src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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_VdwLJ_GeomW3W3_VF_sse2_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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              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     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     /* Setup water-specific parameters */
 120     inr              = nlist->iinr[0];
 121     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 122     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 123     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 124     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 125
 126     jq0              = _mm_set1_ps(charge[inr+0]);
 127     jq1              = _mm_set1_ps(charge[inr+1]);
 128     jq2              = _mm_set1_ps(charge[inr+2]);
 129     vdwjidx0A        = 2*vdwtype[inr+0];
 130     qq00             = _mm_mul_ps(iq0,jq0);
 131     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 132     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 133     qq01             = _mm_mul_ps(iq0,jq1);
 134     qq02             = _mm_mul_ps(iq0,jq2);
 135     qq10             = _mm_mul_ps(iq1,jq0);
 136     qq11             = _mm_mul_ps(iq1,jq1);
 137     qq12             = _mm_mul_ps(iq1,jq2);
 138     qq20             = _mm_mul_ps(iq2,jq0);
 139     qq21             = _mm_mul_ps(iq2,jq1);
 140     qq22             = _mm_mul_ps(iq2,jq2);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = jnrC = jnrD = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146     j_coord_offsetC = 0;
 147     j_coord_offsetD = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 173                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 174
 175         fix0             = _mm_setzero_ps();
 176         fiy0             = _mm_setzero_ps();
 177         fiz0             = _mm_setzero_ps();
 178         fix1             = _mm_setzero_ps();
 179         fiy1             = _mm_setzero_ps();
 180         fiz1             = _mm_setzero_ps();
 181         fix2             = _mm_setzero_ps();
 182         fiy2             = _mm_setzero_ps();
 183         fiz2             = _mm_setzero_ps();
 184
 185         /* Reset potential sums */
 186         velecsum         = _mm_setzero_ps();
 187         vvdwsum          = _mm_setzero_ps();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             jnrC             = jjnr[jidx+2];
 197             jnrD             = jjnr[jidx+3];
 198             j_coord_offsetA  = DIM*jnrA;
 199             j_coord_offsetB  = DIM*jnrB;
 200             j_coord_offsetC  = DIM*jnrC;
 201             j_coord_offsetD  = DIM*jnrD;
 202
 203             /* load j atom coordinates */
 204             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 205                                               x+j_coord_offsetC,x+j_coord_offsetD,
 206                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 207
 208             /* Calculate displacement vector */
 209             dx00             = _mm_sub_ps(ix0,jx0);
 210             dy00             = _mm_sub_ps(iy0,jy0);
 211             dz00             = _mm_sub_ps(iz0,jz0);
 212             dx01             = _mm_sub_ps(ix0,jx1);
 213             dy01             = _mm_sub_ps(iy0,jy1);
 214             dz01             = _mm_sub_ps(iz0,jz1);
 215             dx02             = _mm_sub_ps(ix0,jx2);
 216             dy02             = _mm_sub_ps(iy0,jy2);
 217             dz02             = _mm_sub_ps(iz0,jz2);
 218             dx10             = _mm_sub_ps(ix1,jx0);
 219             dy10             = _mm_sub_ps(iy1,jy0);
 220             dz10             = _mm_sub_ps(iz1,jz0);
 221             dx11             = _mm_sub_ps(ix1,jx1);
 222             dy11             = _mm_sub_ps(iy1,jy1);
 223             dz11             = _mm_sub_ps(iz1,jz1);
 224             dx12             = _mm_sub_ps(ix1,jx2);
 225             dy12             = _mm_sub_ps(iy1,jy2);
 226             dz12             = _mm_sub_ps(iz1,jz2);
 227             dx20             = _mm_sub_ps(ix2,jx0);
 228             dy20             = _mm_sub_ps(iy2,jy0);
 229             dz20             = _mm_sub_ps(iz2,jz0);
 230             dx21             = _mm_sub_ps(ix2,jx1);
 231             dy21             = _mm_sub_ps(iy2,jy1);
 232             dz21             = _mm_sub_ps(iz2,jz1);
 233             dx22             = _mm_sub_ps(ix2,jx2);
 234             dy22             = _mm_sub_ps(iy2,jy2);
 235             dz22             = _mm_sub_ps(iz2,jz2);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 239             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 240             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 241             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 242             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 243             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 244             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 245             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 246             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 247
 248             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 249             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 250             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 251             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 252             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 253             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 254             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 255             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 256             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 257
 258             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 259             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 260             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 261             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 262             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 263             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 264             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 265             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 266             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 267
 268             fjx0             = _mm_setzero_ps();
 269             fjy0             = _mm_setzero_ps();
 270             fjz0             = _mm_setzero_ps();
 271             fjx1             = _mm_setzero_ps();
 272             fjy1             = _mm_setzero_ps();
 273             fjz1             = _mm_setzero_ps();
 274             fjx2             = _mm_setzero_ps();
 275             fjy2             = _mm_setzero_ps();
 276             fjz2             = _mm_setzero_ps();
 277
 278             /**************************
 279              * CALCULATE INTERACTIONS *
 280              **************************/
 281
 282             /* COULOMB ELECTROSTATICS */
 283             velec            = _mm_mul_ps(qq00,rinv00);
 284             felec            = _mm_mul_ps(velec,rinvsq00);
 285
 286             /* LENNARD-JONES DISPERSION/REPULSION */
 287
 288             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 289             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 290             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 291             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 292             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 293
 294             /* Update potential sum for this i atom from the interaction with this j atom. */
 295             velecsum         = _mm_add_ps(velecsum,velec);
 296             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 297
 298             fscal            = _mm_add_ps(felec,fvdw);
 299
 300             /* Calculate temporary vectorial force */
 301             tx               = _mm_mul_ps(fscal,dx00);
 302             ty               = _mm_mul_ps(fscal,dy00);
 303             tz               = _mm_mul_ps(fscal,dz00);
 304
 305             /* Update vectorial force */
 306             fix0             = _mm_add_ps(fix0,tx);
 307             fiy0             = _mm_add_ps(fiy0,ty);
 308             fiz0             = _mm_add_ps(fiz0,tz);
 309
 310             fjx0             = _mm_add_ps(fjx0,tx);
 311             fjy0             = _mm_add_ps(fjy0,ty);
 312             fjz0             = _mm_add_ps(fjz0,tz);
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             /* COULOMB ELECTROSTATICS */
 319             velec            = _mm_mul_ps(qq01,rinv01);
 320             felec            = _mm_mul_ps(velec,rinvsq01);
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             velecsum         = _mm_add_ps(velecsum,velec);
 324
 325             fscal            = felec;
 326
 327             /* Calculate temporary vectorial force */
 328             tx               = _mm_mul_ps(fscal,dx01);
 329             ty               = _mm_mul_ps(fscal,dy01);
 330             tz               = _mm_mul_ps(fscal,dz01);
 331
 332             /* Update vectorial force */
 333             fix0             = _mm_add_ps(fix0,tx);
 334             fiy0             = _mm_add_ps(fiy0,ty);
 335             fiz0             = _mm_add_ps(fiz0,tz);
 336
 337             fjx1             = _mm_add_ps(fjx1,tx);
 338             fjy1             = _mm_add_ps(fjy1,ty);
 339             fjz1             = _mm_add_ps(fjz1,tz);
 340
 341             /**************************
 342              * CALCULATE INTERACTIONS *
 343              **************************/
 344
 345             /* COULOMB ELECTROSTATICS */
 346             velec            = _mm_mul_ps(qq02,rinv02);
 347             felec            = _mm_mul_ps(velec,rinvsq02);
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             velecsum         = _mm_add_ps(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = _mm_mul_ps(fscal,dx02);
 356             ty               = _mm_mul_ps(fscal,dy02);
 357             tz               = _mm_mul_ps(fscal,dz02);
 358
 359             /* Update vectorial force */
 360             fix0             = _mm_add_ps(fix0,tx);
 361             fiy0             = _mm_add_ps(fiy0,ty);
 362             fiz0             = _mm_add_ps(fiz0,tz);
 363
 364             fjx2             = _mm_add_ps(fjx2,tx);
 365             fjy2             = _mm_add_ps(fjy2,ty);
 366             fjz2             = _mm_add_ps(fjz2,tz);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             /* COULOMB ELECTROSTATICS */
 373             velec            = _mm_mul_ps(qq10,rinv10);
 374             felec            = _mm_mul_ps(velec,rinvsq10);
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velecsum         = _mm_add_ps(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm_mul_ps(fscal,dx10);
 383             ty               = _mm_mul_ps(fscal,dy10);
 384             tz               = _mm_mul_ps(fscal,dz10);
 385
 386             /* Update vectorial force */
 387             fix1             = _mm_add_ps(fix1,tx);
 388             fiy1             = _mm_add_ps(fiy1,ty);
 389             fiz1             = _mm_add_ps(fiz1,tz);
 390
 391             fjx0             = _mm_add_ps(fjx0,tx);
 392             fjy0             = _mm_add_ps(fjy0,ty);
 393             fjz0             = _mm_add_ps(fjz0,tz);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* COULOMB ELECTROSTATICS */
 400             velec            = _mm_mul_ps(qq11,rinv11);
 401             felec            = _mm_mul_ps(velec,rinvsq11);
 402
 403             /* Update potential sum for this i atom from the interaction with this j atom. */
 404             velecsum         = _mm_add_ps(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             /* Calculate temporary vectorial force */
 409             tx               = _mm_mul_ps(fscal,dx11);
 410             ty               = _mm_mul_ps(fscal,dy11);
 411             tz               = _mm_mul_ps(fscal,dz11);
 412
 413             /* Update vectorial force */
 414             fix1             = _mm_add_ps(fix1,tx);
 415             fiy1             = _mm_add_ps(fiy1,ty);
 416             fiz1             = _mm_add_ps(fiz1,tz);
 417
 418             fjx1             = _mm_add_ps(fjx1,tx);
 419             fjy1             = _mm_add_ps(fjy1,ty);
 420             fjz1             = _mm_add_ps(fjz1,tz);
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             /* COULOMB ELECTROSTATICS */
 427             velec            = _mm_mul_ps(qq12,rinv12);
 428             felec            = _mm_mul_ps(velec,rinvsq12);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velecsum         = _mm_add_ps(velecsum,velec);
 432
 433             fscal            = felec;
 434
 435             /* Calculate temporary vectorial force */
 436             tx               = _mm_mul_ps(fscal,dx12);
 437             ty               = _mm_mul_ps(fscal,dy12);
 438             tz               = _mm_mul_ps(fscal,dz12);
 439
 440             /* Update vectorial force */
 441             fix1             = _mm_add_ps(fix1,tx);
 442             fiy1             = _mm_add_ps(fiy1,ty);
 443             fiz1             = _mm_add_ps(fiz1,tz);
 444
 445             fjx2             = _mm_add_ps(fjx2,tx);
 446             fjy2             = _mm_add_ps(fjy2,ty);
 447             fjz2             = _mm_add_ps(fjz2,tz);
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             /* COULOMB ELECTROSTATICS */
 454             velec            = _mm_mul_ps(qq20,rinv20);
 455             felec            = _mm_mul_ps(velec,rinvsq20);
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velecsum         = _mm_add_ps(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             /* Calculate temporary vectorial force */
 463             tx               = _mm_mul_ps(fscal,dx20);
 464             ty               = _mm_mul_ps(fscal,dy20);
 465             tz               = _mm_mul_ps(fscal,dz20);
 466
 467             /* Update vectorial force */
 468             fix2             = _mm_add_ps(fix2,tx);
 469             fiy2             = _mm_add_ps(fiy2,ty);
 470             fiz2             = _mm_add_ps(fiz2,tz);
 471
 472             fjx0             = _mm_add_ps(fjx0,tx);
 473             fjy0             = _mm_add_ps(fjy0,ty);
 474             fjz0             = _mm_add_ps(fjz0,tz);
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             /* COULOMB ELECTROSTATICS */
 481             velec            = _mm_mul_ps(qq21,rinv21);
 482             felec            = _mm_mul_ps(velec,rinvsq21);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velecsum         = _mm_add_ps(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             /* Calculate temporary vectorial force */
 490             tx               = _mm_mul_ps(fscal,dx21);
 491             ty               = _mm_mul_ps(fscal,dy21);
 492             tz               = _mm_mul_ps(fscal,dz21);
 493
 494             /* Update vectorial force */
 495             fix2             = _mm_add_ps(fix2,tx);
 496             fiy2             = _mm_add_ps(fiy2,ty);
 497             fiz2             = _mm_add_ps(fiz2,tz);
 498
 499             fjx1             = _mm_add_ps(fjx1,tx);
 500             fjy1             = _mm_add_ps(fjy1,ty);
 501             fjz1             = _mm_add_ps(fjz1,tz);
 502
 503             /**************************
 504              * CALCULATE INTERACTIONS *
 505              **************************/
 506
 507             /* COULOMB ELECTROSTATICS */
 508             velec            = _mm_mul_ps(qq22,rinv22);
 509             felec            = _mm_mul_ps(velec,rinvsq22);
 510
 511             /* Update potential sum for this i atom from the interaction with this j atom. */
 512             velecsum         = _mm_add_ps(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             /* Calculate temporary vectorial force */
 517             tx               = _mm_mul_ps(fscal,dx22);
 518             ty               = _mm_mul_ps(fscal,dy22);
 519             tz               = _mm_mul_ps(fscal,dz22);
 520
 521             /* Update vectorial force */
 522             fix2             = _mm_add_ps(fix2,tx);
 523             fiy2             = _mm_add_ps(fiy2,ty);
 524             fiz2             = _mm_add_ps(fiz2,tz);
 525
 526             fjx2             = _mm_add_ps(fjx2,tx);
 527             fjy2             = _mm_add_ps(fjy2,ty);
 528             fjz2             = _mm_add_ps(fjz2,tz);
 529
 530             fjptrA             = f+j_coord_offsetA;
 531             fjptrB             = f+j_coord_offsetB;
 532             fjptrC             = f+j_coord_offsetC;
 533             fjptrD             = f+j_coord_offsetD;
 534
 535             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 536                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 537
 538             /* Inner loop uses 264 flops */
 539         }
 540
 541         if(jidx<j_index_end)
 542         {
 543
 544             /* Get j neighbor index, and coordinate index */
 545             jnrlistA         = jjnr[jidx];
 546             jnrlistB         = jjnr[jidx+1];
 547             jnrlistC         = jjnr[jidx+2];
 548             jnrlistD         = jjnr[jidx+3];
 549             /* Sign of each element will be negative for non-real atoms.
 550              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 551              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 552              */
 553             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 554             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 555             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 556             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 557             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 558             j_coord_offsetA  = DIM*jnrA;
 559             j_coord_offsetB  = DIM*jnrB;
 560             j_coord_offsetC  = DIM*jnrC;
 561             j_coord_offsetD  = DIM*jnrD;
 562
 563             /* load j atom coordinates */
 564             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 565                                               x+j_coord_offsetC,x+j_coord_offsetD,
 566                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 567
 568             /* Calculate displacement vector */
 569             dx00             = _mm_sub_ps(ix0,jx0);
 570             dy00             = _mm_sub_ps(iy0,jy0);
 571             dz00             = _mm_sub_ps(iz0,jz0);
 572             dx01             = _mm_sub_ps(ix0,jx1);
 573             dy01             = _mm_sub_ps(iy0,jy1);
 574             dz01             = _mm_sub_ps(iz0,jz1);
 575             dx02             = _mm_sub_ps(ix0,jx2);
 576             dy02             = _mm_sub_ps(iy0,jy2);
 577             dz02             = _mm_sub_ps(iz0,jz2);
 578             dx10             = _mm_sub_ps(ix1,jx0);
 579             dy10             = _mm_sub_ps(iy1,jy0);
 580             dz10             = _mm_sub_ps(iz1,jz0);
 581             dx11             = _mm_sub_ps(ix1,jx1);
 582             dy11             = _mm_sub_ps(iy1,jy1);
 583             dz11             = _mm_sub_ps(iz1,jz1);
 584             dx12             = _mm_sub_ps(ix1,jx2);
 585             dy12             = _mm_sub_ps(iy1,jy2);
 586             dz12             = _mm_sub_ps(iz1,jz2);
 587             dx20             = _mm_sub_ps(ix2,jx0);
 588             dy20             = _mm_sub_ps(iy2,jy0);
 589             dz20             = _mm_sub_ps(iz2,jz0);
 590             dx21             = _mm_sub_ps(ix2,jx1);
 591             dy21             = _mm_sub_ps(iy2,jy1);
 592             dz21             = _mm_sub_ps(iz2,jz1);
 593             dx22             = _mm_sub_ps(ix2,jx2);
 594             dy22             = _mm_sub_ps(iy2,jy2);
 595             dz22             = _mm_sub_ps(iz2,jz2);
 596
 597             /* Calculate squared distance and things based on it */
 598             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 599             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 600             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 601             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 602             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 603             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 604             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 605             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 606             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 607
 608             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 609             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 610             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 611             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 612             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 613             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 614             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 615             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 616             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 617
 618             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 619             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 620             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 621             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 622             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 623             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 624             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 625             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 626             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 627
 628             fjx0             = _mm_setzero_ps();
 629             fjy0             = _mm_setzero_ps();
 630             fjz0             = _mm_setzero_ps();
 631             fjx1             = _mm_setzero_ps();
 632             fjy1             = _mm_setzero_ps();
 633             fjz1             = _mm_setzero_ps();
 634             fjx2             = _mm_setzero_ps();
 635             fjy2             = _mm_setzero_ps();
 636             fjz2             = _mm_setzero_ps();
 637
 638             /**************************
 639              * CALCULATE INTERACTIONS *
 640              **************************/
 641
 642             /* COULOMB ELECTROSTATICS */
 643             velec            = _mm_mul_ps(qq00,rinv00);
 644             felec            = _mm_mul_ps(velec,rinvsq00);
 645
 646             /* LENNARD-JONES DISPERSION/REPULSION */
 647
 648             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 649             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 650             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 651             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 652             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 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             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 658             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 659
 660             fscal            = _mm_add_ps(felec,fvdw);
 661
 662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 663
 664             /* Calculate temporary vectorial force */
 665             tx               = _mm_mul_ps(fscal,dx00);
 666             ty               = _mm_mul_ps(fscal,dy00);
 667             tz               = _mm_mul_ps(fscal,dz00);
 668
 669             /* Update vectorial force */
 670             fix0             = _mm_add_ps(fix0,tx);
 671             fiy0             = _mm_add_ps(fiy0,ty);
 672             fiz0             = _mm_add_ps(fiz0,tz);
 673
 674             fjx0             = _mm_add_ps(fjx0,tx);
 675             fjy0             = _mm_add_ps(fjy0,ty);
 676             fjz0             = _mm_add_ps(fjz0,tz);
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             /* COULOMB ELECTROSTATICS */
 683             velec            = _mm_mul_ps(qq01,rinv01);
 684             felec            = _mm_mul_ps(velec,rinvsq01);
 685
 686             /* Update potential sum for this i atom from the interaction with this j atom. */
 687             velec            = _mm_andnot_ps(dummy_mask,velec);
 688             velecsum         = _mm_add_ps(velecsum,velec);
 689
 690             fscal            = felec;
 691
 692             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 693
 694             /* Calculate temporary vectorial force */
 695             tx               = _mm_mul_ps(fscal,dx01);
 696             ty               = _mm_mul_ps(fscal,dy01);
 697             tz               = _mm_mul_ps(fscal,dz01);
 698
 699             /* Update vectorial force */
 700             fix0             = _mm_add_ps(fix0,tx);
 701             fiy0             = _mm_add_ps(fiy0,ty);
 702             fiz0             = _mm_add_ps(fiz0,tz);
 703
 704             fjx1             = _mm_add_ps(fjx1,tx);
 705             fjy1             = _mm_add_ps(fjy1,ty);
 706             fjz1             = _mm_add_ps(fjz1,tz);
 707
 708             /**************************
 709              * CALCULATE INTERACTIONS *
 710              **************************/
 711
 712             /* COULOMB ELECTROSTATICS */
 713             velec            = _mm_mul_ps(qq02,rinv02);
 714             felec            = _mm_mul_ps(velec,rinvsq02);
 715
 716             /* Update potential sum for this i atom from the interaction with this j atom. */
 717             velec            = _mm_andnot_ps(dummy_mask,velec);
 718             velecsum         = _mm_add_ps(velecsum,velec);
 719
 720             fscal            = felec;
 721
 722             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 723
 724             /* Calculate temporary vectorial force */
 725             tx               = _mm_mul_ps(fscal,dx02);
 726             ty               = _mm_mul_ps(fscal,dy02);
 727             tz               = _mm_mul_ps(fscal,dz02);
 728
 729             /* Update vectorial force */
 730             fix0             = _mm_add_ps(fix0,tx);
 731             fiy0             = _mm_add_ps(fiy0,ty);
 732             fiz0             = _mm_add_ps(fiz0,tz);
 733
 734             fjx2             = _mm_add_ps(fjx2,tx);
 735             fjy2             = _mm_add_ps(fjy2,ty);
 736             fjz2             = _mm_add_ps(fjz2,tz);
 737
 738             /**************************
 739              * CALCULATE INTERACTIONS *
 740              **************************/
 741
 742             /* COULOMB ELECTROSTATICS */
 743             velec            = _mm_mul_ps(qq10,rinv10);
 744             felec            = _mm_mul_ps(velec,rinvsq10);
 745
 746             /* Update potential sum for this i atom from the interaction with this j atom. */
 747             velec            = _mm_andnot_ps(dummy_mask,velec);
 748             velecsum         = _mm_add_ps(velecsum,velec);
 749
 750             fscal            = felec;
 751
 752             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 753
 754             /* Calculate temporary vectorial force */
 755             tx               = _mm_mul_ps(fscal,dx10);
 756             ty               = _mm_mul_ps(fscal,dy10);
 757             tz               = _mm_mul_ps(fscal,dz10);
 758
 759             /* Update vectorial force */
 760             fix1             = _mm_add_ps(fix1,tx);
 761             fiy1             = _mm_add_ps(fiy1,ty);
 762             fiz1             = _mm_add_ps(fiz1,tz);
 763
 764             fjx0             = _mm_add_ps(fjx0,tx);
 765             fjy0             = _mm_add_ps(fjy0,ty);
 766             fjz0             = _mm_add_ps(fjz0,tz);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* COULOMB ELECTROSTATICS */
 773             velec            = _mm_mul_ps(qq11,rinv11);
 774             felec            = _mm_mul_ps(velec,rinvsq11);
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_andnot_ps(dummy_mask,velec);
 778             velecsum         = _mm_add_ps(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 783
 784             /* Calculate temporary vectorial force */
 785             tx               = _mm_mul_ps(fscal,dx11);
 786             ty               = _mm_mul_ps(fscal,dy11);
 787             tz               = _mm_mul_ps(fscal,dz11);
 788
 789             /* Update vectorial force */
 790             fix1             = _mm_add_ps(fix1,tx);
 791             fiy1             = _mm_add_ps(fiy1,ty);
 792             fiz1             = _mm_add_ps(fiz1,tz);
 793
 794             fjx1             = _mm_add_ps(fjx1,tx);
 795             fjy1             = _mm_add_ps(fjy1,ty);
 796             fjz1             = _mm_add_ps(fjz1,tz);
 797
 798             /**************************
 799              * CALCULATE INTERACTIONS *
 800              **************************/
 801
 802             /* COULOMB ELECTROSTATICS */
 803             velec            = _mm_mul_ps(qq12,rinv12);
 804             felec            = _mm_mul_ps(velec,rinvsq12);
 805
 806             /* Update potential sum for this i atom from the interaction with this j atom. */
 807             velec            = _mm_andnot_ps(dummy_mask,velec);
 808             velecsum         = _mm_add_ps(velecsum,velec);
 809
 810             fscal            = felec;
 811
 812             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 813
 814             /* Calculate temporary vectorial force */
 815             tx               = _mm_mul_ps(fscal,dx12);
 816             ty               = _mm_mul_ps(fscal,dy12);
 817             tz               = _mm_mul_ps(fscal,dz12);
 818
 819             /* Update vectorial force */
 820             fix1             = _mm_add_ps(fix1,tx);
 821             fiy1             = _mm_add_ps(fiy1,ty);
 822             fiz1             = _mm_add_ps(fiz1,tz);
 823
 824             fjx2             = _mm_add_ps(fjx2,tx);
 825             fjy2             = _mm_add_ps(fjy2,ty);
 826             fjz2             = _mm_add_ps(fjz2,tz);
 827
 828             /**************************
 829              * CALCULATE INTERACTIONS *
 830              **************************/
 831
 832             /* COULOMB ELECTROSTATICS */
 833             velec            = _mm_mul_ps(qq20,rinv20);
 834             felec            = _mm_mul_ps(velec,rinvsq20);
 835
 836             /* Update potential sum for this i atom from the interaction with this j atom. */
 837             velec            = _mm_andnot_ps(dummy_mask,velec);
 838             velecsum         = _mm_add_ps(velecsum,velec);
 839
 840             fscal            = felec;
 841
 842             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 843
 844             /* Calculate temporary vectorial force */
 845             tx               = _mm_mul_ps(fscal,dx20);
 846             ty               = _mm_mul_ps(fscal,dy20);
 847             tz               = _mm_mul_ps(fscal,dz20);
 848
 849             /* Update vectorial force */
 850             fix2             = _mm_add_ps(fix2,tx);
 851             fiy2             = _mm_add_ps(fiy2,ty);
 852             fiz2             = _mm_add_ps(fiz2,tz);
 853
 854             fjx0             = _mm_add_ps(fjx0,tx);
 855             fjy0             = _mm_add_ps(fjy0,ty);
 856             fjz0             = _mm_add_ps(fjz0,tz);
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             /* COULOMB ELECTROSTATICS */
 863             velec            = _mm_mul_ps(qq21,rinv21);
 864             felec            = _mm_mul_ps(velec,rinvsq21);
 865
 866             /* Update potential sum for this i atom from the interaction with this j atom. */
 867             velec            = _mm_andnot_ps(dummy_mask,velec);
 868             velecsum         = _mm_add_ps(velecsum,velec);
 869
 870             fscal            = felec;
 871
 872             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 873
 874             /* Calculate temporary vectorial force */
 875             tx               = _mm_mul_ps(fscal,dx21);
 876             ty               = _mm_mul_ps(fscal,dy21);
 877             tz               = _mm_mul_ps(fscal,dz21);
 878
 879             /* Update vectorial force */
 880             fix2             = _mm_add_ps(fix2,tx);
 881             fiy2             = _mm_add_ps(fiy2,ty);
 882             fiz2             = _mm_add_ps(fiz2,tz);
 883
 884             fjx1             = _mm_add_ps(fjx1,tx);
 885             fjy1             = _mm_add_ps(fjy1,ty);
 886             fjz1             = _mm_add_ps(fjz1,tz);
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             /* COULOMB ELECTROSTATICS */
 893             velec            = _mm_mul_ps(qq22,rinv22);
 894             felec            = _mm_mul_ps(velec,rinvsq22);
 895
 896             /* Update potential sum for this i atom from the interaction with this j atom. */
 897             velec            = _mm_andnot_ps(dummy_mask,velec);
 898             velecsum         = _mm_add_ps(velecsum,velec);
 899
 900             fscal            = felec;
 901
 902             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 903
 904             /* Calculate temporary vectorial force */
 905             tx               = _mm_mul_ps(fscal,dx22);
 906             ty               = _mm_mul_ps(fscal,dy22);
 907             tz               = _mm_mul_ps(fscal,dz22);
 908
 909             /* Update vectorial force */
 910             fix2             = _mm_add_ps(fix2,tx);
 911             fiy2             = _mm_add_ps(fiy2,ty);
 912             fiz2             = _mm_add_ps(fiz2,tz);
 913
 914             fjx2             = _mm_add_ps(fjx2,tx);
 915             fjy2             = _mm_add_ps(fjy2,ty);
 916             fjz2             = _mm_add_ps(fjz2,tz);
 917
 918             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 919             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 920             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 921             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 922
 923             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 924                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 925
 926             /* Inner loop uses 264 flops */
 927         }
 928
 929         /* End of innermost loop */
 930
 931         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 932                                               f+i_coord_offset,fshift+i_shift_offset);
 933
 934         ggid                        = gid[iidx];
 935         /* Update potential energies */
 936         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 937         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 938
 939         /* Increment number of inner iterations */
 940         inneriter                  += j_index_end - j_index_start;
 941
 942         /* Outer loop uses 20 flops */
 943     }
 944
 945     /* Increment number of outer iterations */
 946     outeriter        += nri;
 947
 948     /* Update outer/inner flops */
 949
 950     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*264);
 951 }
 952 /*
 953  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
 954  * Electrostatics interaction: Coulomb
 955  * VdW interaction:            LennardJones
 956  * Geometry:                   Water3-Water3
 957  * Calculate force/pot:        Force
 958  */
 959 void
 960 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
 961                     (t_nblist * gmx_restrict                nlist,
 962                      rvec * gmx_restrict                    xx,
 963                      rvec * gmx_restrict                    ff,
 964                      t_forcerec * gmx_restrict              fr,
 965                      t_mdatoms * gmx_restrict               mdatoms,
 966                      nb_kernel_data_t * gmx_restrict        kernel_data,
 967                      t_nrnb * gmx_restrict                  nrnb)
 968 {
 969     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 970      * just 0 for non-waters.
 971      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 972      * jnr indices corresponding to data put in the four positions in the SIMD register.
 973      */
 974     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 975     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 976     int              jnrA,jnrB,jnrC,jnrD;
 977     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 978     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 979     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 980     real             rcutoff_scalar;
 981     real             *shiftvec,*fshift,*x,*f;
 982     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 983     real             scratch[4*DIM];
 984     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 985     int              vdwioffset0;
 986     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 987     int              vdwioffset1;
 988     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 989     int              vdwioffset2;
 990     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 991     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 992     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 993     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 994     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 995     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 996     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 997     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 998     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 999     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1000     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1001     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1002     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1003     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1004     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1005     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1006     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1007     real             *charge;
1008     int              nvdwtype;
1009     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1010     int              *vdwtype;
1011     real             *vdwparam;
1012     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1013     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1014     __m128           dummy_mask,cutoff_mask;
1015     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1016     __m128           one     = _mm_set1_ps(1.0);
1017     __m128           two     = _mm_set1_ps(2.0);
1018     x                = xx[0];
1019     f                = ff[0];
1020
1021     nri              = nlist->nri;
1022     iinr             = nlist->iinr;
1023     jindex           = nlist->jindex;
1024     jjnr             = nlist->jjnr;
1025     shiftidx         = nlist->shift;
1026     gid              = nlist->gid;
1027     shiftvec         = fr->shift_vec[0];
1028     fshift           = fr->fshift[0];
1029     facel            = _mm_set1_ps(fr->epsfac);
1030     charge           = mdatoms->chargeA;
1031     nvdwtype         = fr->ntype;
1032     vdwparam         = fr->nbfp;
1033     vdwtype          = mdatoms->typeA;
1034
1035     /* Setup water-specific parameters */
1036     inr              = nlist->iinr[0];
1037     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1038     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1039     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1040     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1041
1042     jq0              = _mm_set1_ps(charge[inr+0]);
1043     jq1              = _mm_set1_ps(charge[inr+1]);
1044     jq2              = _mm_set1_ps(charge[inr+2]);
1045     vdwjidx0A        = 2*vdwtype[inr+0];
1046     qq00             = _mm_mul_ps(iq0,jq0);
1047     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1048     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1049     qq01             = _mm_mul_ps(iq0,jq1);
1050     qq02             = _mm_mul_ps(iq0,jq2);
1051     qq10             = _mm_mul_ps(iq1,jq0);
1052     qq11             = _mm_mul_ps(iq1,jq1);
1053     qq12             = _mm_mul_ps(iq1,jq2);
1054     qq20             = _mm_mul_ps(iq2,jq0);
1055     qq21             = _mm_mul_ps(iq2,jq1);
1056     qq22             = _mm_mul_ps(iq2,jq2);
1057
1058     /* Avoid stupid compiler warnings */
1059     jnrA = jnrB = jnrC = jnrD = 0;
1060     j_coord_offsetA = 0;
1061     j_coord_offsetB = 0;
1062     j_coord_offsetC = 0;
1063     j_coord_offsetD = 0;
1064
1065     outeriter        = 0;
1066     inneriter        = 0;
1067
1068     for(iidx=0;iidx<4*DIM;iidx++)
1069     {
1070         scratch[iidx] = 0.0;
1071     }
1072
1073     /* Start outer loop over neighborlists */
1074     for(iidx=0; iidx<nri; iidx++)
1075     {
1076         /* Load shift vector for this list */
1077         i_shift_offset   = DIM*shiftidx[iidx];
1078
1079         /* Load limits for loop over neighbors */
1080         j_index_start    = jindex[iidx];
1081         j_index_end      = jindex[iidx+1];
1082
1083         /* Get outer coordinate index */
1084         inr              = iinr[iidx];
1085         i_coord_offset   = DIM*inr;
1086
1087         /* Load i particle coords and add shift vector */
1088         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1089                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1090
1091         fix0             = _mm_setzero_ps();
1092         fiy0             = _mm_setzero_ps();
1093         fiz0             = _mm_setzero_ps();
1094         fix1             = _mm_setzero_ps();
1095         fiy1             = _mm_setzero_ps();
1096         fiz1             = _mm_setzero_ps();
1097         fix2             = _mm_setzero_ps();
1098         fiy2             = _mm_setzero_ps();
1099         fiz2             = _mm_setzero_ps();
1100
1101         /* Start inner kernel loop */
1102         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1103         {
1104
1105             /* Get j neighbor index, and coordinate index */
1106             jnrA             = jjnr[jidx];
1107             jnrB             = jjnr[jidx+1];
1108             jnrC             = jjnr[jidx+2];
1109             jnrD             = jjnr[jidx+3];
1110             j_coord_offsetA  = DIM*jnrA;
1111             j_coord_offsetB  = DIM*jnrB;
1112             j_coord_offsetC  = DIM*jnrC;
1113             j_coord_offsetD  = DIM*jnrD;
1114
1115             /* load j atom coordinates */
1116             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1117                                               x+j_coord_offsetC,x+j_coord_offsetD,
1118                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1119
1120             /* Calculate displacement vector */
1121             dx00             = _mm_sub_ps(ix0,jx0);
1122             dy00             = _mm_sub_ps(iy0,jy0);
1123             dz00             = _mm_sub_ps(iz0,jz0);
1124             dx01             = _mm_sub_ps(ix0,jx1);
1125             dy01             = _mm_sub_ps(iy0,jy1);
1126             dz01             = _mm_sub_ps(iz0,jz1);
1127             dx02             = _mm_sub_ps(ix0,jx2);
1128             dy02             = _mm_sub_ps(iy0,jy2);
1129             dz02             = _mm_sub_ps(iz0,jz2);
1130             dx10             = _mm_sub_ps(ix1,jx0);
1131             dy10             = _mm_sub_ps(iy1,jy0);
1132             dz10             = _mm_sub_ps(iz1,jz0);
1133             dx11             = _mm_sub_ps(ix1,jx1);
1134             dy11             = _mm_sub_ps(iy1,jy1);
1135             dz11             = _mm_sub_ps(iz1,jz1);
1136             dx12             = _mm_sub_ps(ix1,jx2);
1137             dy12             = _mm_sub_ps(iy1,jy2);
1138             dz12             = _mm_sub_ps(iz1,jz2);
1139             dx20             = _mm_sub_ps(ix2,jx0);
1140             dy20             = _mm_sub_ps(iy2,jy0);
1141             dz20             = _mm_sub_ps(iz2,jz0);
1142             dx21             = _mm_sub_ps(ix2,jx1);
1143             dy21             = _mm_sub_ps(iy2,jy1);
1144             dz21             = _mm_sub_ps(iz2,jz1);
1145             dx22             = _mm_sub_ps(ix2,jx2);
1146             dy22             = _mm_sub_ps(iy2,jy2);
1147             dz22             = _mm_sub_ps(iz2,jz2);
1148
1149             /* Calculate squared distance and things based on it */
1150             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1151             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1152             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1153             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1154             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1155             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1156             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1157             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1158             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1159
1160             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1161             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1162             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1163             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1164             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1165             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1166             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1167             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1168             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1169
1170             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1171             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1172             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1173             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1174             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1175             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1176             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1177             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1178             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1179
1180             fjx0             = _mm_setzero_ps();
1181             fjy0             = _mm_setzero_ps();
1182             fjz0             = _mm_setzero_ps();
1183             fjx1             = _mm_setzero_ps();
1184             fjy1             = _mm_setzero_ps();
1185             fjz1             = _mm_setzero_ps();
1186             fjx2             = _mm_setzero_ps();
1187             fjy2             = _mm_setzero_ps();
1188             fjz2             = _mm_setzero_ps();
1189
1190             /**************************
1191              * CALCULATE INTERACTIONS *
1192              **************************/
1193
1194             /* COULOMB ELECTROSTATICS */
1195             velec            = _mm_mul_ps(qq00,rinv00);
1196             felec            = _mm_mul_ps(velec,rinvsq00);
1197
1198             /* LENNARD-JONES DISPERSION/REPULSION */
1199
1200             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1201             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1202
1203             fscal            = _mm_add_ps(felec,fvdw);
1204
1205             /* Calculate temporary vectorial force */
1206             tx               = _mm_mul_ps(fscal,dx00);
1207             ty               = _mm_mul_ps(fscal,dy00);
1208             tz               = _mm_mul_ps(fscal,dz00);
1209
1210             /* Update vectorial force */
1211             fix0             = _mm_add_ps(fix0,tx);
1212             fiy0             = _mm_add_ps(fiy0,ty);
1213             fiz0             = _mm_add_ps(fiz0,tz);
1214
1215             fjx0             = _mm_add_ps(fjx0,tx);
1216             fjy0             = _mm_add_ps(fjy0,ty);
1217             fjz0             = _mm_add_ps(fjz0,tz);
1218
1219             /**************************
1220              * CALCULATE INTERACTIONS *
1221              **************************/
1222
1223             /* COULOMB ELECTROSTATICS */
1224             velec            = _mm_mul_ps(qq01,rinv01);
1225             felec            = _mm_mul_ps(velec,rinvsq01);
1226
1227             fscal            = felec;
1228
1229             /* Calculate temporary vectorial force */
1230             tx               = _mm_mul_ps(fscal,dx01);
1231             ty               = _mm_mul_ps(fscal,dy01);
1232             tz               = _mm_mul_ps(fscal,dz01);
1233
1234             /* Update vectorial force */
1235             fix0             = _mm_add_ps(fix0,tx);
1236             fiy0             = _mm_add_ps(fiy0,ty);
1237             fiz0             = _mm_add_ps(fiz0,tz);
1238
1239             fjx1             = _mm_add_ps(fjx1,tx);
1240             fjy1             = _mm_add_ps(fjy1,ty);
1241             fjz1             = _mm_add_ps(fjz1,tz);
1242
1243             /**************************
1244              * CALCULATE INTERACTIONS *
1245              **************************/
1246
1247             /* COULOMB ELECTROSTATICS */
1248             velec            = _mm_mul_ps(qq02,rinv02);
1249             felec            = _mm_mul_ps(velec,rinvsq02);
1250
1251             fscal            = felec;
1252
1253             /* Calculate temporary vectorial force */
1254             tx               = _mm_mul_ps(fscal,dx02);
1255             ty               = _mm_mul_ps(fscal,dy02);
1256             tz               = _mm_mul_ps(fscal,dz02);
1257
1258             /* Update vectorial force */
1259             fix0             = _mm_add_ps(fix0,tx);
1260             fiy0             = _mm_add_ps(fiy0,ty);
1261             fiz0             = _mm_add_ps(fiz0,tz);
1262
1263             fjx2             = _mm_add_ps(fjx2,tx);
1264             fjy2             = _mm_add_ps(fjy2,ty);
1265             fjz2             = _mm_add_ps(fjz2,tz);
1266
1267             /**************************
1268              * CALCULATE INTERACTIONS *
1269              **************************/
1270
1271             /* COULOMB ELECTROSTATICS */
1272             velec            = _mm_mul_ps(qq10,rinv10);
1273             felec            = _mm_mul_ps(velec,rinvsq10);
1274
1275             fscal            = felec;
1276
1277             /* Calculate temporary vectorial force */
1278             tx               = _mm_mul_ps(fscal,dx10);
1279             ty               = _mm_mul_ps(fscal,dy10);
1280             tz               = _mm_mul_ps(fscal,dz10);
1281
1282             /* Update vectorial force */
1283             fix1             = _mm_add_ps(fix1,tx);
1284             fiy1             = _mm_add_ps(fiy1,ty);
1285             fiz1             = _mm_add_ps(fiz1,tz);
1286
1287             fjx0             = _mm_add_ps(fjx0,tx);
1288             fjy0             = _mm_add_ps(fjy0,ty);
1289             fjz0             = _mm_add_ps(fjz0,tz);
1290
1291             /**************************
1292              * CALCULATE INTERACTIONS *
1293              **************************/
1294
1295             /* COULOMB ELECTROSTATICS */
1296             velec            = _mm_mul_ps(qq11,rinv11);
1297             felec            = _mm_mul_ps(velec,rinvsq11);
1298
1299             fscal            = felec;
1300
1301             /* Calculate temporary vectorial force */
1302             tx               = _mm_mul_ps(fscal,dx11);
1303             ty               = _mm_mul_ps(fscal,dy11);
1304             tz               = _mm_mul_ps(fscal,dz11);
1305
1306             /* Update vectorial force */
1307             fix1             = _mm_add_ps(fix1,tx);
1308             fiy1             = _mm_add_ps(fiy1,ty);
1309             fiz1             = _mm_add_ps(fiz1,tz);
1310
1311             fjx1             = _mm_add_ps(fjx1,tx);
1312             fjy1             = _mm_add_ps(fjy1,ty);
1313             fjz1             = _mm_add_ps(fjz1,tz);
1314
1315             /**************************
1316              * CALCULATE INTERACTIONS *
1317              **************************/
1318
1319             /* COULOMB ELECTROSTATICS */
1320             velec            = _mm_mul_ps(qq12,rinv12);
1321             felec            = _mm_mul_ps(velec,rinvsq12);
1322
1323             fscal            = felec;
1324
1325             /* Calculate temporary vectorial force */
1326             tx               = _mm_mul_ps(fscal,dx12);
1327             ty               = _mm_mul_ps(fscal,dy12);
1328             tz               = _mm_mul_ps(fscal,dz12);
1329
1330             /* Update vectorial force */
1331             fix1             = _mm_add_ps(fix1,tx);
1332             fiy1             = _mm_add_ps(fiy1,ty);
1333             fiz1             = _mm_add_ps(fiz1,tz);
1334
1335             fjx2             = _mm_add_ps(fjx2,tx);
1336             fjy2             = _mm_add_ps(fjy2,ty);
1337             fjz2             = _mm_add_ps(fjz2,tz);
1338
1339             /**************************
1340              * CALCULATE INTERACTIONS *
1341              **************************/
1342
1343             /* COULOMB ELECTROSTATICS */
1344             velec            = _mm_mul_ps(qq20,rinv20);
1345             felec            = _mm_mul_ps(velec,rinvsq20);
1346
1347             fscal            = felec;
1348
1349             /* Calculate temporary vectorial force */
1350             tx               = _mm_mul_ps(fscal,dx20);
1351             ty               = _mm_mul_ps(fscal,dy20);
1352             tz               = _mm_mul_ps(fscal,dz20);
1353
1354             /* Update vectorial force */
1355             fix2             = _mm_add_ps(fix2,tx);
1356             fiy2             = _mm_add_ps(fiy2,ty);
1357             fiz2             = _mm_add_ps(fiz2,tz);
1358
1359             fjx0             = _mm_add_ps(fjx0,tx);
1360             fjy0             = _mm_add_ps(fjy0,ty);
1361             fjz0             = _mm_add_ps(fjz0,tz);
1362
1363             /**************************
1364              * CALCULATE INTERACTIONS *
1365              **************************/
1366
1367             /* COULOMB ELECTROSTATICS */
1368             velec            = _mm_mul_ps(qq21,rinv21);
1369             felec            = _mm_mul_ps(velec,rinvsq21);
1370
1371             fscal            = felec;
1372
1373             /* Calculate temporary vectorial force */
1374             tx               = _mm_mul_ps(fscal,dx21);
1375             ty               = _mm_mul_ps(fscal,dy21);
1376             tz               = _mm_mul_ps(fscal,dz21);
1377
1378             /* Update vectorial force */
1379             fix2             = _mm_add_ps(fix2,tx);
1380             fiy2             = _mm_add_ps(fiy2,ty);
1381             fiz2             = _mm_add_ps(fiz2,tz);
1382
1383             fjx1             = _mm_add_ps(fjx1,tx);
1384             fjy1             = _mm_add_ps(fjy1,ty);
1385             fjz1             = _mm_add_ps(fjz1,tz);
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             /* COULOMB ELECTROSTATICS */
1392             velec            = _mm_mul_ps(qq22,rinv22);
1393             felec            = _mm_mul_ps(velec,rinvsq22);
1394
1395             fscal            = felec;
1396
1397             /* Calculate temporary vectorial force */
1398             tx               = _mm_mul_ps(fscal,dx22);
1399             ty               = _mm_mul_ps(fscal,dy22);
1400             tz               = _mm_mul_ps(fscal,dz22);
1401
1402             /* Update vectorial force */
1403             fix2             = _mm_add_ps(fix2,tx);
1404             fiy2             = _mm_add_ps(fiy2,ty);
1405             fiz2             = _mm_add_ps(fiz2,tz);
1406
1407             fjx2             = _mm_add_ps(fjx2,tx);
1408             fjy2             = _mm_add_ps(fjy2,ty);
1409             fjz2             = _mm_add_ps(fjz2,tz);
1410
1411             fjptrA             = f+j_coord_offsetA;
1412             fjptrB             = f+j_coord_offsetB;
1413             fjptrC             = f+j_coord_offsetC;
1414             fjptrD             = f+j_coord_offsetD;
1415
1416             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1417                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1418
1419             /* Inner loop uses 250 flops */
1420         }
1421
1422         if(jidx<j_index_end)
1423         {
1424
1425             /* Get j neighbor index, and coordinate index */
1426             jnrlistA         = jjnr[jidx];
1427             jnrlistB         = jjnr[jidx+1];
1428             jnrlistC         = jjnr[jidx+2];
1429             jnrlistD         = jjnr[jidx+3];
1430             /* Sign of each element will be negative for non-real atoms.
1431              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1432              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1433              */
1434             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1435             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1436             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1437             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1438             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1439             j_coord_offsetA  = DIM*jnrA;
1440             j_coord_offsetB  = DIM*jnrB;
1441             j_coord_offsetC  = DIM*jnrC;
1442             j_coord_offsetD  = DIM*jnrD;
1443
1444             /* load j atom coordinates */
1445             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1446                                               x+j_coord_offsetC,x+j_coord_offsetD,
1447                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1448
1449             /* Calculate displacement vector */
1450             dx00             = _mm_sub_ps(ix0,jx0);
1451             dy00             = _mm_sub_ps(iy0,jy0);
1452             dz00             = _mm_sub_ps(iz0,jz0);
1453             dx01             = _mm_sub_ps(ix0,jx1);
1454             dy01             = _mm_sub_ps(iy0,jy1);
1455             dz01             = _mm_sub_ps(iz0,jz1);
1456             dx02             = _mm_sub_ps(ix0,jx2);
1457             dy02             = _mm_sub_ps(iy0,jy2);
1458             dz02             = _mm_sub_ps(iz0,jz2);
1459             dx10             = _mm_sub_ps(ix1,jx0);
1460             dy10             = _mm_sub_ps(iy1,jy0);
1461             dz10             = _mm_sub_ps(iz1,jz0);
1462             dx11             = _mm_sub_ps(ix1,jx1);
1463             dy11             = _mm_sub_ps(iy1,jy1);
1464             dz11             = _mm_sub_ps(iz1,jz1);
1465             dx12             = _mm_sub_ps(ix1,jx2);
1466             dy12             = _mm_sub_ps(iy1,jy2);
1467             dz12             = _mm_sub_ps(iz1,jz2);
1468             dx20             = _mm_sub_ps(ix2,jx0);
1469             dy20             = _mm_sub_ps(iy2,jy0);
1470             dz20             = _mm_sub_ps(iz2,jz0);
1471             dx21             = _mm_sub_ps(ix2,jx1);
1472             dy21             = _mm_sub_ps(iy2,jy1);
1473             dz21             = _mm_sub_ps(iz2,jz1);
1474             dx22             = _mm_sub_ps(ix2,jx2);
1475             dy22             = _mm_sub_ps(iy2,jy2);
1476             dz22             = _mm_sub_ps(iz2,jz2);
1477
1478             /* Calculate squared distance and things based on it */
1479             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1480             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1481             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1482             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1483             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1484             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1485             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1486             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1487             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1488
1489             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1490             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1491             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1492             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1493             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1494             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1495             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1496             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1497             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1498
1499             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1500             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1501             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1502             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1503             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1504             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1505             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1506             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1507             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1508
1509             fjx0             = _mm_setzero_ps();
1510             fjy0             = _mm_setzero_ps();
1511             fjz0             = _mm_setzero_ps();
1512             fjx1             = _mm_setzero_ps();
1513             fjy1             = _mm_setzero_ps();
1514             fjz1             = _mm_setzero_ps();
1515             fjx2             = _mm_setzero_ps();
1516             fjy2             = _mm_setzero_ps();
1517             fjz2             = _mm_setzero_ps();
1518
1519             /**************************
1520              * CALCULATE INTERACTIONS *
1521              **************************/
1522
1523             /* COULOMB ELECTROSTATICS */
1524             velec            = _mm_mul_ps(qq00,rinv00);
1525             felec            = _mm_mul_ps(velec,rinvsq00);
1526
1527             /* LENNARD-JONES DISPERSION/REPULSION */
1528
1529             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1530             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1531
1532             fscal            = _mm_add_ps(felec,fvdw);
1533
1534             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1535
1536             /* Calculate temporary vectorial force */
1537             tx               = _mm_mul_ps(fscal,dx00);
1538             ty               = _mm_mul_ps(fscal,dy00);
1539             tz               = _mm_mul_ps(fscal,dz00);
1540
1541             /* Update vectorial force */
1542             fix0             = _mm_add_ps(fix0,tx);
1543             fiy0             = _mm_add_ps(fiy0,ty);
1544             fiz0             = _mm_add_ps(fiz0,tz);
1545
1546             fjx0             = _mm_add_ps(fjx0,tx);
1547             fjy0             = _mm_add_ps(fjy0,ty);
1548             fjz0             = _mm_add_ps(fjz0,tz);
1549
1550             /**************************
1551              * CALCULATE INTERACTIONS *
1552              **************************/
1553
1554             /* COULOMB ELECTROSTATICS */
1555             velec            = _mm_mul_ps(qq01,rinv01);
1556             felec            = _mm_mul_ps(velec,rinvsq01);
1557
1558             fscal            = felec;
1559
1560             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1561
1562             /* Calculate temporary vectorial force */
1563             tx               = _mm_mul_ps(fscal,dx01);
1564             ty               = _mm_mul_ps(fscal,dy01);
1565             tz               = _mm_mul_ps(fscal,dz01);
1566
1567             /* Update vectorial force */
1568             fix0             = _mm_add_ps(fix0,tx);
1569             fiy0             = _mm_add_ps(fiy0,ty);
1570             fiz0             = _mm_add_ps(fiz0,tz);
1571
1572             fjx1             = _mm_add_ps(fjx1,tx);
1573             fjy1             = _mm_add_ps(fjy1,ty);
1574             fjz1             = _mm_add_ps(fjz1,tz);
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             /* COULOMB ELECTROSTATICS */
1581             velec            = _mm_mul_ps(qq02,rinv02);
1582             felec            = _mm_mul_ps(velec,rinvsq02);
1583
1584             fscal            = felec;
1585
1586             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1587
1588             /* Calculate temporary vectorial force */
1589             tx               = _mm_mul_ps(fscal,dx02);
1590             ty               = _mm_mul_ps(fscal,dy02);
1591             tz               = _mm_mul_ps(fscal,dz02);
1592
1593             /* Update vectorial force */
1594             fix0             = _mm_add_ps(fix0,tx);
1595             fiy0             = _mm_add_ps(fiy0,ty);
1596             fiz0             = _mm_add_ps(fiz0,tz);
1597
1598             fjx2             = _mm_add_ps(fjx2,tx);
1599             fjy2             = _mm_add_ps(fjy2,ty);
1600             fjz2             = _mm_add_ps(fjz2,tz);
1601
1602             /**************************
1603              * CALCULATE INTERACTIONS *
1604              **************************/
1605
1606             /* COULOMB ELECTROSTATICS */
1607             velec            = _mm_mul_ps(qq10,rinv10);
1608             felec            = _mm_mul_ps(velec,rinvsq10);
1609
1610             fscal            = felec;
1611
1612             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1613
1614             /* Calculate temporary vectorial force */
1615             tx               = _mm_mul_ps(fscal,dx10);
1616             ty               = _mm_mul_ps(fscal,dy10);
1617             tz               = _mm_mul_ps(fscal,dz10);
1618
1619             /* Update vectorial force */
1620             fix1             = _mm_add_ps(fix1,tx);
1621             fiy1             = _mm_add_ps(fiy1,ty);
1622             fiz1             = _mm_add_ps(fiz1,tz);
1623
1624             fjx0             = _mm_add_ps(fjx0,tx);
1625             fjy0             = _mm_add_ps(fjy0,ty);
1626             fjz0             = _mm_add_ps(fjz0,tz);
1627
1628             /**************************
1629              * CALCULATE INTERACTIONS *
1630              **************************/
1631
1632             /* COULOMB ELECTROSTATICS */
1633             velec            = _mm_mul_ps(qq11,rinv11);
1634             felec            = _mm_mul_ps(velec,rinvsq11);
1635
1636             fscal            = felec;
1637
1638             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1639
1640             /* Calculate temporary vectorial force */
1641             tx               = _mm_mul_ps(fscal,dx11);
1642             ty               = _mm_mul_ps(fscal,dy11);
1643             tz               = _mm_mul_ps(fscal,dz11);
1644
1645             /* Update vectorial force */
1646             fix1             = _mm_add_ps(fix1,tx);
1647             fiy1             = _mm_add_ps(fiy1,ty);
1648             fiz1             = _mm_add_ps(fiz1,tz);
1649
1650             fjx1             = _mm_add_ps(fjx1,tx);
1651             fjy1             = _mm_add_ps(fjy1,ty);
1652             fjz1             = _mm_add_ps(fjz1,tz);
1653
1654             /**************************
1655              * CALCULATE INTERACTIONS *
1656              **************************/
1657
1658             /* COULOMB ELECTROSTATICS */
1659             velec            = _mm_mul_ps(qq12,rinv12);
1660             felec            = _mm_mul_ps(velec,rinvsq12);
1661
1662             fscal            = felec;
1663
1664             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1665
1666             /* Calculate temporary vectorial force */
1667             tx               = _mm_mul_ps(fscal,dx12);
1668             ty               = _mm_mul_ps(fscal,dy12);
1669             tz               = _mm_mul_ps(fscal,dz12);
1670
1671             /* Update vectorial force */
1672             fix1             = _mm_add_ps(fix1,tx);
1673             fiy1             = _mm_add_ps(fiy1,ty);
1674             fiz1             = _mm_add_ps(fiz1,tz);
1675
1676             fjx2             = _mm_add_ps(fjx2,tx);
1677             fjy2             = _mm_add_ps(fjy2,ty);
1678             fjz2             = _mm_add_ps(fjz2,tz);
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             /* COULOMB ELECTROSTATICS */
1685             velec            = _mm_mul_ps(qq20,rinv20);
1686             felec            = _mm_mul_ps(velec,rinvsq20);
1687
1688             fscal            = felec;
1689
1690             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1691
1692             /* Calculate temporary vectorial force */
1693             tx               = _mm_mul_ps(fscal,dx20);
1694             ty               = _mm_mul_ps(fscal,dy20);
1695             tz               = _mm_mul_ps(fscal,dz20);
1696
1697             /* Update vectorial force */
1698             fix2             = _mm_add_ps(fix2,tx);
1699             fiy2             = _mm_add_ps(fiy2,ty);
1700             fiz2             = _mm_add_ps(fiz2,tz);
1701
1702             fjx0             = _mm_add_ps(fjx0,tx);
1703             fjy0             = _mm_add_ps(fjy0,ty);
1704             fjz0             = _mm_add_ps(fjz0,tz);
1705
1706             /**************************
1707              * CALCULATE INTERACTIONS *
1708              **************************/
1709
1710             /* COULOMB ELECTROSTATICS */
1711             velec            = _mm_mul_ps(qq21,rinv21);
1712             felec            = _mm_mul_ps(velec,rinvsq21);
1713
1714             fscal            = felec;
1715
1716             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1717
1718             /* Calculate temporary vectorial force */
1719             tx               = _mm_mul_ps(fscal,dx21);
1720             ty               = _mm_mul_ps(fscal,dy21);
1721             tz               = _mm_mul_ps(fscal,dz21);
1722
1723             /* Update vectorial force */
1724             fix2             = _mm_add_ps(fix2,tx);
1725             fiy2             = _mm_add_ps(fiy2,ty);
1726             fiz2             = _mm_add_ps(fiz2,tz);
1727
1728             fjx1             = _mm_add_ps(fjx1,tx);
1729             fjy1             = _mm_add_ps(fjy1,ty);
1730             fjz1             = _mm_add_ps(fjz1,tz);
1731
1732             /**************************
1733              * CALCULATE INTERACTIONS *
1734              **************************/
1735
1736             /* COULOMB ELECTROSTATICS */
1737             velec            = _mm_mul_ps(qq22,rinv22);
1738             felec            = _mm_mul_ps(velec,rinvsq22);
1739
1740             fscal            = felec;
1741
1742             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1743
1744             /* Calculate temporary vectorial force */
1745             tx               = _mm_mul_ps(fscal,dx22);
1746             ty               = _mm_mul_ps(fscal,dy22);
1747             tz               = _mm_mul_ps(fscal,dz22);
1748
1749             /* Update vectorial force */
1750             fix2             = _mm_add_ps(fix2,tx);
1751             fiy2             = _mm_add_ps(fiy2,ty);
1752             fiz2             = _mm_add_ps(fiz2,tz);
1753
1754             fjx2             = _mm_add_ps(fjx2,tx);
1755             fjy2             = _mm_add_ps(fjy2,ty);
1756             fjz2             = _mm_add_ps(fjz2,tz);
1757
1758             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1759             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1760             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1761             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1762
1763             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1764                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1765
1766             /* Inner loop uses 250 flops */
1767         }
1768
1769         /* End of innermost loop */
1770
1771         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1772                                               f+i_coord_offset,fshift+i_shift_offset);
1773
1774         /* Increment number of inner iterations */
1775         inneriter                  += j_index_end - j_index_start;
1776
1777         /* Outer loop uses 18 flops */
1778     }
1779
1780     /* Increment number of outer iterations */
1781     outeriter        += nri;
1782
1783     /* Update outer/inner flops */
1784
1785     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1786 }