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  78     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  79     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  80     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  81     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  82     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  83     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  84     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  85     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  86     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128           dummy_mask,cutoff_mask;
  90     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  91     __m128           one     = _mm_set1_ps(1.0);
  92     __m128           two     = _mm_set1_ps(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_ps(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106
 107     /* Setup water-specific parameters */
 108     inr              = nlist->iinr[0];
 109     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 110     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 111     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 112
 113     jq1              = _mm_set1_ps(charge[inr+1]);
 114     jq2              = _mm_set1_ps(charge[inr+2]);
 115     jq3              = _mm_set1_ps(charge[inr+3]);
 116     qq11             = _mm_mul_ps(iq1,jq1);
 117     qq12             = _mm_mul_ps(iq1,jq2);
 118     qq13             = _mm_mul_ps(iq1,jq3);
 119     qq21             = _mm_mul_ps(iq2,jq1);
 120     qq22             = _mm_mul_ps(iq2,jq2);
 121     qq23             = _mm_mul_ps(iq2,jq3);
 122     qq31             = _mm_mul_ps(iq3,jq1);
 123     qq32             = _mm_mul_ps(iq3,jq2);
 124     qq33             = _mm_mul_ps(iq3,jq3);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = jnrC = jnrD = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130     j_coord_offsetC = 0;
 131     j_coord_offsetD = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141         shX              = shiftvec[i_shift_offset+XX];
 142         shY              = shiftvec[i_shift_offset+YY];
 143         shZ              = shiftvec[i_shift_offset+ZZ];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 155         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 156         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 157         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 158         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 159         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 160         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
 161         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
 162         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
 163
 164         fix1             = _mm_setzero_ps();
 165         fiy1             = _mm_setzero_ps();
 166         fiz1             = _mm_setzero_ps();
 167         fix2             = _mm_setzero_ps();
 168         fiy2             = _mm_setzero_ps();
 169         fiz2             = _mm_setzero_ps();
 170         fix3             = _mm_setzero_ps();
 171         fiy3             = _mm_setzero_ps();
 172         fiz3             = _mm_setzero_ps();
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_ps();
 176
 177         /* Start inner kernel loop */
 178         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 179         {
 180
 181             /* Get j neighbor index, and coordinate index */
 182             jnrA             = jjnr[jidx];
 183             jnrB             = jjnr[jidx+1];
 184             jnrC             = jjnr[jidx+2];
 185             jnrD             = jjnr[jidx+3];
 186
 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             /* COULOMB ELECTROSTATICS */
 272             velec            = _mm_mul_ps(qq11,rinv11);
 273             felec            = _mm_mul_ps(velec,rinvsq11);
 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             /* Calculate temporary vectorial force */
 281             tx               = _mm_mul_ps(fscal,dx11);
 282             ty               = _mm_mul_ps(fscal,dy11);
 283             tz               = _mm_mul_ps(fscal,dz11);
 284
 285             /* Update vectorial force */
 286             fix1             = _mm_add_ps(fix1,tx);
 287             fiy1             = _mm_add_ps(fiy1,ty);
 288             fiz1             = _mm_add_ps(fiz1,tz);
 289
 290             fjx1             = _mm_add_ps(fjx1,tx);
 291             fjy1             = _mm_add_ps(fjy1,ty);
 292             fjz1             = _mm_add_ps(fjz1,tz);
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             /* COULOMB ELECTROSTATICS */
 299             velec            = _mm_mul_ps(qq12,rinv12);
 300             felec            = _mm_mul_ps(velec,rinvsq12);
 301
 302             /* Update potential sum for this i atom from the interaction with this j atom. */
 303             velecsum         = _mm_add_ps(velecsum,velec);
 304
 305             fscal            = felec;
 306
 307             /* Calculate temporary vectorial force */
 308             tx               = _mm_mul_ps(fscal,dx12);
 309             ty               = _mm_mul_ps(fscal,dy12);
 310             tz               = _mm_mul_ps(fscal,dz12);
 311
 312             /* Update vectorial force */
 313             fix1             = _mm_add_ps(fix1,tx);
 314             fiy1             = _mm_add_ps(fiy1,ty);
 315             fiz1             = _mm_add_ps(fiz1,tz);
 316
 317             fjx2             = _mm_add_ps(fjx2,tx);
 318             fjy2             = _mm_add_ps(fjy2,ty);
 319             fjz2             = _mm_add_ps(fjz2,tz);
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             /* COULOMB ELECTROSTATICS */
 326             velec            = _mm_mul_ps(qq13,rinv13);
 327             felec            = _mm_mul_ps(velec,rinvsq13);
 328
 329             /* Update potential sum for this i atom from the interaction with this j atom. */
 330             velecsum         = _mm_add_ps(velecsum,velec);
 331
 332             fscal            = felec;
 333
 334             /* Calculate temporary vectorial force */
 335             tx               = _mm_mul_ps(fscal,dx13);
 336             ty               = _mm_mul_ps(fscal,dy13);
 337             tz               = _mm_mul_ps(fscal,dz13);
 338
 339             /* Update vectorial force */
 340             fix1             = _mm_add_ps(fix1,tx);
 341             fiy1             = _mm_add_ps(fiy1,ty);
 342             fiz1             = _mm_add_ps(fiz1,tz);
 343
 344             fjx3             = _mm_add_ps(fjx3,tx);
 345             fjy3             = _mm_add_ps(fjy3,ty);
 346             fjz3             = _mm_add_ps(fjz3,tz);
 347
 348             /**************************
 349              * CALCULATE INTERACTIONS *
 350              **************************/
 351
 352             /* COULOMB ELECTROSTATICS */
 353             velec            = _mm_mul_ps(qq21,rinv21);
 354             felec            = _mm_mul_ps(velec,rinvsq21);
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velecsum         = _mm_add_ps(velecsum,velec);
 358
 359             fscal            = felec;
 360
 361             /* Calculate temporary vectorial force */
 362             tx               = _mm_mul_ps(fscal,dx21);
 363             ty               = _mm_mul_ps(fscal,dy21);
 364             tz               = _mm_mul_ps(fscal,dz21);
 365
 366             /* Update vectorial force */
 367             fix2             = _mm_add_ps(fix2,tx);
 368             fiy2             = _mm_add_ps(fiy2,ty);
 369             fiz2             = _mm_add_ps(fiz2,tz);
 370
 371             fjx1             = _mm_add_ps(fjx1,tx);
 372             fjy1             = _mm_add_ps(fjy1,ty);
 373             fjz1             = _mm_add_ps(fjz1,tz);
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             /* COULOMB ELECTROSTATICS */
 380             velec            = _mm_mul_ps(qq22,rinv22);
 381             felec            = _mm_mul_ps(velec,rinvsq22);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velecsum         = _mm_add_ps(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             /* Calculate temporary vectorial force */
 389             tx               = _mm_mul_ps(fscal,dx22);
 390             ty               = _mm_mul_ps(fscal,dy22);
 391             tz               = _mm_mul_ps(fscal,dz22);
 392
 393             /* Update vectorial force */
 394             fix2             = _mm_add_ps(fix2,tx);
 395             fiy2             = _mm_add_ps(fiy2,ty);
 396             fiz2             = _mm_add_ps(fiz2,tz);
 397
 398             fjx2             = _mm_add_ps(fjx2,tx);
 399             fjy2             = _mm_add_ps(fjy2,ty);
 400             fjz2             = _mm_add_ps(fjz2,tz);
 401
 402             /**************************
 403              * CALCULATE INTERACTIONS *
 404              **************************/
 405
 406             /* COULOMB ELECTROSTATICS */
 407             velec            = _mm_mul_ps(qq23,rinv23);
 408             felec            = _mm_mul_ps(velec,rinvsq23);
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velecsum         = _mm_add_ps(velecsum,velec);
 412
 413             fscal            = felec;
 414
 415             /* Calculate temporary vectorial force */
 416             tx               = _mm_mul_ps(fscal,dx23);
 417             ty               = _mm_mul_ps(fscal,dy23);
 418             tz               = _mm_mul_ps(fscal,dz23);
 419
 420             /* Update vectorial force */
 421             fix2             = _mm_add_ps(fix2,tx);
 422             fiy2             = _mm_add_ps(fiy2,ty);
 423             fiz2             = _mm_add_ps(fiz2,tz);
 424
 425             fjx3             = _mm_add_ps(fjx3,tx);
 426             fjy3             = _mm_add_ps(fjy3,ty);
 427             fjz3             = _mm_add_ps(fjz3,tz);
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             /* COULOMB ELECTROSTATICS */
 434             velec            = _mm_mul_ps(qq31,rinv31);
 435             felec            = _mm_mul_ps(velec,rinvsq31);
 436
 437             /* Update potential sum for this i atom from the interaction with this j atom. */
 438             velecsum         = _mm_add_ps(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             /* Calculate temporary vectorial force */
 443             tx               = _mm_mul_ps(fscal,dx31);
 444             ty               = _mm_mul_ps(fscal,dy31);
 445             tz               = _mm_mul_ps(fscal,dz31);
 446
 447             /* Update vectorial force */
 448             fix3             = _mm_add_ps(fix3,tx);
 449             fiy3             = _mm_add_ps(fiy3,ty);
 450             fiz3             = _mm_add_ps(fiz3,tz);
 451
 452             fjx1             = _mm_add_ps(fjx1,tx);
 453             fjy1             = _mm_add_ps(fjy1,ty);
 454             fjz1             = _mm_add_ps(fjz1,tz);
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             /* COULOMB ELECTROSTATICS */
 461             velec            = _mm_mul_ps(qq32,rinv32);
 462             felec            = _mm_mul_ps(velec,rinvsq32);
 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             /* Calculate temporary vectorial force */
 470             tx               = _mm_mul_ps(fscal,dx32);
 471             ty               = _mm_mul_ps(fscal,dy32);
 472             tz               = _mm_mul_ps(fscal,dz32);
 473
 474             /* Update vectorial force */
 475             fix3             = _mm_add_ps(fix3,tx);
 476             fiy3             = _mm_add_ps(fiy3,ty);
 477             fiz3             = _mm_add_ps(fiz3,tz);
 478
 479             fjx2             = _mm_add_ps(fjx2,tx);
 480             fjy2             = _mm_add_ps(fjy2,ty);
 481             fjz2             = _mm_add_ps(fjz2,tz);
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             /* COULOMB ELECTROSTATICS */
 488             velec            = _mm_mul_ps(qq33,rinv33);
 489             felec            = _mm_mul_ps(velec,rinvsq33);
 490
 491             /* Update potential sum for this i atom from the interaction with this j atom. */
 492             velecsum         = _mm_add_ps(velecsum,velec);
 493
 494             fscal            = felec;
 495
 496             /* Calculate temporary vectorial force */
 497             tx               = _mm_mul_ps(fscal,dx33);
 498             ty               = _mm_mul_ps(fscal,dy33);
 499             tz               = _mm_mul_ps(fscal,dz33);
 500
 501             /* Update vectorial force */
 502             fix3             = _mm_add_ps(fix3,tx);
 503             fiy3             = _mm_add_ps(fiy3,ty);
 504             fiz3             = _mm_add_ps(fiz3,tz);
 505
 506             fjx3             = _mm_add_ps(fjx3,tx);
 507             fjy3             = _mm_add_ps(fjy3,ty);
 508             fjz3             = _mm_add_ps(fjz3,tz);
 509
 510             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
 511                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
 512                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 513
 514             /* Inner loop uses 252 flops */
 515         }
 516
 517         if(jidx<j_index_end)
 518         {
 519
 520             /* Get j neighbor index, and coordinate index */
 521             jnrA             = jjnr[jidx];
 522             jnrB             = jjnr[jidx+1];
 523             jnrC             = jjnr[jidx+2];
 524             jnrD             = jjnr[jidx+3];
 525
 526             /* Sign of each element will be negative for non-real atoms.
 527              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 528              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 529              */
 530             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 531             jnrA       = (jnrA>=0) ? jnrA : 0;
 532             jnrB       = (jnrB>=0) ? jnrB : 0;
 533             jnrC       = (jnrC>=0) ? jnrC : 0;
 534             jnrD       = (jnrD>=0) ? jnrD : 0;
 535
 536             j_coord_offsetA  = DIM*jnrA;
 537             j_coord_offsetB  = DIM*jnrB;
 538             j_coord_offsetC  = DIM*jnrC;
 539             j_coord_offsetD  = DIM*jnrD;
 540
 541             /* load j atom coordinates */
 542             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 543                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 544                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 545
 546             /* Calculate displacement vector */
 547             dx11             = _mm_sub_ps(ix1,jx1);
 548             dy11             = _mm_sub_ps(iy1,jy1);
 549             dz11             = _mm_sub_ps(iz1,jz1);
 550             dx12             = _mm_sub_ps(ix1,jx2);
 551             dy12             = _mm_sub_ps(iy1,jy2);
 552             dz12             = _mm_sub_ps(iz1,jz2);
 553             dx13             = _mm_sub_ps(ix1,jx3);
 554             dy13             = _mm_sub_ps(iy1,jy3);
 555             dz13             = _mm_sub_ps(iz1,jz3);
 556             dx21             = _mm_sub_ps(ix2,jx1);
 557             dy21             = _mm_sub_ps(iy2,jy1);
 558             dz21             = _mm_sub_ps(iz2,jz1);
 559             dx22             = _mm_sub_ps(ix2,jx2);
 560             dy22             = _mm_sub_ps(iy2,jy2);
 561             dz22             = _mm_sub_ps(iz2,jz2);
 562             dx23             = _mm_sub_ps(ix2,jx3);
 563             dy23             = _mm_sub_ps(iy2,jy3);
 564             dz23             = _mm_sub_ps(iz2,jz3);
 565             dx31             = _mm_sub_ps(ix3,jx1);
 566             dy31             = _mm_sub_ps(iy3,jy1);
 567             dz31             = _mm_sub_ps(iz3,jz1);
 568             dx32             = _mm_sub_ps(ix3,jx2);
 569             dy32             = _mm_sub_ps(iy3,jy2);
 570             dz32             = _mm_sub_ps(iz3,jz2);
 571             dx33             = _mm_sub_ps(ix3,jx3);
 572             dy33             = _mm_sub_ps(iy3,jy3);
 573             dz33             = _mm_sub_ps(iz3,jz3);
 574
 575             /* Calculate squared distance and things based on it */
 576             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 577             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 578             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 579             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 580             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 581             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 582             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 583             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 584             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 585
 586             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 587             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 588             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 589             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 590             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 591             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 592             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 593             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 594             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 595
 596             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 597             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 598             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 599             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 600             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 601             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 602             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 603             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 604             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 605
 606             fjx1             = _mm_setzero_ps();
 607             fjy1             = _mm_setzero_ps();
 608             fjz1             = _mm_setzero_ps();
 609             fjx2             = _mm_setzero_ps();
 610             fjy2             = _mm_setzero_ps();
 611             fjz2             = _mm_setzero_ps();
 612             fjx3             = _mm_setzero_ps();
 613             fjy3             = _mm_setzero_ps();
 614             fjz3             = _mm_setzero_ps();
 615
 616             /**************************
 617              * CALCULATE INTERACTIONS *
 618              **************************/
 619
 620             /* COULOMB ELECTROSTATICS */
 621             velec            = _mm_mul_ps(qq11,rinv11);
 622             felec            = _mm_mul_ps(velec,rinvsq11);
 623
 624             /* Update potential sum for this i atom from the interaction with this j atom. */
 625             velec            = _mm_andnot_ps(dummy_mask,velec);
 626             velecsum         = _mm_add_ps(velecsum,velec);
 627
 628             fscal            = felec;
 629
 630             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 631
 632             /* Calculate temporary vectorial force */
 633             tx               = _mm_mul_ps(fscal,dx11);
 634             ty               = _mm_mul_ps(fscal,dy11);
 635             tz               = _mm_mul_ps(fscal,dz11);
 636
 637             /* Update vectorial force */
 638             fix1             = _mm_add_ps(fix1,tx);
 639             fiy1             = _mm_add_ps(fiy1,ty);
 640             fiz1             = _mm_add_ps(fiz1,tz);
 641
 642             fjx1             = _mm_add_ps(fjx1,tx);
 643             fjy1             = _mm_add_ps(fjy1,ty);
 644             fjz1             = _mm_add_ps(fjz1,tz);
 645
 646             /**************************
 647              * CALCULATE INTERACTIONS *
 648              **************************/
 649
 650             /* COULOMB ELECTROSTATICS */
 651             velec            = _mm_mul_ps(qq12,rinv12);
 652             felec            = _mm_mul_ps(velec,rinvsq12);
 653
 654             /* Update potential sum for this i atom from the interaction with this j atom. */
 655             velec            = _mm_andnot_ps(dummy_mask,velec);
 656             velecsum         = _mm_add_ps(velecsum,velec);
 657
 658             fscal            = felec;
 659
 660             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 661
 662             /* Calculate temporary vectorial force */
 663             tx               = _mm_mul_ps(fscal,dx12);
 664             ty               = _mm_mul_ps(fscal,dy12);
 665             tz               = _mm_mul_ps(fscal,dz12);
 666
 667             /* Update vectorial force */
 668             fix1             = _mm_add_ps(fix1,tx);
 669             fiy1             = _mm_add_ps(fiy1,ty);
 670             fiz1             = _mm_add_ps(fiz1,tz);
 671
 672             fjx2             = _mm_add_ps(fjx2,tx);
 673             fjy2             = _mm_add_ps(fjy2,ty);
 674             fjz2             = _mm_add_ps(fjz2,tz);
 675
 676             /**************************
 677              * CALCULATE INTERACTIONS *
 678              **************************/
 679
 680             /* COULOMB ELECTROSTATICS */
 681             velec            = _mm_mul_ps(qq13,rinv13);
 682             felec            = _mm_mul_ps(velec,rinvsq13);
 683
 684             /* Update potential sum for this i atom from the interaction with this j atom. */
 685             velec            = _mm_andnot_ps(dummy_mask,velec);
 686             velecsum         = _mm_add_ps(velecsum,velec);
 687
 688             fscal            = felec;
 689
 690             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 691
 692             /* Calculate temporary vectorial force */
 693             tx               = _mm_mul_ps(fscal,dx13);
 694             ty               = _mm_mul_ps(fscal,dy13);
 695             tz               = _mm_mul_ps(fscal,dz13);
 696
 697             /* Update vectorial force */
 698             fix1             = _mm_add_ps(fix1,tx);
 699             fiy1             = _mm_add_ps(fiy1,ty);
 700             fiz1             = _mm_add_ps(fiz1,tz);
 701
 702             fjx3             = _mm_add_ps(fjx3,tx);
 703             fjy3             = _mm_add_ps(fjy3,ty);
 704             fjz3             = _mm_add_ps(fjz3,tz);
 705
 706             /**************************
 707              * CALCULATE INTERACTIONS *
 708              **************************/
 709
 710             /* COULOMB ELECTROSTATICS */
 711             velec            = _mm_mul_ps(qq21,rinv21);
 712             felec            = _mm_mul_ps(velec,rinvsq21);
 713
 714             /* Update potential sum for this i atom from the interaction with this j atom. */
 715             velec            = _mm_andnot_ps(dummy_mask,velec);
 716             velecsum         = _mm_add_ps(velecsum,velec);
 717
 718             fscal            = felec;
 719
 720             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 721
 722             /* Calculate temporary vectorial force */
 723             tx               = _mm_mul_ps(fscal,dx21);
 724             ty               = _mm_mul_ps(fscal,dy21);
 725             tz               = _mm_mul_ps(fscal,dz21);
 726
 727             /* Update vectorial force */
 728             fix2             = _mm_add_ps(fix2,tx);
 729             fiy2             = _mm_add_ps(fiy2,ty);
 730             fiz2             = _mm_add_ps(fiz2,tz);
 731
 732             fjx1             = _mm_add_ps(fjx1,tx);
 733             fjy1             = _mm_add_ps(fjy1,ty);
 734             fjz1             = _mm_add_ps(fjz1,tz);
 735
 736             /**************************
 737              * CALCULATE INTERACTIONS *
 738              **************************/
 739
 740             /* COULOMB ELECTROSTATICS */
 741             velec            = _mm_mul_ps(qq22,rinv22);
 742             felec            = _mm_mul_ps(velec,rinvsq22);
 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             /* Calculate temporary vectorial force */
 753             tx               = _mm_mul_ps(fscal,dx22);
 754             ty               = _mm_mul_ps(fscal,dy22);
 755             tz               = _mm_mul_ps(fscal,dz22);
 756
 757             /* Update vectorial force */
 758             fix2             = _mm_add_ps(fix2,tx);
 759             fiy2             = _mm_add_ps(fiy2,ty);
 760             fiz2             = _mm_add_ps(fiz2,tz);
 761
 762             fjx2             = _mm_add_ps(fjx2,tx);
 763             fjy2             = _mm_add_ps(fjy2,ty);
 764             fjz2             = _mm_add_ps(fjz2,tz);
 765
 766             /**************************
 767              * CALCULATE INTERACTIONS *
 768              **************************/
 769
 770             /* COULOMB ELECTROSTATICS */
 771             velec            = _mm_mul_ps(qq23,rinv23);
 772             felec            = _mm_mul_ps(velec,rinvsq23);
 773
 774             /* Update potential sum for this i atom from the interaction with this j atom. */
 775             velec            = _mm_andnot_ps(dummy_mask,velec);
 776             velecsum         = _mm_add_ps(velecsum,velec);
 777
 778             fscal            = felec;
 779
 780             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 781
 782             /* Calculate temporary vectorial force */
 783             tx               = _mm_mul_ps(fscal,dx23);
 784             ty               = _mm_mul_ps(fscal,dy23);
 785             tz               = _mm_mul_ps(fscal,dz23);
 786
 787             /* Update vectorial force */
 788             fix2             = _mm_add_ps(fix2,tx);
 789             fiy2             = _mm_add_ps(fiy2,ty);
 790             fiz2             = _mm_add_ps(fiz2,tz);
 791
 792             fjx3             = _mm_add_ps(fjx3,tx);
 793             fjy3             = _mm_add_ps(fjy3,ty);
 794             fjz3             = _mm_add_ps(fjz3,tz);
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             /* COULOMB ELECTROSTATICS */
 801             velec            = _mm_mul_ps(qq31,rinv31);
 802             felec            = _mm_mul_ps(velec,rinvsq31);
 803
 804             /* Update potential sum for this i atom from the interaction with this j atom. */
 805             velec            = _mm_andnot_ps(dummy_mask,velec);
 806             velecsum         = _mm_add_ps(velecsum,velec);
 807
 808             fscal            = felec;
 809
 810             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 811
 812             /* Calculate temporary vectorial force */
 813             tx               = _mm_mul_ps(fscal,dx31);
 814             ty               = _mm_mul_ps(fscal,dy31);
 815             tz               = _mm_mul_ps(fscal,dz31);
 816
 817             /* Update vectorial force */
 818             fix3             = _mm_add_ps(fix3,tx);
 819             fiy3             = _mm_add_ps(fiy3,ty);
 820             fiz3             = _mm_add_ps(fiz3,tz);
 821
 822             fjx1             = _mm_add_ps(fjx1,tx);
 823             fjy1             = _mm_add_ps(fjy1,ty);
 824             fjz1             = _mm_add_ps(fjz1,tz);
 825
 826             /**************************
 827              * CALCULATE INTERACTIONS *
 828              **************************/
 829
 830             /* COULOMB ELECTROSTATICS */
 831             velec            = _mm_mul_ps(qq32,rinv32);
 832             felec            = _mm_mul_ps(velec,rinvsq32);
 833
 834             /* Update potential sum for this i atom from the interaction with this j atom. */
 835             velec            = _mm_andnot_ps(dummy_mask,velec);
 836             velecsum         = _mm_add_ps(velecsum,velec);
 837
 838             fscal            = felec;
 839
 840             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 841
 842             /* Calculate temporary vectorial force */
 843             tx               = _mm_mul_ps(fscal,dx32);
 844             ty               = _mm_mul_ps(fscal,dy32);
 845             tz               = _mm_mul_ps(fscal,dz32);
 846
 847             /* Update vectorial force */
 848             fix3             = _mm_add_ps(fix3,tx);
 849             fiy3             = _mm_add_ps(fiy3,ty);
 850             fiz3             = _mm_add_ps(fiz3,tz);
 851
 852             fjx2             = _mm_add_ps(fjx2,tx);
 853             fjy2             = _mm_add_ps(fjy2,ty);
 854             fjz2             = _mm_add_ps(fjz2,tz);
 855
 856             /**************************
 857              * CALCULATE INTERACTIONS *
 858              **************************/
 859
 860             /* COULOMB ELECTROSTATICS */
 861             velec            = _mm_mul_ps(qq33,rinv33);
 862             felec            = _mm_mul_ps(velec,rinvsq33);
 863
 864             /* Update potential sum for this i atom from the interaction with this j atom. */
 865             velec            = _mm_andnot_ps(dummy_mask,velec);
 866             velecsum         = _mm_add_ps(velecsum,velec);
 867
 868             fscal            = felec;
 869
 870             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 871
 872             /* Calculate temporary vectorial force */
 873             tx               = _mm_mul_ps(fscal,dx33);
 874             ty               = _mm_mul_ps(fscal,dy33);
 875             tz               = _mm_mul_ps(fscal,dz33);
 876
 877             /* Update vectorial force */
 878             fix3             = _mm_add_ps(fix3,tx);
 879             fiy3             = _mm_add_ps(fiy3,ty);
 880             fiz3             = _mm_add_ps(fiz3,tz);
 881
 882             fjx3             = _mm_add_ps(fjx3,tx);
 883             fjy3             = _mm_add_ps(fjy3,ty);
 884             fjz3             = _mm_add_ps(fjz3,tz);
 885
 886             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
 887                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
 888                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 889
 890             /* Inner loop uses 252 flops */
 891         }
 892
 893         /* End of innermost loop */
 894
 895         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 896                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 897
 898         ggid                        = gid[iidx];
 899         /* Update potential energies */
 900         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 901
 902         /* Increment number of inner iterations */
 903         inneriter                  += j_index_end - j_index_start;
 904
 905         /* Outer loop uses 28 flops */
 906     }
 907
 908     /* Increment number of outer iterations */
 909     outeriter        += nri;
 910
 911     /* Update outer/inner flops */
 912
 913     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*28 + inneriter*252);
 914 }
 915 /*
 916  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse2_single
 917  * Electrostatics interaction: Coulomb
 918  * VdW interaction:            None
 919  * Geometry:                   Water4-Water4
 920  * Calculate force/pot:        Force
 921  */
 922 void
 923 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse2_single
 924                     (t_nblist * gmx_restrict                nlist,
 925                      rvec * gmx_restrict                    xx,
 926                      rvec * gmx_restrict                    ff,
 927                      t_forcerec * gmx_restrict              fr,
 928                      t_mdatoms * gmx_restrict               mdatoms,
 929                      nb_kernel_data_t * gmx_restrict        kernel_data,
 930                      t_nrnb * gmx_restrict                  nrnb)
 931 {
 932     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 933      * just 0 for non-waters.
 934      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 935      * jnr indices corresponding to data put in the four positions in the SIMD register.
 936      */
 937     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 938     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 939     int              jnrA,jnrB,jnrC,jnrD;
 940     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 941     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 942     real             shX,shY,shZ,rcutoff_scalar;
 943     real             *shiftvec,*fshift,*x,*f;
 944     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 945     int              vdwioffset1;
 946     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 947     int              vdwioffset2;
 948     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 949     int              vdwioffset3;
 950     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 951     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 952     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 953     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 954     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 955     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 956     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 957     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 958     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 959     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 960     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 961     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 962     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 963     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 964     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 965     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 966     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 967     real             *charge;
 968     __m128           dummy_mask,cutoff_mask;
 969     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 970     __m128           one     = _mm_set1_ps(1.0);
 971     __m128           two     = _mm_set1_ps(2.0);
 972     x                = xx[0];
 973     f                = ff[0];
 974
 975     nri              = nlist->nri;
 976     iinr             = nlist->iinr;
 977     jindex           = nlist->jindex;
 978     jjnr             = nlist->jjnr;
 979     shiftidx         = nlist->shift;
 980     gid              = nlist->gid;
 981     shiftvec         = fr->shift_vec[0];
 982     fshift           = fr->fshift[0];
 983     facel            = _mm_set1_ps(fr->epsfac);
 984     charge           = mdatoms->chargeA;
 985
 986     /* Setup water-specific parameters */
 987     inr              = nlist->iinr[0];
 988     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 989     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 990     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 991
 992     jq1              = _mm_set1_ps(charge[inr+1]);
 993     jq2              = _mm_set1_ps(charge[inr+2]);
 994     jq3              = _mm_set1_ps(charge[inr+3]);
 995     qq11             = _mm_mul_ps(iq1,jq1);
 996     qq12             = _mm_mul_ps(iq1,jq2);
 997     qq13             = _mm_mul_ps(iq1,jq3);
 998     qq21             = _mm_mul_ps(iq2,jq1);
 999     qq22             = _mm_mul_ps(iq2,jq2);
1000     qq23             = _mm_mul_ps(iq2,jq3);
1001     qq31             = _mm_mul_ps(iq3,jq1);
1002     qq32             = _mm_mul_ps(iq3,jq2);
1003     qq33             = _mm_mul_ps(iq3,jq3);
1004
1005     /* Avoid stupid compiler warnings */
1006     jnrA = jnrB = jnrC = jnrD = 0;
1007     j_coord_offsetA = 0;
1008     j_coord_offsetB = 0;
1009     j_coord_offsetC = 0;
1010     j_coord_offsetD = 0;
1011
1012     outeriter        = 0;
1013     inneriter        = 0;
1014
1015     /* Start outer loop over neighborlists */
1016     for(iidx=0; iidx<nri; iidx++)
1017     {
1018         /* Load shift vector for this list */
1019         i_shift_offset   = DIM*shiftidx[iidx];
1020         shX              = shiftvec[i_shift_offset+XX];
1021         shY              = shiftvec[i_shift_offset+YY];
1022         shZ              = shiftvec[i_shift_offset+ZZ];
1023
1024         /* Load limits for loop over neighbors */
1025         j_index_start    = jindex[iidx];
1026         j_index_end      = jindex[iidx+1];
1027
1028         /* Get outer coordinate index */
1029         inr              = iinr[iidx];
1030         i_coord_offset   = DIM*inr;
1031
1032         /* Load i particle coords and add shift vector */
1033         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1034         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1035         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1036         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1037         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1038         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1039         ix3              = _mm_set1_ps(shX + x[i_coord_offset+DIM*3+XX]);
1040         iy3              = _mm_set1_ps(shY + x[i_coord_offset+DIM*3+YY]);
1041         iz3              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*3+ZZ]);
1042
1043         fix1             = _mm_setzero_ps();
1044         fiy1             = _mm_setzero_ps();
1045         fiz1             = _mm_setzero_ps();
1046         fix2             = _mm_setzero_ps();
1047         fiy2             = _mm_setzero_ps();
1048         fiz2             = _mm_setzero_ps();
1049         fix3             = _mm_setzero_ps();
1050         fiy3             = _mm_setzero_ps();
1051         fiz3             = _mm_setzero_ps();
1052
1053         /* Start inner kernel loop */
1054         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1055         {
1056
1057             /* Get j neighbor index, and coordinate index */
1058             jnrA             = jjnr[jidx];
1059             jnrB             = jjnr[jidx+1];
1060             jnrC             = jjnr[jidx+2];
1061             jnrD             = jjnr[jidx+3];
1062
1063             j_coord_offsetA  = DIM*jnrA;
1064             j_coord_offsetB  = DIM*jnrB;
1065             j_coord_offsetC  = DIM*jnrC;
1066             j_coord_offsetD  = DIM*jnrD;
1067
1068             /* load j atom coordinates */
1069             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1070                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1071                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1072
1073             /* Calculate displacement vector */
1074             dx11             = _mm_sub_ps(ix1,jx1);
1075             dy11             = _mm_sub_ps(iy1,jy1);
1076             dz11             = _mm_sub_ps(iz1,jz1);
1077             dx12             = _mm_sub_ps(ix1,jx2);
1078             dy12             = _mm_sub_ps(iy1,jy2);
1079             dz12             = _mm_sub_ps(iz1,jz2);
1080             dx13             = _mm_sub_ps(ix1,jx3);
1081             dy13             = _mm_sub_ps(iy1,jy3);
1082             dz13             = _mm_sub_ps(iz1,jz3);
1083             dx21             = _mm_sub_ps(ix2,jx1);
1084             dy21             = _mm_sub_ps(iy2,jy1);
1085             dz21             = _mm_sub_ps(iz2,jz1);
1086             dx22             = _mm_sub_ps(ix2,jx2);
1087             dy22             = _mm_sub_ps(iy2,jy2);
1088             dz22             = _mm_sub_ps(iz2,jz2);
1089             dx23             = _mm_sub_ps(ix2,jx3);
1090             dy23             = _mm_sub_ps(iy2,jy3);
1091             dz23             = _mm_sub_ps(iz2,jz3);
1092             dx31             = _mm_sub_ps(ix3,jx1);
1093             dy31             = _mm_sub_ps(iy3,jy1);
1094             dz31             = _mm_sub_ps(iz3,jz1);
1095             dx32             = _mm_sub_ps(ix3,jx2);
1096             dy32             = _mm_sub_ps(iy3,jy2);
1097             dz32             = _mm_sub_ps(iz3,jz2);
1098             dx33             = _mm_sub_ps(ix3,jx3);
1099             dy33             = _mm_sub_ps(iy3,jy3);
1100             dz33             = _mm_sub_ps(iz3,jz3);
1101
1102             /* Calculate squared distance and things based on it */
1103             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1104             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1105             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1106             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1107             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1108             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1109             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1110             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1111             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1112
1113             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1114             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1115             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1116             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1117             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1118             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1119             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1120             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1121             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1122
1123             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1124             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1125             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1126             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1127             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1128             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1129             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1130             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1131             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1132
1133             fjx1             = _mm_setzero_ps();
1134             fjy1             = _mm_setzero_ps();
1135             fjz1             = _mm_setzero_ps();
1136             fjx2             = _mm_setzero_ps();
1137             fjy2             = _mm_setzero_ps();
1138             fjz2             = _mm_setzero_ps();
1139             fjx3             = _mm_setzero_ps();
1140             fjy3             = _mm_setzero_ps();
1141             fjz3             = _mm_setzero_ps();
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             /* COULOMB ELECTROSTATICS */
1148             velec            = _mm_mul_ps(qq11,rinv11);
1149             felec            = _mm_mul_ps(velec,rinvsq11);
1150
1151             fscal            = felec;
1152
1153             /* Calculate temporary vectorial force */
1154             tx               = _mm_mul_ps(fscal,dx11);
1155             ty               = _mm_mul_ps(fscal,dy11);
1156             tz               = _mm_mul_ps(fscal,dz11);
1157
1158             /* Update vectorial force */
1159             fix1             = _mm_add_ps(fix1,tx);
1160             fiy1             = _mm_add_ps(fiy1,ty);
1161             fiz1             = _mm_add_ps(fiz1,tz);
1162
1163             fjx1             = _mm_add_ps(fjx1,tx);
1164             fjy1             = _mm_add_ps(fjy1,ty);
1165             fjz1             = _mm_add_ps(fjz1,tz);
1166
1167             /**************************
1168              * CALCULATE INTERACTIONS *
1169              **************************/
1170
1171             /* COULOMB ELECTROSTATICS */
1172             velec            = _mm_mul_ps(qq12,rinv12);
1173             felec            = _mm_mul_ps(velec,rinvsq12);
1174
1175             fscal            = felec;
1176
1177             /* Calculate temporary vectorial force */
1178             tx               = _mm_mul_ps(fscal,dx12);
1179             ty               = _mm_mul_ps(fscal,dy12);
1180             tz               = _mm_mul_ps(fscal,dz12);
1181
1182             /* Update vectorial force */
1183             fix1             = _mm_add_ps(fix1,tx);
1184             fiy1             = _mm_add_ps(fiy1,ty);
1185             fiz1             = _mm_add_ps(fiz1,tz);
1186
1187             fjx2             = _mm_add_ps(fjx2,tx);
1188             fjy2             = _mm_add_ps(fjy2,ty);
1189             fjz2             = _mm_add_ps(fjz2,tz);
1190
1191             /**************************
1192              * CALCULATE INTERACTIONS *
1193              **************************/
1194
1195             /* COULOMB ELECTROSTATICS */
1196             velec            = _mm_mul_ps(qq13,rinv13);
1197             felec            = _mm_mul_ps(velec,rinvsq13);
1198
1199             fscal            = felec;
1200
1201             /* Calculate temporary vectorial force */
1202             tx               = _mm_mul_ps(fscal,dx13);
1203             ty               = _mm_mul_ps(fscal,dy13);
1204             tz               = _mm_mul_ps(fscal,dz13);
1205
1206             /* Update vectorial force */
1207             fix1             = _mm_add_ps(fix1,tx);
1208             fiy1             = _mm_add_ps(fiy1,ty);
1209             fiz1             = _mm_add_ps(fiz1,tz);
1210
1211             fjx3             = _mm_add_ps(fjx3,tx);
1212             fjy3             = _mm_add_ps(fjy3,ty);
1213             fjz3             = _mm_add_ps(fjz3,tz);
1214
1215             /**************************
1216              * CALCULATE INTERACTIONS *
1217              **************************/
1218
1219             /* COULOMB ELECTROSTATICS */
1220             velec            = _mm_mul_ps(qq21,rinv21);
1221             felec            = _mm_mul_ps(velec,rinvsq21);
1222
1223             fscal            = felec;
1224
1225             /* Calculate temporary vectorial force */
1226             tx               = _mm_mul_ps(fscal,dx21);
1227             ty               = _mm_mul_ps(fscal,dy21);
1228             tz               = _mm_mul_ps(fscal,dz21);
1229
1230             /* Update vectorial force */
1231             fix2             = _mm_add_ps(fix2,tx);
1232             fiy2             = _mm_add_ps(fiy2,ty);
1233             fiz2             = _mm_add_ps(fiz2,tz);
1234
1235             fjx1             = _mm_add_ps(fjx1,tx);
1236             fjy1             = _mm_add_ps(fjy1,ty);
1237             fjz1             = _mm_add_ps(fjz1,tz);
1238
1239             /**************************
1240              * CALCULATE INTERACTIONS *
1241              **************************/
1242
1243             /* COULOMB ELECTROSTATICS */
1244             velec            = _mm_mul_ps(qq22,rinv22);
1245             felec            = _mm_mul_ps(velec,rinvsq22);
1246
1247             fscal            = felec;
1248
1249             /* Calculate temporary vectorial force */
1250             tx               = _mm_mul_ps(fscal,dx22);
1251             ty               = _mm_mul_ps(fscal,dy22);
1252             tz               = _mm_mul_ps(fscal,dz22);
1253
1254             /* Update vectorial force */
1255             fix2             = _mm_add_ps(fix2,tx);
1256             fiy2             = _mm_add_ps(fiy2,ty);
1257             fiz2             = _mm_add_ps(fiz2,tz);
1258
1259             fjx2             = _mm_add_ps(fjx2,tx);
1260             fjy2             = _mm_add_ps(fjy2,ty);
1261             fjz2             = _mm_add_ps(fjz2,tz);
1262
1263             /**************************
1264              * CALCULATE INTERACTIONS *
1265              **************************/
1266
1267             /* COULOMB ELECTROSTATICS */
1268             velec            = _mm_mul_ps(qq23,rinv23);
1269             felec            = _mm_mul_ps(velec,rinvsq23);
1270
1271             fscal            = felec;
1272
1273             /* Calculate temporary vectorial force */
1274             tx               = _mm_mul_ps(fscal,dx23);
1275             ty               = _mm_mul_ps(fscal,dy23);
1276             tz               = _mm_mul_ps(fscal,dz23);
1277
1278             /* Update vectorial force */
1279             fix2             = _mm_add_ps(fix2,tx);
1280             fiy2             = _mm_add_ps(fiy2,ty);
1281             fiz2             = _mm_add_ps(fiz2,tz);
1282
1283             fjx3             = _mm_add_ps(fjx3,tx);
1284             fjy3             = _mm_add_ps(fjy3,ty);
1285             fjz3             = _mm_add_ps(fjz3,tz);
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* COULOMB ELECTROSTATICS */
1292             velec            = _mm_mul_ps(qq31,rinv31);
1293             felec            = _mm_mul_ps(velec,rinvsq31);
1294
1295             fscal            = felec;
1296
1297             /* Calculate temporary vectorial force */
1298             tx               = _mm_mul_ps(fscal,dx31);
1299             ty               = _mm_mul_ps(fscal,dy31);
1300             tz               = _mm_mul_ps(fscal,dz31);
1301
1302             /* Update vectorial force */
1303             fix3             = _mm_add_ps(fix3,tx);
1304             fiy3             = _mm_add_ps(fiy3,ty);
1305             fiz3             = _mm_add_ps(fiz3,tz);
1306
1307             fjx1             = _mm_add_ps(fjx1,tx);
1308             fjy1             = _mm_add_ps(fjy1,ty);
1309             fjz1             = _mm_add_ps(fjz1,tz);
1310
1311             /**************************
1312              * CALCULATE INTERACTIONS *
1313              **************************/
1314
1315             /* COULOMB ELECTROSTATICS */
1316             velec            = _mm_mul_ps(qq32,rinv32);
1317             felec            = _mm_mul_ps(velec,rinvsq32);
1318
1319             fscal            = felec;
1320
1321             /* Calculate temporary vectorial force */
1322             tx               = _mm_mul_ps(fscal,dx32);
1323             ty               = _mm_mul_ps(fscal,dy32);
1324             tz               = _mm_mul_ps(fscal,dz32);
1325
1326             /* Update vectorial force */
1327             fix3             = _mm_add_ps(fix3,tx);
1328             fiy3             = _mm_add_ps(fiy3,ty);
1329             fiz3             = _mm_add_ps(fiz3,tz);
1330
1331             fjx2             = _mm_add_ps(fjx2,tx);
1332             fjy2             = _mm_add_ps(fjy2,ty);
1333             fjz2             = _mm_add_ps(fjz2,tz);
1334
1335             /**************************
1336              * CALCULATE INTERACTIONS *
1337              **************************/
1338
1339             /* COULOMB ELECTROSTATICS */
1340             velec            = _mm_mul_ps(qq33,rinv33);
1341             felec            = _mm_mul_ps(velec,rinvsq33);
1342
1343             fscal            = felec;
1344
1345             /* Calculate temporary vectorial force */
1346             tx               = _mm_mul_ps(fscal,dx33);
1347             ty               = _mm_mul_ps(fscal,dy33);
1348             tz               = _mm_mul_ps(fscal,dz33);
1349
1350             /* Update vectorial force */
1351             fix3             = _mm_add_ps(fix3,tx);
1352             fiy3             = _mm_add_ps(fiy3,ty);
1353             fiz3             = _mm_add_ps(fiz3,tz);
1354
1355             fjx3             = _mm_add_ps(fjx3,tx);
1356             fjy3             = _mm_add_ps(fjy3,ty);
1357             fjz3             = _mm_add_ps(fjz3,tz);
1358
1359             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
1360                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
1361                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1362
1363             /* Inner loop uses 243 flops */
1364         }
1365
1366         if(jidx<j_index_end)
1367         {
1368
1369             /* Get j neighbor index, and coordinate index */
1370             jnrA             = jjnr[jidx];
1371             jnrB             = jjnr[jidx+1];
1372             jnrC             = jjnr[jidx+2];
1373             jnrD             = jjnr[jidx+3];
1374
1375             /* Sign of each element will be negative for non-real atoms.
1376              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1377              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1378              */
1379             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1380             jnrA       = (jnrA>=0) ? jnrA : 0;
1381             jnrB       = (jnrB>=0) ? jnrB : 0;
1382             jnrC       = (jnrC>=0) ? jnrC : 0;
1383             jnrD       = (jnrD>=0) ? jnrD : 0;
1384
1385             j_coord_offsetA  = DIM*jnrA;
1386             j_coord_offsetB  = DIM*jnrB;
1387             j_coord_offsetC  = DIM*jnrC;
1388             j_coord_offsetD  = DIM*jnrD;
1389
1390             /* load j atom coordinates */
1391             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1392                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1393                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1394
1395             /* Calculate displacement vector */
1396             dx11             = _mm_sub_ps(ix1,jx1);
1397             dy11             = _mm_sub_ps(iy1,jy1);
1398             dz11             = _mm_sub_ps(iz1,jz1);
1399             dx12             = _mm_sub_ps(ix1,jx2);
1400             dy12             = _mm_sub_ps(iy1,jy2);
1401             dz12             = _mm_sub_ps(iz1,jz2);
1402             dx13             = _mm_sub_ps(ix1,jx3);
1403             dy13             = _mm_sub_ps(iy1,jy3);
1404             dz13             = _mm_sub_ps(iz1,jz3);
1405             dx21             = _mm_sub_ps(ix2,jx1);
1406             dy21             = _mm_sub_ps(iy2,jy1);
1407             dz21             = _mm_sub_ps(iz2,jz1);
1408             dx22             = _mm_sub_ps(ix2,jx2);
1409             dy22             = _mm_sub_ps(iy2,jy2);
1410             dz22             = _mm_sub_ps(iz2,jz2);
1411             dx23             = _mm_sub_ps(ix2,jx3);
1412             dy23             = _mm_sub_ps(iy2,jy3);
1413             dz23             = _mm_sub_ps(iz2,jz3);
1414             dx31             = _mm_sub_ps(ix3,jx1);
1415             dy31             = _mm_sub_ps(iy3,jy1);
1416             dz31             = _mm_sub_ps(iz3,jz1);
1417             dx32             = _mm_sub_ps(ix3,jx2);
1418             dy32             = _mm_sub_ps(iy3,jy2);
1419             dz32             = _mm_sub_ps(iz3,jz2);
1420             dx33             = _mm_sub_ps(ix3,jx3);
1421             dy33             = _mm_sub_ps(iy3,jy3);
1422             dz33             = _mm_sub_ps(iz3,jz3);
1423
1424             /* Calculate squared distance and things based on it */
1425             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1426             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1427             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1428             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1429             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1430             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1431             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1432             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1433             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1434
1435             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1436             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1437             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1438             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1439             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1440             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1441             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1442             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1443             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1444
1445             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1446             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1447             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1448             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1449             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1450             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1451             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1452             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1453             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1454
1455             fjx1             = _mm_setzero_ps();
1456             fjy1             = _mm_setzero_ps();
1457             fjz1             = _mm_setzero_ps();
1458             fjx2             = _mm_setzero_ps();
1459             fjy2             = _mm_setzero_ps();
1460             fjz2             = _mm_setzero_ps();
1461             fjx3             = _mm_setzero_ps();
1462             fjy3             = _mm_setzero_ps();
1463             fjz3             = _mm_setzero_ps();
1464
1465             /**************************
1466              * CALCULATE INTERACTIONS *
1467              **************************/
1468
1469             /* COULOMB ELECTROSTATICS */
1470             velec            = _mm_mul_ps(qq11,rinv11);
1471             felec            = _mm_mul_ps(velec,rinvsq11);
1472
1473             fscal            = felec;
1474
1475             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1476
1477             /* Calculate temporary vectorial force */
1478             tx               = _mm_mul_ps(fscal,dx11);
1479             ty               = _mm_mul_ps(fscal,dy11);
1480             tz               = _mm_mul_ps(fscal,dz11);
1481
1482             /* Update vectorial force */
1483             fix1             = _mm_add_ps(fix1,tx);
1484             fiy1             = _mm_add_ps(fiy1,ty);
1485             fiz1             = _mm_add_ps(fiz1,tz);
1486
1487             fjx1             = _mm_add_ps(fjx1,tx);
1488             fjy1             = _mm_add_ps(fjy1,ty);
1489             fjz1             = _mm_add_ps(fjz1,tz);
1490
1491             /**************************
1492              * CALCULATE INTERACTIONS *
1493              **************************/
1494
1495             /* COULOMB ELECTROSTATICS */
1496             velec            = _mm_mul_ps(qq12,rinv12);
1497             felec            = _mm_mul_ps(velec,rinvsq12);
1498
1499             fscal            = felec;
1500
1501             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1502
1503             /* Calculate temporary vectorial force */
1504             tx               = _mm_mul_ps(fscal,dx12);
1505             ty               = _mm_mul_ps(fscal,dy12);
1506             tz               = _mm_mul_ps(fscal,dz12);
1507
1508             /* Update vectorial force */
1509             fix1             = _mm_add_ps(fix1,tx);
1510             fiy1             = _mm_add_ps(fiy1,ty);
1511             fiz1             = _mm_add_ps(fiz1,tz);
1512
1513             fjx2             = _mm_add_ps(fjx2,tx);
1514             fjy2             = _mm_add_ps(fjy2,ty);
1515             fjz2             = _mm_add_ps(fjz2,tz);
1516
1517             /**************************
1518              * CALCULATE INTERACTIONS *
1519              **************************/
1520
1521             /* COULOMB ELECTROSTATICS */
1522             velec            = _mm_mul_ps(qq13,rinv13);
1523             felec            = _mm_mul_ps(velec,rinvsq13);
1524
1525             fscal            = felec;
1526
1527             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1528
1529             /* Calculate temporary vectorial force */
1530             tx               = _mm_mul_ps(fscal,dx13);
1531             ty               = _mm_mul_ps(fscal,dy13);
1532             tz               = _mm_mul_ps(fscal,dz13);
1533
1534             /* Update vectorial force */
1535             fix1             = _mm_add_ps(fix1,tx);
1536             fiy1             = _mm_add_ps(fiy1,ty);
1537             fiz1             = _mm_add_ps(fiz1,tz);
1538
1539             fjx3             = _mm_add_ps(fjx3,tx);
1540             fjy3             = _mm_add_ps(fjy3,ty);
1541             fjz3             = _mm_add_ps(fjz3,tz);
1542
1543             /**************************
1544              * CALCULATE INTERACTIONS *
1545              **************************/
1546
1547             /* COULOMB ELECTROSTATICS */
1548             velec            = _mm_mul_ps(qq21,rinv21);
1549             felec            = _mm_mul_ps(velec,rinvsq21);
1550
1551             fscal            = felec;
1552
1553             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1554
1555             /* Calculate temporary vectorial force */
1556             tx               = _mm_mul_ps(fscal,dx21);
1557             ty               = _mm_mul_ps(fscal,dy21);
1558             tz               = _mm_mul_ps(fscal,dz21);
1559
1560             /* Update vectorial force */
1561             fix2             = _mm_add_ps(fix2,tx);
1562             fiy2             = _mm_add_ps(fiy2,ty);
1563             fiz2             = _mm_add_ps(fiz2,tz);
1564
1565             fjx1             = _mm_add_ps(fjx1,tx);
1566             fjy1             = _mm_add_ps(fjy1,ty);
1567             fjz1             = _mm_add_ps(fjz1,tz);
1568
1569             /**************************
1570              * CALCULATE INTERACTIONS *
1571              **************************/
1572
1573             /* COULOMB ELECTROSTATICS */
1574             velec            = _mm_mul_ps(qq22,rinv22);
1575             felec            = _mm_mul_ps(velec,rinvsq22);
1576
1577             fscal            = felec;
1578
1579             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1580
1581             /* Calculate temporary vectorial force */
1582             tx               = _mm_mul_ps(fscal,dx22);
1583             ty               = _mm_mul_ps(fscal,dy22);
1584             tz               = _mm_mul_ps(fscal,dz22);
1585
1586             /* Update vectorial force */
1587             fix2             = _mm_add_ps(fix2,tx);
1588             fiy2             = _mm_add_ps(fiy2,ty);
1589             fiz2             = _mm_add_ps(fiz2,tz);
1590
1591             fjx2             = _mm_add_ps(fjx2,tx);
1592             fjy2             = _mm_add_ps(fjy2,ty);
1593             fjz2             = _mm_add_ps(fjz2,tz);
1594
1595             /**************************
1596              * CALCULATE INTERACTIONS *
1597              **************************/
1598
1599             /* COULOMB ELECTROSTATICS */
1600             velec            = _mm_mul_ps(qq23,rinv23);
1601             felec            = _mm_mul_ps(velec,rinvsq23);
1602
1603             fscal            = felec;
1604
1605             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1606
1607             /* Calculate temporary vectorial force */
1608             tx               = _mm_mul_ps(fscal,dx23);
1609             ty               = _mm_mul_ps(fscal,dy23);
1610             tz               = _mm_mul_ps(fscal,dz23);
1611
1612             /* Update vectorial force */
1613             fix2             = _mm_add_ps(fix2,tx);
1614             fiy2             = _mm_add_ps(fiy2,ty);
1615             fiz2             = _mm_add_ps(fiz2,tz);
1616
1617             fjx3             = _mm_add_ps(fjx3,tx);
1618             fjy3             = _mm_add_ps(fjy3,ty);
1619             fjz3             = _mm_add_ps(fjz3,tz);
1620
1621             /**************************
1622              * CALCULATE INTERACTIONS *
1623              **************************/
1624
1625             /* COULOMB ELECTROSTATICS */
1626             velec            = _mm_mul_ps(qq31,rinv31);
1627             felec            = _mm_mul_ps(velec,rinvsq31);
1628
1629             fscal            = felec;
1630
1631             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1632
1633             /* Calculate temporary vectorial force */
1634             tx               = _mm_mul_ps(fscal,dx31);
1635             ty               = _mm_mul_ps(fscal,dy31);
1636             tz               = _mm_mul_ps(fscal,dz31);
1637
1638             /* Update vectorial force */
1639             fix3             = _mm_add_ps(fix3,tx);
1640             fiy3             = _mm_add_ps(fiy3,ty);
1641             fiz3             = _mm_add_ps(fiz3,tz);
1642
1643             fjx1             = _mm_add_ps(fjx1,tx);
1644             fjy1             = _mm_add_ps(fjy1,ty);
1645             fjz1             = _mm_add_ps(fjz1,tz);
1646
1647             /**************************
1648              * CALCULATE INTERACTIONS *
1649              **************************/
1650
1651             /* COULOMB ELECTROSTATICS */
1652             velec            = _mm_mul_ps(qq32,rinv32);
1653             felec            = _mm_mul_ps(velec,rinvsq32);
1654
1655             fscal            = felec;
1656
1657             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1658
1659             /* Calculate temporary vectorial force */
1660             tx               = _mm_mul_ps(fscal,dx32);
1661             ty               = _mm_mul_ps(fscal,dy32);
1662             tz               = _mm_mul_ps(fscal,dz32);
1663
1664             /* Update vectorial force */
1665             fix3             = _mm_add_ps(fix3,tx);
1666             fiy3             = _mm_add_ps(fiy3,ty);
1667             fiz3             = _mm_add_ps(fiz3,tz);
1668
1669             fjx2             = _mm_add_ps(fjx2,tx);
1670             fjy2             = _mm_add_ps(fjy2,ty);
1671             fjz2             = _mm_add_ps(fjz2,tz);
1672
1673             /**************************
1674              * CALCULATE INTERACTIONS *
1675              **************************/
1676
1677             /* COULOMB ELECTROSTATICS */
1678             velec            = _mm_mul_ps(qq33,rinv33);
1679             felec            = _mm_mul_ps(velec,rinvsq33);
1680
1681             fscal            = felec;
1682
1683             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1684
1685             /* Calculate temporary vectorial force */
1686             tx               = _mm_mul_ps(fscal,dx33);
1687             ty               = _mm_mul_ps(fscal,dy33);
1688             tz               = _mm_mul_ps(fscal,dz33);
1689
1690             /* Update vectorial force */
1691             fix3             = _mm_add_ps(fix3,tx);
1692             fiy3             = _mm_add_ps(fiy3,ty);
1693             fiz3             = _mm_add_ps(fiz3,tz);
1694
1695             fjx3             = _mm_add_ps(fjx3,tx);
1696             fjy3             = _mm_add_ps(fjy3,ty);
1697             fjz3             = _mm_add_ps(fjz3,tz);
1698
1699             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,
1700                                                    f+j_coord_offsetC+DIM,f+j_coord_offsetD+DIM,
1701                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1702
1703             /* Inner loop uses 243 flops */
1704         }
1705
1706         /* End of innermost loop */
1707
1708         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1709                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1710
1711         /* Increment number of inner iterations */
1712         inneriter                  += j_index_end - j_index_start;
1713
1714         /* Outer loop uses 27 flops */
1715     }
1716
1717     /* Increment number of outer iterations */
1718     outeriter        += nri;
1719
1720     /* Update outer/inner flops */
1721
1722     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*27 + inneriter*243);
1723 }