src/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecCoul_VdwNone_GeomW3W3_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_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_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_sse4_1_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_sse4_1_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              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           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  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
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 113     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 114     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 115
 116     jq0              = _mm_set1_ps(charge[inr+0]);
 117     jq1              = _mm_set1_ps(charge[inr+1]);
 118     jq2              = _mm_set1_ps(charge[inr+2]);
 119     qq00             = _mm_mul_ps(iq0,jq0);
 120     qq01             = _mm_mul_ps(iq0,jq1);
 121     qq02             = _mm_mul_ps(iq0,jq2);
 122     qq10             = _mm_mul_ps(iq1,jq0);
 123     qq11             = _mm_mul_ps(iq1,jq1);
 124     qq12             = _mm_mul_ps(iq1,jq2);
 125     qq20             = _mm_mul_ps(iq2,jq0);
 126     qq21             = _mm_mul_ps(iq2,jq1);
 127     qq22             = _mm_mul_ps(iq2,jq2);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 160                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 161
 162         fix0             = _mm_setzero_ps();
 163         fiy0             = _mm_setzero_ps();
 164         fiz0             = _mm_setzero_ps();
 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
 172         /* Reset potential sums */
 173         velecsum         = _mm_setzero_ps();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                               x+j_coord_offsetC,x+j_coord_offsetD,
 192                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm_sub_ps(ix0,jx0);
 196             dy00             = _mm_sub_ps(iy0,jy0);
 197             dz00             = _mm_sub_ps(iz0,jz0);
 198             dx01             = _mm_sub_ps(ix0,jx1);
 199             dy01             = _mm_sub_ps(iy0,jy1);
 200             dz01             = _mm_sub_ps(iz0,jz1);
 201             dx02             = _mm_sub_ps(ix0,jx2);
 202             dy02             = _mm_sub_ps(iy0,jy2);
 203             dz02             = _mm_sub_ps(iz0,jz2);
 204             dx10             = _mm_sub_ps(ix1,jx0);
 205             dy10             = _mm_sub_ps(iy1,jy0);
 206             dz10             = _mm_sub_ps(iz1,jz0);
 207             dx11             = _mm_sub_ps(ix1,jx1);
 208             dy11             = _mm_sub_ps(iy1,jy1);
 209             dz11             = _mm_sub_ps(iz1,jz1);
 210             dx12             = _mm_sub_ps(ix1,jx2);
 211             dy12             = _mm_sub_ps(iy1,jy2);
 212             dz12             = _mm_sub_ps(iz1,jz2);
 213             dx20             = _mm_sub_ps(ix2,jx0);
 214             dy20             = _mm_sub_ps(iy2,jy0);
 215             dz20             = _mm_sub_ps(iz2,jz0);
 216             dx21             = _mm_sub_ps(ix2,jx1);
 217             dy21             = _mm_sub_ps(iy2,jy1);
 218             dz21             = _mm_sub_ps(iz2,jz1);
 219             dx22             = _mm_sub_ps(ix2,jx2);
 220             dy22             = _mm_sub_ps(iy2,jy2);
 221             dz22             = _mm_sub_ps(iz2,jz2);
 222
 223             /* Calculate squared distance and things based on it */
 224             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 225             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 226             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 227             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 228             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 229             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 230             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 231             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 232             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 233
 234             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 235             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 236             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 237             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 238             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 239             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 240             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 241             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 242             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 243
 244             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 245             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 246             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 247             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 248             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 249             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 250             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 251             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 252             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 253
 254             fjx0             = _mm_setzero_ps();
 255             fjy0             = _mm_setzero_ps();
 256             fjz0             = _mm_setzero_ps();
 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
 264             /**************************
 265              * CALCULATE INTERACTIONS *
 266              **************************/
 267
 268             /* COULOMB ELECTROSTATICS */
 269             velec            = _mm_mul_ps(qq00,rinv00);
 270             felec            = _mm_mul_ps(velec,rinvsq00);
 271
 272             /* Update potential sum for this i atom from the interaction with this j atom. */
 273             velecsum         = _mm_add_ps(velecsum,velec);
 274
 275             fscal            = felec;
 276
 277             /* Calculate temporary vectorial force */
 278             tx               = _mm_mul_ps(fscal,dx00);
 279             ty               = _mm_mul_ps(fscal,dy00);
 280             tz               = _mm_mul_ps(fscal,dz00);
 281
 282             /* Update vectorial force */
 283             fix0             = _mm_add_ps(fix0,tx);
 284             fiy0             = _mm_add_ps(fiy0,ty);
 285             fiz0             = _mm_add_ps(fiz0,tz);
 286
 287             fjx0             = _mm_add_ps(fjx0,tx);
 288             fjy0             = _mm_add_ps(fjy0,ty);
 289             fjz0             = _mm_add_ps(fjz0,tz);
 290
 291             /**************************
 292              * CALCULATE INTERACTIONS *
 293              **************************/
 294
 295             /* COULOMB ELECTROSTATICS */
 296             velec            = _mm_mul_ps(qq01,rinv01);
 297             felec            = _mm_mul_ps(velec,rinvsq01);
 298
 299             /* Update potential sum for this i atom from the interaction with this j atom. */
 300             velecsum         = _mm_add_ps(velecsum,velec);
 301
 302             fscal            = felec;
 303
 304             /* Calculate temporary vectorial force */
 305             tx               = _mm_mul_ps(fscal,dx01);
 306             ty               = _mm_mul_ps(fscal,dy01);
 307             tz               = _mm_mul_ps(fscal,dz01);
 308
 309             /* Update vectorial force */
 310             fix0             = _mm_add_ps(fix0,tx);
 311             fiy0             = _mm_add_ps(fiy0,ty);
 312             fiz0             = _mm_add_ps(fiz0,tz);
 313
 314             fjx1             = _mm_add_ps(fjx1,tx);
 315             fjy1             = _mm_add_ps(fjy1,ty);
 316             fjz1             = _mm_add_ps(fjz1,tz);
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             /* COULOMB ELECTROSTATICS */
 323             velec            = _mm_mul_ps(qq02,rinv02);
 324             felec            = _mm_mul_ps(velec,rinvsq02);
 325
 326             /* Update potential sum for this i atom from the interaction with this j atom. */
 327             velecsum         = _mm_add_ps(velecsum,velec);
 328
 329             fscal            = felec;
 330
 331             /* Calculate temporary vectorial force */
 332             tx               = _mm_mul_ps(fscal,dx02);
 333             ty               = _mm_mul_ps(fscal,dy02);
 334             tz               = _mm_mul_ps(fscal,dz02);
 335
 336             /* Update vectorial force */
 337             fix0             = _mm_add_ps(fix0,tx);
 338             fiy0             = _mm_add_ps(fiy0,ty);
 339             fiz0             = _mm_add_ps(fiz0,tz);
 340
 341             fjx2             = _mm_add_ps(fjx2,tx);
 342             fjy2             = _mm_add_ps(fjy2,ty);
 343             fjz2             = _mm_add_ps(fjz2,tz);
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             /* COULOMB ELECTROSTATICS */
 350             velec            = _mm_mul_ps(qq10,rinv10);
 351             felec            = _mm_mul_ps(velec,rinvsq10);
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velecsum         = _mm_add_ps(velecsum,velec);
 355
 356             fscal            = felec;
 357
 358             /* Calculate temporary vectorial force */
 359             tx               = _mm_mul_ps(fscal,dx10);
 360             ty               = _mm_mul_ps(fscal,dy10);
 361             tz               = _mm_mul_ps(fscal,dz10);
 362
 363             /* Update vectorial force */
 364             fix1             = _mm_add_ps(fix1,tx);
 365             fiy1             = _mm_add_ps(fiy1,ty);
 366             fiz1             = _mm_add_ps(fiz1,tz);
 367
 368             fjx0             = _mm_add_ps(fjx0,tx);
 369             fjy0             = _mm_add_ps(fjy0,ty);
 370             fjz0             = _mm_add_ps(fjz0,tz);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             /* COULOMB ELECTROSTATICS */
 377             velec            = _mm_mul_ps(qq11,rinv11);
 378             felec            = _mm_mul_ps(velec,rinvsq11);
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velecsum         = _mm_add_ps(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             /* Calculate temporary vectorial force */
 386             tx               = _mm_mul_ps(fscal,dx11);
 387             ty               = _mm_mul_ps(fscal,dy11);
 388             tz               = _mm_mul_ps(fscal,dz11);
 389
 390             /* Update vectorial force */
 391             fix1             = _mm_add_ps(fix1,tx);
 392             fiy1             = _mm_add_ps(fiy1,ty);
 393             fiz1             = _mm_add_ps(fiz1,tz);
 394
 395             fjx1             = _mm_add_ps(fjx1,tx);
 396             fjy1             = _mm_add_ps(fjy1,ty);
 397             fjz1             = _mm_add_ps(fjz1,tz);
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             /* COULOMB ELECTROSTATICS */
 404             velec            = _mm_mul_ps(qq12,rinv12);
 405             felec            = _mm_mul_ps(velec,rinvsq12);
 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             /* Calculate temporary vectorial force */
 413             tx               = _mm_mul_ps(fscal,dx12);
 414             ty               = _mm_mul_ps(fscal,dy12);
 415             tz               = _mm_mul_ps(fscal,dz12);
 416
 417             /* Update vectorial force */
 418             fix1             = _mm_add_ps(fix1,tx);
 419             fiy1             = _mm_add_ps(fiy1,ty);
 420             fiz1             = _mm_add_ps(fiz1,tz);
 421
 422             fjx2             = _mm_add_ps(fjx2,tx);
 423             fjy2             = _mm_add_ps(fjy2,ty);
 424             fjz2             = _mm_add_ps(fjz2,tz);
 425
 426             /**************************
 427              * CALCULATE INTERACTIONS *
 428              **************************/
 429
 430             /* COULOMB ELECTROSTATICS */
 431             velec            = _mm_mul_ps(qq20,rinv20);
 432             felec            = _mm_mul_ps(velec,rinvsq20);
 433
 434             /* Update potential sum for this i atom from the interaction with this j atom. */
 435             velecsum         = _mm_add_ps(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             /* Calculate temporary vectorial force */
 440             tx               = _mm_mul_ps(fscal,dx20);
 441             ty               = _mm_mul_ps(fscal,dy20);
 442             tz               = _mm_mul_ps(fscal,dz20);
 443
 444             /* Update vectorial force */
 445             fix2             = _mm_add_ps(fix2,tx);
 446             fiy2             = _mm_add_ps(fiy2,ty);
 447             fiz2             = _mm_add_ps(fiz2,tz);
 448
 449             fjx0             = _mm_add_ps(fjx0,tx);
 450             fjy0             = _mm_add_ps(fjy0,ty);
 451             fjz0             = _mm_add_ps(fjz0,tz);
 452
 453             /**************************
 454              * CALCULATE INTERACTIONS *
 455              **************************/
 456
 457             /* COULOMB ELECTROSTATICS */
 458             velec            = _mm_mul_ps(qq21,rinv21);
 459             felec            = _mm_mul_ps(velec,rinvsq21);
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velecsum         = _mm_add_ps(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466             /* Calculate temporary vectorial force */
 467             tx               = _mm_mul_ps(fscal,dx21);
 468             ty               = _mm_mul_ps(fscal,dy21);
 469             tz               = _mm_mul_ps(fscal,dz21);
 470
 471             /* Update vectorial force */
 472             fix2             = _mm_add_ps(fix2,tx);
 473             fiy2             = _mm_add_ps(fiy2,ty);
 474             fiz2             = _mm_add_ps(fiz2,tz);
 475
 476             fjx1             = _mm_add_ps(fjx1,tx);
 477             fjy1             = _mm_add_ps(fjy1,ty);
 478             fjz1             = _mm_add_ps(fjz1,tz);
 479
 480             /**************************
 481              * CALCULATE INTERACTIONS *
 482              **************************/
 483
 484             /* COULOMB ELECTROSTATICS */
 485             velec            = _mm_mul_ps(qq22,rinv22);
 486             felec            = _mm_mul_ps(velec,rinvsq22);
 487
 488             /* Update potential sum for this i atom from the interaction with this j atom. */
 489             velecsum         = _mm_add_ps(velecsum,velec);
 490
 491             fscal            = felec;
 492
 493             /* Calculate temporary vectorial force */
 494             tx               = _mm_mul_ps(fscal,dx22);
 495             ty               = _mm_mul_ps(fscal,dy22);
 496             tz               = _mm_mul_ps(fscal,dz22);
 497
 498             /* Update vectorial force */
 499             fix2             = _mm_add_ps(fix2,tx);
 500             fiy2             = _mm_add_ps(fiy2,ty);
 501             fiz2             = _mm_add_ps(fiz2,tz);
 502
 503             fjx2             = _mm_add_ps(fjx2,tx);
 504             fjy2             = _mm_add_ps(fjy2,ty);
 505             fjz2             = _mm_add_ps(fjz2,tz);
 506
 507             fjptrA             = f+j_coord_offsetA;
 508             fjptrB             = f+j_coord_offsetB;
 509             fjptrC             = f+j_coord_offsetC;
 510             fjptrD             = f+j_coord_offsetD;
 511
 512             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 513                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 514
 515             /* Inner loop uses 252 flops */
 516         }
 517
 518         if(jidx<j_index_end)
 519         {
 520
 521             /* Get j neighbor index, and coordinate index */
 522             jnrlistA         = jjnr[jidx];
 523             jnrlistB         = jjnr[jidx+1];
 524             jnrlistC         = jjnr[jidx+2];
 525             jnrlistD         = jjnr[jidx+3];
 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       = (jnrlistA>=0) ? jnrlistA : 0;
 532             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 533             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 534             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 535             j_coord_offsetA  = DIM*jnrA;
 536             j_coord_offsetB  = DIM*jnrB;
 537             j_coord_offsetC  = DIM*jnrC;
 538             j_coord_offsetD  = DIM*jnrD;
 539
 540             /* load j atom coordinates */
 541             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 542                                               x+j_coord_offsetC,x+j_coord_offsetD,
 543                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 544
 545             /* Calculate displacement vector */
 546             dx00             = _mm_sub_ps(ix0,jx0);
 547             dy00             = _mm_sub_ps(iy0,jy0);
 548             dz00             = _mm_sub_ps(iz0,jz0);
 549             dx01             = _mm_sub_ps(ix0,jx1);
 550             dy01             = _mm_sub_ps(iy0,jy1);
 551             dz01             = _mm_sub_ps(iz0,jz1);
 552             dx02             = _mm_sub_ps(ix0,jx2);
 553             dy02             = _mm_sub_ps(iy0,jy2);
 554             dz02             = _mm_sub_ps(iz0,jz2);
 555             dx10             = _mm_sub_ps(ix1,jx0);
 556             dy10             = _mm_sub_ps(iy1,jy0);
 557             dz10             = _mm_sub_ps(iz1,jz0);
 558             dx11             = _mm_sub_ps(ix1,jx1);
 559             dy11             = _mm_sub_ps(iy1,jy1);
 560             dz11             = _mm_sub_ps(iz1,jz1);
 561             dx12             = _mm_sub_ps(ix1,jx2);
 562             dy12             = _mm_sub_ps(iy1,jy2);
 563             dz12             = _mm_sub_ps(iz1,jz2);
 564             dx20             = _mm_sub_ps(ix2,jx0);
 565             dy20             = _mm_sub_ps(iy2,jy0);
 566             dz20             = _mm_sub_ps(iz2,jz0);
 567             dx21             = _mm_sub_ps(ix2,jx1);
 568             dy21             = _mm_sub_ps(iy2,jy1);
 569             dz21             = _mm_sub_ps(iz2,jz1);
 570             dx22             = _mm_sub_ps(ix2,jx2);
 571             dy22             = _mm_sub_ps(iy2,jy2);
 572             dz22             = _mm_sub_ps(iz2,jz2);
 573
 574             /* Calculate squared distance and things based on it */
 575             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 576             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 577             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 578             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 579             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 580             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 581             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 582             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 583             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 584
 585             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 586             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 587             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 588             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 589             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 590             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 591             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 592             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 593             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 594
 595             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 596             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 597             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 598             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 599             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 600             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 601             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 602             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 603             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 604
 605             fjx0             = _mm_setzero_ps();
 606             fjy0             = _mm_setzero_ps();
 607             fjz0             = _mm_setzero_ps();
 608             fjx1             = _mm_setzero_ps();
 609             fjy1             = _mm_setzero_ps();
 610             fjz1             = _mm_setzero_ps();
 611             fjx2             = _mm_setzero_ps();
 612             fjy2             = _mm_setzero_ps();
 613             fjz2             = _mm_setzero_ps();
 614
 615             /**************************
 616              * CALCULATE INTERACTIONS *
 617              **************************/
 618
 619             /* COULOMB ELECTROSTATICS */
 620             velec            = _mm_mul_ps(qq00,rinv00);
 621             felec            = _mm_mul_ps(velec,rinvsq00);
 622
 623             /* Update potential sum for this i atom from the interaction with this j atom. */
 624             velec            = _mm_andnot_ps(dummy_mask,velec);
 625             velecsum         = _mm_add_ps(velecsum,velec);
 626
 627             fscal            = felec;
 628
 629             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 630
 631             /* Calculate temporary vectorial force */
 632             tx               = _mm_mul_ps(fscal,dx00);
 633             ty               = _mm_mul_ps(fscal,dy00);
 634             tz               = _mm_mul_ps(fscal,dz00);
 635
 636             /* Update vectorial force */
 637             fix0             = _mm_add_ps(fix0,tx);
 638             fiy0             = _mm_add_ps(fiy0,ty);
 639             fiz0             = _mm_add_ps(fiz0,tz);
 640
 641             fjx0             = _mm_add_ps(fjx0,tx);
 642             fjy0             = _mm_add_ps(fjy0,ty);
 643             fjz0             = _mm_add_ps(fjz0,tz);
 644
 645             /**************************
 646              * CALCULATE INTERACTIONS *
 647              **************************/
 648
 649             /* COULOMB ELECTROSTATICS */
 650             velec            = _mm_mul_ps(qq01,rinv01);
 651             felec            = _mm_mul_ps(velec,rinvsq01);
 652
 653             /* Update potential sum for this i atom from the interaction with this j atom. */
 654             velec            = _mm_andnot_ps(dummy_mask,velec);
 655             velecsum         = _mm_add_ps(velecsum,velec);
 656
 657             fscal            = felec;
 658
 659             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 660
 661             /* Calculate temporary vectorial force */
 662             tx               = _mm_mul_ps(fscal,dx01);
 663             ty               = _mm_mul_ps(fscal,dy01);
 664             tz               = _mm_mul_ps(fscal,dz01);
 665
 666             /* Update vectorial force */
 667             fix0             = _mm_add_ps(fix0,tx);
 668             fiy0             = _mm_add_ps(fiy0,ty);
 669             fiz0             = _mm_add_ps(fiz0,tz);
 670
 671             fjx1             = _mm_add_ps(fjx1,tx);
 672             fjy1             = _mm_add_ps(fjy1,ty);
 673             fjz1             = _mm_add_ps(fjz1,tz);
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             /* COULOMB ELECTROSTATICS */
 680             velec            = _mm_mul_ps(qq02,rinv02);
 681             felec            = _mm_mul_ps(velec,rinvsq02);
 682
 683             /* Update potential sum for this i atom from the interaction with this j atom. */
 684             velec            = _mm_andnot_ps(dummy_mask,velec);
 685             velecsum         = _mm_add_ps(velecsum,velec);
 686
 687             fscal            = felec;
 688
 689             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 690
 691             /* Calculate temporary vectorial force */
 692             tx               = _mm_mul_ps(fscal,dx02);
 693             ty               = _mm_mul_ps(fscal,dy02);
 694             tz               = _mm_mul_ps(fscal,dz02);
 695
 696             /* Update vectorial force */
 697             fix0             = _mm_add_ps(fix0,tx);
 698             fiy0             = _mm_add_ps(fiy0,ty);
 699             fiz0             = _mm_add_ps(fiz0,tz);
 700
 701             fjx2             = _mm_add_ps(fjx2,tx);
 702             fjy2             = _mm_add_ps(fjy2,ty);
 703             fjz2             = _mm_add_ps(fjz2,tz);
 704
 705             /**************************
 706              * CALCULATE INTERACTIONS *
 707              **************************/
 708
 709             /* COULOMB ELECTROSTATICS */
 710             velec            = _mm_mul_ps(qq10,rinv10);
 711             felec            = _mm_mul_ps(velec,rinvsq10);
 712
 713             /* Update potential sum for this i atom from the interaction with this j atom. */
 714             velec            = _mm_andnot_ps(dummy_mask,velec);
 715             velecsum         = _mm_add_ps(velecsum,velec);
 716
 717             fscal            = felec;
 718
 719             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 720
 721             /* Calculate temporary vectorial force */
 722             tx               = _mm_mul_ps(fscal,dx10);
 723             ty               = _mm_mul_ps(fscal,dy10);
 724             tz               = _mm_mul_ps(fscal,dz10);
 725
 726             /* Update vectorial force */
 727             fix1             = _mm_add_ps(fix1,tx);
 728             fiy1             = _mm_add_ps(fiy1,ty);
 729             fiz1             = _mm_add_ps(fiz1,tz);
 730
 731             fjx0             = _mm_add_ps(fjx0,tx);
 732             fjy0             = _mm_add_ps(fjy0,ty);
 733             fjz0             = _mm_add_ps(fjz0,tz);
 734
 735             /**************************
 736              * CALCULATE INTERACTIONS *
 737              **************************/
 738
 739             /* COULOMB ELECTROSTATICS */
 740             velec            = _mm_mul_ps(qq11,rinv11);
 741             felec            = _mm_mul_ps(velec,rinvsq11);
 742
 743             /* Update potential sum for this i atom from the interaction with this j atom. */
 744             velec            = _mm_andnot_ps(dummy_mask,velec);
 745             velecsum         = _mm_add_ps(velecsum,velec);
 746
 747             fscal            = felec;
 748
 749             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 750
 751             /* Calculate temporary vectorial force */
 752             tx               = _mm_mul_ps(fscal,dx11);
 753             ty               = _mm_mul_ps(fscal,dy11);
 754             tz               = _mm_mul_ps(fscal,dz11);
 755
 756             /* Update vectorial force */
 757             fix1             = _mm_add_ps(fix1,tx);
 758             fiy1             = _mm_add_ps(fiy1,ty);
 759             fiz1             = _mm_add_ps(fiz1,tz);
 760
 761             fjx1             = _mm_add_ps(fjx1,tx);
 762             fjy1             = _mm_add_ps(fjy1,ty);
 763             fjz1             = _mm_add_ps(fjz1,tz);
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             /* COULOMB ELECTROSTATICS */
 770             velec            = _mm_mul_ps(qq12,rinv12);
 771             felec            = _mm_mul_ps(velec,rinvsq12);
 772
 773             /* Update potential sum for this i atom from the interaction with this j atom. */
 774             velec            = _mm_andnot_ps(dummy_mask,velec);
 775             velecsum         = _mm_add_ps(velecsum,velec);
 776
 777             fscal            = felec;
 778
 779             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 780
 781             /* Calculate temporary vectorial force */
 782             tx               = _mm_mul_ps(fscal,dx12);
 783             ty               = _mm_mul_ps(fscal,dy12);
 784             tz               = _mm_mul_ps(fscal,dz12);
 785
 786             /* Update vectorial force */
 787             fix1             = _mm_add_ps(fix1,tx);
 788             fiy1             = _mm_add_ps(fiy1,ty);
 789             fiz1             = _mm_add_ps(fiz1,tz);
 790
 791             fjx2             = _mm_add_ps(fjx2,tx);
 792             fjy2             = _mm_add_ps(fjy2,ty);
 793             fjz2             = _mm_add_ps(fjz2,tz);
 794
 795             /**************************
 796              * CALCULATE INTERACTIONS *
 797              **************************/
 798
 799             /* COULOMB ELECTROSTATICS */
 800             velec            = _mm_mul_ps(qq20,rinv20);
 801             felec            = _mm_mul_ps(velec,rinvsq20);
 802
 803             /* Update potential sum for this i atom from the interaction with this j atom. */
 804             velec            = _mm_andnot_ps(dummy_mask,velec);
 805             velecsum         = _mm_add_ps(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 810
 811             /* Calculate temporary vectorial force */
 812             tx               = _mm_mul_ps(fscal,dx20);
 813             ty               = _mm_mul_ps(fscal,dy20);
 814             tz               = _mm_mul_ps(fscal,dz20);
 815
 816             /* Update vectorial force */
 817             fix2             = _mm_add_ps(fix2,tx);
 818             fiy2             = _mm_add_ps(fiy2,ty);
 819             fiz2             = _mm_add_ps(fiz2,tz);
 820
 821             fjx0             = _mm_add_ps(fjx0,tx);
 822             fjy0             = _mm_add_ps(fjy0,ty);
 823             fjz0             = _mm_add_ps(fjz0,tz);
 824
 825             /**************************
 826              * CALCULATE INTERACTIONS *
 827              **************************/
 828
 829             /* COULOMB ELECTROSTATICS */
 830             velec            = _mm_mul_ps(qq21,rinv21);
 831             felec            = _mm_mul_ps(velec,rinvsq21);
 832
 833             /* Update potential sum for this i atom from the interaction with this j atom. */
 834             velec            = _mm_andnot_ps(dummy_mask,velec);
 835             velecsum         = _mm_add_ps(velecsum,velec);
 836
 837             fscal            = felec;
 838
 839             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 840
 841             /* Calculate temporary vectorial force */
 842             tx               = _mm_mul_ps(fscal,dx21);
 843             ty               = _mm_mul_ps(fscal,dy21);
 844             tz               = _mm_mul_ps(fscal,dz21);
 845
 846             /* Update vectorial force */
 847             fix2             = _mm_add_ps(fix2,tx);
 848             fiy2             = _mm_add_ps(fiy2,ty);
 849             fiz2             = _mm_add_ps(fiz2,tz);
 850
 851             fjx1             = _mm_add_ps(fjx1,tx);
 852             fjy1             = _mm_add_ps(fjy1,ty);
 853             fjz1             = _mm_add_ps(fjz1,tz);
 854
 855             /**************************
 856              * CALCULATE INTERACTIONS *
 857              **************************/
 858
 859             /* COULOMB ELECTROSTATICS */
 860             velec            = _mm_mul_ps(qq22,rinv22);
 861             felec            = _mm_mul_ps(velec,rinvsq22);
 862
 863             /* Update potential sum for this i atom from the interaction with this j atom. */
 864             velec            = _mm_andnot_ps(dummy_mask,velec);
 865             velecsum         = _mm_add_ps(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 870
 871             /* Calculate temporary vectorial force */
 872             tx               = _mm_mul_ps(fscal,dx22);
 873             ty               = _mm_mul_ps(fscal,dy22);
 874             tz               = _mm_mul_ps(fscal,dz22);
 875
 876             /* Update vectorial force */
 877             fix2             = _mm_add_ps(fix2,tx);
 878             fiy2             = _mm_add_ps(fiy2,ty);
 879             fiz2             = _mm_add_ps(fiz2,tz);
 880
 881             fjx2             = _mm_add_ps(fjx2,tx);
 882             fjy2             = _mm_add_ps(fjy2,ty);
 883             fjz2             = _mm_add_ps(fjz2,tz);
 884
 885             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 886             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 887             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 888             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 889
 890             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 891                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 892
 893             /* Inner loop uses 252 flops */
 894         }
 895
 896         /* End of innermost loop */
 897
 898         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 899                                               f+i_coord_offset,fshift+i_shift_offset);
 900
 901         ggid                        = gid[iidx];
 902         /* Update potential energies */
 903         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 904
 905         /* Increment number of inner iterations */
 906         inneriter                  += j_index_end - j_index_start;
 907
 908         /* Outer loop uses 19 flops */
 909     }
 910
 911     /* Increment number of outer iterations */
 912     outeriter        += nri;
 913
 914     /* Update outer/inner flops */
 915
 916     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*252);
 917 }
 918 /*
 919  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_sse4_1_single
 920  * Electrostatics interaction: Coulomb
 921  * VdW interaction:            None
 922  * Geometry:                   Water3-Water3
 923  * Calculate force/pot:        Force
 924  */
 925 void
 926 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_sse4_1_single
 927                     (t_nblist * gmx_restrict                nlist,
 928                      rvec * gmx_restrict                    xx,
 929                      rvec * gmx_restrict                    ff,
 930                      t_forcerec * gmx_restrict              fr,
 931                      t_mdatoms * gmx_restrict               mdatoms,
 932                      nb_kernel_data_t * gmx_restrict        kernel_data,
 933                      t_nrnb * gmx_restrict                  nrnb)
 934 {
 935     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 936      * just 0 for non-waters.
 937      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 938      * jnr indices corresponding to data put in the four positions in the SIMD register.
 939      */
 940     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 941     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 942     int              jnrA,jnrB,jnrC,jnrD;
 943     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 944     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 945     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 946     real             rcutoff_scalar;
 947     real             *shiftvec,*fshift,*x,*f;
 948     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 949     real             scratch[4*DIM];
 950     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 951     int              vdwioffset0;
 952     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 953     int              vdwioffset1;
 954     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 955     int              vdwioffset2;
 956     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 957     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 958     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 959     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 960     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 961     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 962     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 963     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 964     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 965     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 966     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 967     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 968     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 969     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 970     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 971     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 972     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 973     real             *charge;
 974     __m128           dummy_mask,cutoff_mask;
 975     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 976     __m128           one     = _mm_set1_ps(1.0);
 977     __m128           two     = _mm_set1_ps(2.0);
 978     x                = xx[0];
 979     f                = ff[0];
 980
 981     nri              = nlist->nri;
 982     iinr             = nlist->iinr;
 983     jindex           = nlist->jindex;
 984     jjnr             = nlist->jjnr;
 985     shiftidx         = nlist->shift;
 986     gid              = nlist->gid;
 987     shiftvec         = fr->shift_vec[0];
 988     fshift           = fr->fshift[0];
 989     facel            = _mm_set1_ps(fr->epsfac);
 990     charge           = mdatoms->chargeA;
 991
 992     /* Setup water-specific parameters */
 993     inr              = nlist->iinr[0];
 994     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 995     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 996     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 997
 998     jq0              = _mm_set1_ps(charge[inr+0]);
 999     jq1              = _mm_set1_ps(charge[inr+1]);
1000     jq2              = _mm_set1_ps(charge[inr+2]);
1001     qq00             = _mm_mul_ps(iq0,jq0);
1002     qq01             = _mm_mul_ps(iq0,jq1);
1003     qq02             = _mm_mul_ps(iq0,jq2);
1004     qq10             = _mm_mul_ps(iq1,jq0);
1005     qq11             = _mm_mul_ps(iq1,jq1);
1006     qq12             = _mm_mul_ps(iq1,jq2);
1007     qq20             = _mm_mul_ps(iq2,jq0);
1008     qq21             = _mm_mul_ps(iq2,jq1);
1009     qq22             = _mm_mul_ps(iq2,jq2);
1010
1011     /* Avoid stupid compiler warnings */
1012     jnrA = jnrB = jnrC = jnrD = 0;
1013     j_coord_offsetA = 0;
1014     j_coord_offsetB = 0;
1015     j_coord_offsetC = 0;
1016     j_coord_offsetD = 0;
1017
1018     outeriter        = 0;
1019     inneriter        = 0;
1020
1021     for(iidx=0;iidx<4*DIM;iidx++)
1022     {
1023         scratch[iidx] = 0.0;
1024     }
1025
1026     /* Start outer loop over neighborlists */
1027     for(iidx=0; iidx<nri; iidx++)
1028     {
1029         /* Load shift vector for this list */
1030         i_shift_offset   = DIM*shiftidx[iidx];
1031
1032         /* Load limits for loop over neighbors */
1033         j_index_start    = jindex[iidx];
1034         j_index_end      = jindex[iidx+1];
1035
1036         /* Get outer coordinate index */
1037         inr              = iinr[iidx];
1038         i_coord_offset   = DIM*inr;
1039
1040         /* Load i particle coords and add shift vector */
1041         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1042                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1043
1044         fix0             = _mm_setzero_ps();
1045         fiy0             = _mm_setzero_ps();
1046         fiz0             = _mm_setzero_ps();
1047         fix1             = _mm_setzero_ps();
1048         fiy1             = _mm_setzero_ps();
1049         fiz1             = _mm_setzero_ps();
1050         fix2             = _mm_setzero_ps();
1051         fiy2             = _mm_setzero_ps();
1052         fiz2             = _mm_setzero_ps();
1053
1054         /* Start inner kernel loop */
1055         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1056         {
1057
1058             /* Get j neighbor index, and coordinate index */
1059             jnrA             = jjnr[jidx];
1060             jnrB             = jjnr[jidx+1];
1061             jnrC             = jjnr[jidx+2];
1062             jnrD             = jjnr[jidx+3];
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,x+j_coord_offsetB,
1070                                               x+j_coord_offsetC,x+j_coord_offsetD,
1071                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1072
1073             /* Calculate displacement vector */
1074             dx00             = _mm_sub_ps(ix0,jx0);
1075             dy00             = _mm_sub_ps(iy0,jy0);
1076             dz00             = _mm_sub_ps(iz0,jz0);
1077             dx01             = _mm_sub_ps(ix0,jx1);
1078             dy01             = _mm_sub_ps(iy0,jy1);
1079             dz01             = _mm_sub_ps(iz0,jz1);
1080             dx02             = _mm_sub_ps(ix0,jx2);
1081             dy02             = _mm_sub_ps(iy0,jy2);
1082             dz02             = _mm_sub_ps(iz0,jz2);
1083             dx10             = _mm_sub_ps(ix1,jx0);
1084             dy10             = _mm_sub_ps(iy1,jy0);
1085             dz10             = _mm_sub_ps(iz1,jz0);
1086             dx11             = _mm_sub_ps(ix1,jx1);
1087             dy11             = _mm_sub_ps(iy1,jy1);
1088             dz11             = _mm_sub_ps(iz1,jz1);
1089             dx12             = _mm_sub_ps(ix1,jx2);
1090             dy12             = _mm_sub_ps(iy1,jy2);
1091             dz12             = _mm_sub_ps(iz1,jz2);
1092             dx20             = _mm_sub_ps(ix2,jx0);
1093             dy20             = _mm_sub_ps(iy2,jy0);
1094             dz20             = _mm_sub_ps(iz2,jz0);
1095             dx21             = _mm_sub_ps(ix2,jx1);
1096             dy21             = _mm_sub_ps(iy2,jy1);
1097             dz21             = _mm_sub_ps(iz2,jz1);
1098             dx22             = _mm_sub_ps(ix2,jx2);
1099             dy22             = _mm_sub_ps(iy2,jy2);
1100             dz22             = _mm_sub_ps(iz2,jz2);
1101
1102             /* Calculate squared distance and things based on it */
1103             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1104             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1105             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1106             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1107             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1108             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1109             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1110             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1111             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1112
1113             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1114             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1115             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1116             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1117             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1118             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1119             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1120             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1121             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1122
1123             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1124             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1125             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1126             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1127             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1128             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1129             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1130             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1131             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1132
1133             fjx0             = _mm_setzero_ps();
1134             fjy0             = _mm_setzero_ps();
1135             fjz0             = _mm_setzero_ps();
1136             fjx1             = _mm_setzero_ps();
1137             fjy1             = _mm_setzero_ps();
1138             fjz1             = _mm_setzero_ps();
1139             fjx2             = _mm_setzero_ps();
1140             fjy2             = _mm_setzero_ps();
1141             fjz2             = _mm_setzero_ps();
1142
1143             /**************************
1144              * CALCULATE INTERACTIONS *
1145              **************************/
1146
1147             /* COULOMB ELECTROSTATICS */
1148             velec            = _mm_mul_ps(qq00,rinv00);
1149             felec            = _mm_mul_ps(velec,rinvsq00);
1150
1151             fscal            = felec;
1152
1153             /* Calculate temporary vectorial force */
1154             tx               = _mm_mul_ps(fscal,dx00);
1155             ty               = _mm_mul_ps(fscal,dy00);
1156             tz               = _mm_mul_ps(fscal,dz00);
1157
1158             /* Update vectorial force */
1159             fix0             = _mm_add_ps(fix0,tx);
1160             fiy0             = _mm_add_ps(fiy0,ty);
1161             fiz0             = _mm_add_ps(fiz0,tz);
1162
1163             fjx0             = _mm_add_ps(fjx0,tx);
1164             fjy0             = _mm_add_ps(fjy0,ty);
1165             fjz0             = _mm_add_ps(fjz0,tz);
1166
1167             /**************************
1168              * CALCULATE INTERACTIONS *
1169              **************************/
1170
1171             /* COULOMB ELECTROSTATICS */
1172             velec            = _mm_mul_ps(qq01,rinv01);
1173             felec            = _mm_mul_ps(velec,rinvsq01);
1174
1175             fscal            = felec;
1176
1177             /* Calculate temporary vectorial force */
1178             tx               = _mm_mul_ps(fscal,dx01);
1179             ty               = _mm_mul_ps(fscal,dy01);
1180             tz               = _mm_mul_ps(fscal,dz01);
1181
1182             /* Update vectorial force */
1183             fix0             = _mm_add_ps(fix0,tx);
1184             fiy0             = _mm_add_ps(fiy0,ty);
1185             fiz0             = _mm_add_ps(fiz0,tz);
1186
1187             fjx1             = _mm_add_ps(fjx1,tx);
1188             fjy1             = _mm_add_ps(fjy1,ty);
1189             fjz1             = _mm_add_ps(fjz1,tz);
1190
1191             /**************************
1192              * CALCULATE INTERACTIONS *
1193              **************************/
1194
1195             /* COULOMB ELECTROSTATICS */
1196             velec            = _mm_mul_ps(qq02,rinv02);
1197             felec            = _mm_mul_ps(velec,rinvsq02);
1198
1199             fscal            = felec;
1200
1201             /* Calculate temporary vectorial force */
1202             tx               = _mm_mul_ps(fscal,dx02);
1203             ty               = _mm_mul_ps(fscal,dy02);
1204             tz               = _mm_mul_ps(fscal,dz02);
1205
1206             /* Update vectorial force */
1207             fix0             = _mm_add_ps(fix0,tx);
1208             fiy0             = _mm_add_ps(fiy0,ty);
1209             fiz0             = _mm_add_ps(fiz0,tz);
1210
1211             fjx2             = _mm_add_ps(fjx2,tx);
1212             fjy2             = _mm_add_ps(fjy2,ty);
1213             fjz2             = _mm_add_ps(fjz2,tz);
1214
1215             /**************************
1216              * CALCULATE INTERACTIONS *
1217              **************************/
1218
1219             /* COULOMB ELECTROSTATICS */
1220             velec            = _mm_mul_ps(qq10,rinv10);
1221             felec            = _mm_mul_ps(velec,rinvsq10);
1222
1223             fscal            = felec;
1224
1225             /* Calculate temporary vectorial force */
1226             tx               = _mm_mul_ps(fscal,dx10);
1227             ty               = _mm_mul_ps(fscal,dy10);
1228             tz               = _mm_mul_ps(fscal,dz10);
1229
1230             /* Update vectorial force */
1231             fix1             = _mm_add_ps(fix1,tx);
1232             fiy1             = _mm_add_ps(fiy1,ty);
1233             fiz1             = _mm_add_ps(fiz1,tz);
1234
1235             fjx0             = _mm_add_ps(fjx0,tx);
1236             fjy0             = _mm_add_ps(fjy0,ty);
1237             fjz0             = _mm_add_ps(fjz0,tz);
1238
1239             /**************************
1240              * CALCULATE INTERACTIONS *
1241              **************************/
1242
1243             /* COULOMB ELECTROSTATICS */
1244             velec            = _mm_mul_ps(qq11,rinv11);
1245             felec            = _mm_mul_ps(velec,rinvsq11);
1246
1247             fscal            = felec;
1248
1249             /* Calculate temporary vectorial force */
1250             tx               = _mm_mul_ps(fscal,dx11);
1251             ty               = _mm_mul_ps(fscal,dy11);
1252             tz               = _mm_mul_ps(fscal,dz11);
1253
1254             /* Update vectorial force */
1255             fix1             = _mm_add_ps(fix1,tx);
1256             fiy1             = _mm_add_ps(fiy1,ty);
1257             fiz1             = _mm_add_ps(fiz1,tz);
1258
1259             fjx1             = _mm_add_ps(fjx1,tx);
1260             fjy1             = _mm_add_ps(fjy1,ty);
1261             fjz1             = _mm_add_ps(fjz1,tz);
1262
1263             /**************************
1264              * CALCULATE INTERACTIONS *
1265              **************************/
1266
1267             /* COULOMB ELECTROSTATICS */
1268             velec            = _mm_mul_ps(qq12,rinv12);
1269             felec            = _mm_mul_ps(velec,rinvsq12);
1270
1271             fscal            = felec;
1272
1273             /* Calculate temporary vectorial force */
1274             tx               = _mm_mul_ps(fscal,dx12);
1275             ty               = _mm_mul_ps(fscal,dy12);
1276             tz               = _mm_mul_ps(fscal,dz12);
1277
1278             /* Update vectorial force */
1279             fix1             = _mm_add_ps(fix1,tx);
1280             fiy1             = _mm_add_ps(fiy1,ty);
1281             fiz1             = _mm_add_ps(fiz1,tz);
1282
1283             fjx2             = _mm_add_ps(fjx2,tx);
1284             fjy2             = _mm_add_ps(fjy2,ty);
1285             fjz2             = _mm_add_ps(fjz2,tz);
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* COULOMB ELECTROSTATICS */
1292             velec            = _mm_mul_ps(qq20,rinv20);
1293             felec            = _mm_mul_ps(velec,rinvsq20);
1294
1295             fscal            = felec;
1296
1297             /* Calculate temporary vectorial force */
1298             tx               = _mm_mul_ps(fscal,dx20);
1299             ty               = _mm_mul_ps(fscal,dy20);
1300             tz               = _mm_mul_ps(fscal,dz20);
1301
1302             /* Update vectorial force */
1303             fix2             = _mm_add_ps(fix2,tx);
1304             fiy2             = _mm_add_ps(fiy2,ty);
1305             fiz2             = _mm_add_ps(fiz2,tz);
1306
1307             fjx0             = _mm_add_ps(fjx0,tx);
1308             fjy0             = _mm_add_ps(fjy0,ty);
1309             fjz0             = _mm_add_ps(fjz0,tz);
1310
1311             /**************************
1312              * CALCULATE INTERACTIONS *
1313              **************************/
1314
1315             /* COULOMB ELECTROSTATICS */
1316             velec            = _mm_mul_ps(qq21,rinv21);
1317             felec            = _mm_mul_ps(velec,rinvsq21);
1318
1319             fscal            = felec;
1320
1321             /* Calculate temporary vectorial force */
1322             tx               = _mm_mul_ps(fscal,dx21);
1323             ty               = _mm_mul_ps(fscal,dy21);
1324             tz               = _mm_mul_ps(fscal,dz21);
1325
1326             /* Update vectorial force */
1327             fix2             = _mm_add_ps(fix2,tx);
1328             fiy2             = _mm_add_ps(fiy2,ty);
1329             fiz2             = _mm_add_ps(fiz2,tz);
1330
1331             fjx1             = _mm_add_ps(fjx1,tx);
1332             fjy1             = _mm_add_ps(fjy1,ty);
1333             fjz1             = _mm_add_ps(fjz1,tz);
1334
1335             /**************************
1336              * CALCULATE INTERACTIONS *
1337              **************************/
1338
1339             /* COULOMB ELECTROSTATICS */
1340             velec            = _mm_mul_ps(qq22,rinv22);
1341             felec            = _mm_mul_ps(velec,rinvsq22);
1342
1343             fscal            = felec;
1344
1345             /* Calculate temporary vectorial force */
1346             tx               = _mm_mul_ps(fscal,dx22);
1347             ty               = _mm_mul_ps(fscal,dy22);
1348             tz               = _mm_mul_ps(fscal,dz22);
1349
1350             /* Update vectorial force */
1351             fix2             = _mm_add_ps(fix2,tx);
1352             fiy2             = _mm_add_ps(fiy2,ty);
1353             fiz2             = _mm_add_ps(fiz2,tz);
1354
1355             fjx2             = _mm_add_ps(fjx2,tx);
1356             fjy2             = _mm_add_ps(fjy2,ty);
1357             fjz2             = _mm_add_ps(fjz2,tz);
1358
1359             fjptrA             = f+j_coord_offsetA;
1360             fjptrB             = f+j_coord_offsetB;
1361             fjptrC             = f+j_coord_offsetC;
1362             fjptrD             = f+j_coord_offsetD;
1363
1364             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1365                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1366
1367             /* Inner loop uses 243 flops */
1368         }
1369
1370         if(jidx<j_index_end)
1371         {
1372
1373             /* Get j neighbor index, and coordinate index */
1374             jnrlistA         = jjnr[jidx];
1375             jnrlistB         = jjnr[jidx+1];
1376             jnrlistC         = jjnr[jidx+2];
1377             jnrlistD         = jjnr[jidx+3];
1378             /* Sign of each element will be negative for non-real atoms.
1379              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1380              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1381              */
1382             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1383             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1384             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1385             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1386             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1387             j_coord_offsetA  = DIM*jnrA;
1388             j_coord_offsetB  = DIM*jnrB;
1389             j_coord_offsetC  = DIM*jnrC;
1390             j_coord_offsetD  = DIM*jnrD;
1391
1392             /* load j atom coordinates */
1393             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1394                                               x+j_coord_offsetC,x+j_coord_offsetD,
1395                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1396
1397             /* Calculate displacement vector */
1398             dx00             = _mm_sub_ps(ix0,jx0);
1399             dy00             = _mm_sub_ps(iy0,jy0);
1400             dz00             = _mm_sub_ps(iz0,jz0);
1401             dx01             = _mm_sub_ps(ix0,jx1);
1402             dy01             = _mm_sub_ps(iy0,jy1);
1403             dz01             = _mm_sub_ps(iz0,jz1);
1404             dx02             = _mm_sub_ps(ix0,jx2);
1405             dy02             = _mm_sub_ps(iy0,jy2);
1406             dz02             = _mm_sub_ps(iz0,jz2);
1407             dx10             = _mm_sub_ps(ix1,jx0);
1408             dy10             = _mm_sub_ps(iy1,jy0);
1409             dz10             = _mm_sub_ps(iz1,jz0);
1410             dx11             = _mm_sub_ps(ix1,jx1);
1411             dy11             = _mm_sub_ps(iy1,jy1);
1412             dz11             = _mm_sub_ps(iz1,jz1);
1413             dx12             = _mm_sub_ps(ix1,jx2);
1414             dy12             = _mm_sub_ps(iy1,jy2);
1415             dz12             = _mm_sub_ps(iz1,jz2);
1416             dx20             = _mm_sub_ps(ix2,jx0);
1417             dy20             = _mm_sub_ps(iy2,jy0);
1418             dz20             = _mm_sub_ps(iz2,jz0);
1419             dx21             = _mm_sub_ps(ix2,jx1);
1420             dy21             = _mm_sub_ps(iy2,jy1);
1421             dz21             = _mm_sub_ps(iz2,jz1);
1422             dx22             = _mm_sub_ps(ix2,jx2);
1423             dy22             = _mm_sub_ps(iy2,jy2);
1424             dz22             = _mm_sub_ps(iz2,jz2);
1425
1426             /* Calculate squared distance and things based on it */
1427             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1428             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1429             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1430             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1431             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1432             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1433             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1434             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1435             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1436
1437             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1438             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1439             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1440             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1441             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1442             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1443             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1444             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1445             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1446
1447             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1448             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1449             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1450             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1451             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1452             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1453             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1454             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1455             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1456
1457             fjx0             = _mm_setzero_ps();
1458             fjy0             = _mm_setzero_ps();
1459             fjz0             = _mm_setzero_ps();
1460             fjx1             = _mm_setzero_ps();
1461             fjy1             = _mm_setzero_ps();
1462             fjz1             = _mm_setzero_ps();
1463             fjx2             = _mm_setzero_ps();
1464             fjy2             = _mm_setzero_ps();
1465             fjz2             = _mm_setzero_ps();
1466
1467             /**************************
1468              * CALCULATE INTERACTIONS *
1469              **************************/
1470
1471             /* COULOMB ELECTROSTATICS */
1472             velec            = _mm_mul_ps(qq00,rinv00);
1473             felec            = _mm_mul_ps(velec,rinvsq00);
1474
1475             fscal            = felec;
1476
1477             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1478
1479             /* Calculate temporary vectorial force */
1480             tx               = _mm_mul_ps(fscal,dx00);
1481             ty               = _mm_mul_ps(fscal,dy00);
1482             tz               = _mm_mul_ps(fscal,dz00);
1483
1484             /* Update vectorial force */
1485             fix0             = _mm_add_ps(fix0,tx);
1486             fiy0             = _mm_add_ps(fiy0,ty);
1487             fiz0             = _mm_add_ps(fiz0,tz);
1488
1489             fjx0             = _mm_add_ps(fjx0,tx);
1490             fjy0             = _mm_add_ps(fjy0,ty);
1491             fjz0             = _mm_add_ps(fjz0,tz);
1492
1493             /**************************
1494              * CALCULATE INTERACTIONS *
1495              **************************/
1496
1497             /* COULOMB ELECTROSTATICS */
1498             velec            = _mm_mul_ps(qq01,rinv01);
1499             felec            = _mm_mul_ps(velec,rinvsq01);
1500
1501             fscal            = felec;
1502
1503             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1504
1505             /* Calculate temporary vectorial force */
1506             tx               = _mm_mul_ps(fscal,dx01);
1507             ty               = _mm_mul_ps(fscal,dy01);
1508             tz               = _mm_mul_ps(fscal,dz01);
1509
1510             /* Update vectorial force */
1511             fix0             = _mm_add_ps(fix0,tx);
1512             fiy0             = _mm_add_ps(fiy0,ty);
1513             fiz0             = _mm_add_ps(fiz0,tz);
1514
1515             fjx1             = _mm_add_ps(fjx1,tx);
1516             fjy1             = _mm_add_ps(fjy1,ty);
1517             fjz1             = _mm_add_ps(fjz1,tz);
1518
1519             /**************************
1520              * CALCULATE INTERACTIONS *
1521              **************************/
1522
1523             /* COULOMB ELECTROSTATICS */
1524             velec            = _mm_mul_ps(qq02,rinv02);
1525             felec            = _mm_mul_ps(velec,rinvsq02);
1526
1527             fscal            = felec;
1528
1529             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1530
1531             /* Calculate temporary vectorial force */
1532             tx               = _mm_mul_ps(fscal,dx02);
1533             ty               = _mm_mul_ps(fscal,dy02);
1534             tz               = _mm_mul_ps(fscal,dz02);
1535
1536             /* Update vectorial force */
1537             fix0             = _mm_add_ps(fix0,tx);
1538             fiy0             = _mm_add_ps(fiy0,ty);
1539             fiz0             = _mm_add_ps(fiz0,tz);
1540
1541             fjx2             = _mm_add_ps(fjx2,tx);
1542             fjy2             = _mm_add_ps(fjy2,ty);
1543             fjz2             = _mm_add_ps(fjz2,tz);
1544
1545             /**************************
1546              * CALCULATE INTERACTIONS *
1547              **************************/
1548
1549             /* COULOMB ELECTROSTATICS */
1550             velec            = _mm_mul_ps(qq10,rinv10);
1551             felec            = _mm_mul_ps(velec,rinvsq10);
1552
1553             fscal            = felec;
1554
1555             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1556
1557             /* Calculate temporary vectorial force */
1558             tx               = _mm_mul_ps(fscal,dx10);
1559             ty               = _mm_mul_ps(fscal,dy10);
1560             tz               = _mm_mul_ps(fscal,dz10);
1561
1562             /* Update vectorial force */
1563             fix1             = _mm_add_ps(fix1,tx);
1564             fiy1             = _mm_add_ps(fiy1,ty);
1565             fiz1             = _mm_add_ps(fiz1,tz);
1566
1567             fjx0             = _mm_add_ps(fjx0,tx);
1568             fjy0             = _mm_add_ps(fjy0,ty);
1569             fjz0             = _mm_add_ps(fjz0,tz);
1570
1571             /**************************
1572              * CALCULATE INTERACTIONS *
1573              **************************/
1574
1575             /* COULOMB ELECTROSTATICS */
1576             velec            = _mm_mul_ps(qq11,rinv11);
1577             felec            = _mm_mul_ps(velec,rinvsq11);
1578
1579             fscal            = felec;
1580
1581             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1582
1583             /* Calculate temporary vectorial force */
1584             tx               = _mm_mul_ps(fscal,dx11);
1585             ty               = _mm_mul_ps(fscal,dy11);
1586             tz               = _mm_mul_ps(fscal,dz11);
1587
1588             /* Update vectorial force */
1589             fix1             = _mm_add_ps(fix1,tx);
1590             fiy1             = _mm_add_ps(fiy1,ty);
1591             fiz1             = _mm_add_ps(fiz1,tz);
1592
1593             fjx1             = _mm_add_ps(fjx1,tx);
1594             fjy1             = _mm_add_ps(fjy1,ty);
1595             fjz1             = _mm_add_ps(fjz1,tz);
1596
1597             /**************************
1598              * CALCULATE INTERACTIONS *
1599              **************************/
1600
1601             /* COULOMB ELECTROSTATICS */
1602             velec            = _mm_mul_ps(qq12,rinv12);
1603             felec            = _mm_mul_ps(velec,rinvsq12);
1604
1605             fscal            = felec;
1606
1607             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1608
1609             /* Calculate temporary vectorial force */
1610             tx               = _mm_mul_ps(fscal,dx12);
1611             ty               = _mm_mul_ps(fscal,dy12);
1612             tz               = _mm_mul_ps(fscal,dz12);
1613
1614             /* Update vectorial force */
1615             fix1             = _mm_add_ps(fix1,tx);
1616             fiy1             = _mm_add_ps(fiy1,ty);
1617             fiz1             = _mm_add_ps(fiz1,tz);
1618
1619             fjx2             = _mm_add_ps(fjx2,tx);
1620             fjy2             = _mm_add_ps(fjy2,ty);
1621             fjz2             = _mm_add_ps(fjz2,tz);
1622
1623             /**************************
1624              * CALCULATE INTERACTIONS *
1625              **************************/
1626
1627             /* COULOMB ELECTROSTATICS */
1628             velec            = _mm_mul_ps(qq20,rinv20);
1629             felec            = _mm_mul_ps(velec,rinvsq20);
1630
1631             fscal            = felec;
1632
1633             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1634
1635             /* Calculate temporary vectorial force */
1636             tx               = _mm_mul_ps(fscal,dx20);
1637             ty               = _mm_mul_ps(fscal,dy20);
1638             tz               = _mm_mul_ps(fscal,dz20);
1639
1640             /* Update vectorial force */
1641             fix2             = _mm_add_ps(fix2,tx);
1642             fiy2             = _mm_add_ps(fiy2,ty);
1643             fiz2             = _mm_add_ps(fiz2,tz);
1644
1645             fjx0             = _mm_add_ps(fjx0,tx);
1646             fjy0             = _mm_add_ps(fjy0,ty);
1647             fjz0             = _mm_add_ps(fjz0,tz);
1648
1649             /**************************
1650              * CALCULATE INTERACTIONS *
1651              **************************/
1652
1653             /* COULOMB ELECTROSTATICS */
1654             velec            = _mm_mul_ps(qq21,rinv21);
1655             felec            = _mm_mul_ps(velec,rinvsq21);
1656
1657             fscal            = felec;
1658
1659             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1660
1661             /* Calculate temporary vectorial force */
1662             tx               = _mm_mul_ps(fscal,dx21);
1663             ty               = _mm_mul_ps(fscal,dy21);
1664             tz               = _mm_mul_ps(fscal,dz21);
1665
1666             /* Update vectorial force */
1667             fix2             = _mm_add_ps(fix2,tx);
1668             fiy2             = _mm_add_ps(fiy2,ty);
1669             fiz2             = _mm_add_ps(fiz2,tz);
1670
1671             fjx1             = _mm_add_ps(fjx1,tx);
1672             fjy1             = _mm_add_ps(fjy1,ty);
1673             fjz1             = _mm_add_ps(fjz1,tz);
1674
1675             /**************************
1676              * CALCULATE INTERACTIONS *
1677              **************************/
1678
1679             /* COULOMB ELECTROSTATICS */
1680             velec            = _mm_mul_ps(qq22,rinv22);
1681             felec            = _mm_mul_ps(velec,rinvsq22);
1682
1683             fscal            = felec;
1684
1685             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1686
1687             /* Calculate temporary vectorial force */
1688             tx               = _mm_mul_ps(fscal,dx22);
1689             ty               = _mm_mul_ps(fscal,dy22);
1690             tz               = _mm_mul_ps(fscal,dz22);
1691
1692             /* Update vectorial force */
1693             fix2             = _mm_add_ps(fix2,tx);
1694             fiy2             = _mm_add_ps(fiy2,ty);
1695             fiz2             = _mm_add_ps(fiz2,tz);
1696
1697             fjx2             = _mm_add_ps(fjx2,tx);
1698             fjy2             = _mm_add_ps(fjy2,ty);
1699             fjz2             = _mm_add_ps(fjz2,tz);
1700
1701             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1702             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1703             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1704             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1705
1706             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1707                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1708
1709             /* Inner loop uses 243 flops */
1710         }
1711
1712         /* End of innermost loop */
1713
1714         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1715                                               f+i_coord_offset,fshift+i_shift_offset);
1716
1717         /* Increment number of inner iterations */
1718         inneriter                  += j_index_end - j_index_start;
1719
1720         /* Outer loop uses 18 flops */
1721     }
1722
1723     /* Increment number of outer iterations */
1724     outeriter        += nri;
1725
1726     /* Update outer/inner flops */
1727
1728     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1729 }