src/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double 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_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B;
  75     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B;
  77     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B;
  79     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B;
  81     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  99     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 100     __m128d          dummy_mask,cutoff_mask;
 101     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 102     __m128d          one     = _mm_set1_pd(1.0);
 103     __m128d          two     = _mm_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_pd(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     /* Setup water-specific parameters */
 122     inr              = nlist->iinr[0];
 123     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 124     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 125     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 126     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 127
 128     jq1              = _mm_set1_pd(charge[inr+1]);
 129     jq2              = _mm_set1_pd(charge[inr+2]);
 130     jq3              = _mm_set1_pd(charge[inr+3]);
 131     vdwjidx0A        = 2*vdwtype[inr+0];
 132     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 133     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 134     qq11             = _mm_mul_pd(iq1,jq1);
 135     qq12             = _mm_mul_pd(iq1,jq2);
 136     qq13             = _mm_mul_pd(iq1,jq3);
 137     qq21             = _mm_mul_pd(iq2,jq1);
 138     qq22             = _mm_mul_pd(iq2,jq2);
 139     qq23             = _mm_mul_pd(iq2,jq3);
 140     qq31             = _mm_mul_pd(iq3,jq1);
 141     qq32             = _mm_mul_pd(iq3,jq2);
 142     qq33             = _mm_mul_pd(iq3,jq3);
 143
 144     /* Avoid stupid compiler warnings */
 145     jnrA = jnrB = 0;
 146     j_coord_offsetA = 0;
 147     j_coord_offsetB = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 168                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 169
 170         fix0             = _mm_setzero_pd();
 171         fiy0             = _mm_setzero_pd();
 172         fiz0             = _mm_setzero_pd();
 173         fix1             = _mm_setzero_pd();
 174         fiy1             = _mm_setzero_pd();
 175         fiz1             = _mm_setzero_pd();
 176         fix2             = _mm_setzero_pd();
 177         fiy2             = _mm_setzero_pd();
 178         fiz2             = _mm_setzero_pd();
 179         fix3             = _mm_setzero_pd();
 180         fiy3             = _mm_setzero_pd();
 181         fiz3             = _mm_setzero_pd();
 182
 183         /* Reset potential sums */
 184         velecsum         = _mm_setzero_pd();
 185         vvdwsum          = _mm_setzero_pd();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 200                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm_sub_pd(ix0,jx0);
 204             dy00             = _mm_sub_pd(iy0,jy0);
 205             dz00             = _mm_sub_pd(iz0,jz0);
 206             dx11             = _mm_sub_pd(ix1,jx1);
 207             dy11             = _mm_sub_pd(iy1,jy1);
 208             dz11             = _mm_sub_pd(iz1,jz1);
 209             dx12             = _mm_sub_pd(ix1,jx2);
 210             dy12             = _mm_sub_pd(iy1,jy2);
 211             dz12             = _mm_sub_pd(iz1,jz2);
 212             dx13             = _mm_sub_pd(ix1,jx3);
 213             dy13             = _mm_sub_pd(iy1,jy3);
 214             dz13             = _mm_sub_pd(iz1,jz3);
 215             dx21             = _mm_sub_pd(ix2,jx1);
 216             dy21             = _mm_sub_pd(iy2,jy1);
 217             dz21             = _mm_sub_pd(iz2,jz1);
 218             dx22             = _mm_sub_pd(ix2,jx2);
 219             dy22             = _mm_sub_pd(iy2,jy2);
 220             dz22             = _mm_sub_pd(iz2,jz2);
 221             dx23             = _mm_sub_pd(ix2,jx3);
 222             dy23             = _mm_sub_pd(iy2,jy3);
 223             dz23             = _mm_sub_pd(iz2,jz3);
 224             dx31             = _mm_sub_pd(ix3,jx1);
 225             dy31             = _mm_sub_pd(iy3,jy1);
 226             dz31             = _mm_sub_pd(iz3,jz1);
 227             dx32             = _mm_sub_pd(ix3,jx2);
 228             dy32             = _mm_sub_pd(iy3,jy2);
 229             dz32             = _mm_sub_pd(iz3,jz2);
 230             dx33             = _mm_sub_pd(ix3,jx3);
 231             dy33             = _mm_sub_pd(iy3,jy3);
 232             dz33             = _mm_sub_pd(iz3,jz3);
 233
 234             /* Calculate squared distance and things based on it */
 235             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 236             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 237             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 238             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 239             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 240             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 241             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 242             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 243             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 244             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 245
 246             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 247             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 248             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 249             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 250             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 251             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 252             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 253             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 254             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 255
 256             rinvsq00         = gmx_mm_inv_pd(rsq00);
 257             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 258             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 259             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 260             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 261             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 262             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 263             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 264             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 265             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 266
 267             fjx0             = _mm_setzero_pd();
 268             fjy0             = _mm_setzero_pd();
 269             fjz0             = _mm_setzero_pd();
 270             fjx1             = _mm_setzero_pd();
 271             fjy1             = _mm_setzero_pd();
 272             fjz1             = _mm_setzero_pd();
 273             fjx2             = _mm_setzero_pd();
 274             fjy2             = _mm_setzero_pd();
 275             fjz2             = _mm_setzero_pd();
 276             fjx3             = _mm_setzero_pd();
 277             fjy3             = _mm_setzero_pd();
 278             fjz3             = _mm_setzero_pd();
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             /* LENNARD-JONES DISPERSION/REPULSION */
 285
 286             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 287             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 288             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 289             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 290             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 294
 295             fscal            = fvdw;
 296
 297             /* Calculate temporary vectorial force */
 298             tx               = _mm_mul_pd(fscal,dx00);
 299             ty               = _mm_mul_pd(fscal,dy00);
 300             tz               = _mm_mul_pd(fscal,dz00);
 301
 302             /* Update vectorial force */
 303             fix0             = _mm_add_pd(fix0,tx);
 304             fiy0             = _mm_add_pd(fiy0,ty);
 305             fiz0             = _mm_add_pd(fiz0,tz);
 306
 307             fjx0             = _mm_add_pd(fjx0,tx);
 308             fjy0             = _mm_add_pd(fjy0,ty);
 309             fjz0             = _mm_add_pd(fjz0,tz);
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* COULOMB ELECTROSTATICS */
 316             velec            = _mm_mul_pd(qq11,rinv11);
 317             felec            = _mm_mul_pd(velec,rinvsq11);
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             velecsum         = _mm_add_pd(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             /* Calculate temporary vectorial force */
 325             tx               = _mm_mul_pd(fscal,dx11);
 326             ty               = _mm_mul_pd(fscal,dy11);
 327             tz               = _mm_mul_pd(fscal,dz11);
 328
 329             /* Update vectorial force */
 330             fix1             = _mm_add_pd(fix1,tx);
 331             fiy1             = _mm_add_pd(fiy1,ty);
 332             fiz1             = _mm_add_pd(fiz1,tz);
 333
 334             fjx1             = _mm_add_pd(fjx1,tx);
 335             fjy1             = _mm_add_pd(fjy1,ty);
 336             fjz1             = _mm_add_pd(fjz1,tz);
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             /* COULOMB ELECTROSTATICS */
 343             velec            = _mm_mul_pd(qq12,rinv12);
 344             felec            = _mm_mul_pd(velec,rinvsq12);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velecsum         = _mm_add_pd(velecsum,velec);
 348
 349             fscal            = felec;
 350
 351             /* Calculate temporary vectorial force */
 352             tx               = _mm_mul_pd(fscal,dx12);
 353             ty               = _mm_mul_pd(fscal,dy12);
 354             tz               = _mm_mul_pd(fscal,dz12);
 355
 356             /* Update vectorial force */
 357             fix1             = _mm_add_pd(fix1,tx);
 358             fiy1             = _mm_add_pd(fiy1,ty);
 359             fiz1             = _mm_add_pd(fiz1,tz);
 360
 361             fjx2             = _mm_add_pd(fjx2,tx);
 362             fjy2             = _mm_add_pd(fjy2,ty);
 363             fjz2             = _mm_add_pd(fjz2,tz);
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             /* COULOMB ELECTROSTATICS */
 370             velec            = _mm_mul_pd(qq13,rinv13);
 371             felec            = _mm_mul_pd(velec,rinvsq13);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             velecsum         = _mm_add_pd(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm_mul_pd(fscal,dx13);
 380             ty               = _mm_mul_pd(fscal,dy13);
 381             tz               = _mm_mul_pd(fscal,dz13);
 382
 383             /* Update vectorial force */
 384             fix1             = _mm_add_pd(fix1,tx);
 385             fiy1             = _mm_add_pd(fiy1,ty);
 386             fiz1             = _mm_add_pd(fiz1,tz);
 387
 388             fjx3             = _mm_add_pd(fjx3,tx);
 389             fjy3             = _mm_add_pd(fjy3,ty);
 390             fjz3             = _mm_add_pd(fjz3,tz);
 391
 392             /**************************
 393              * CALCULATE INTERACTIONS *
 394              **************************/
 395
 396             /* COULOMB ELECTROSTATICS */
 397             velec            = _mm_mul_pd(qq21,rinv21);
 398             felec            = _mm_mul_pd(velec,rinvsq21);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velecsum         = _mm_add_pd(velecsum,velec);
 402
 403             fscal            = felec;
 404
 405             /* Calculate temporary vectorial force */
 406             tx               = _mm_mul_pd(fscal,dx21);
 407             ty               = _mm_mul_pd(fscal,dy21);
 408             tz               = _mm_mul_pd(fscal,dz21);
 409
 410             /* Update vectorial force */
 411             fix2             = _mm_add_pd(fix2,tx);
 412             fiy2             = _mm_add_pd(fiy2,ty);
 413             fiz2             = _mm_add_pd(fiz2,tz);
 414
 415             fjx1             = _mm_add_pd(fjx1,tx);
 416             fjy1             = _mm_add_pd(fjy1,ty);
 417             fjz1             = _mm_add_pd(fjz1,tz);
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = _mm_mul_pd(qq22,rinv22);
 425             felec            = _mm_mul_pd(velec,rinvsq22);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velecsum         = _mm_add_pd(velecsum,velec);
 429
 430             fscal            = felec;
 431
 432             /* Calculate temporary vectorial force */
 433             tx               = _mm_mul_pd(fscal,dx22);
 434             ty               = _mm_mul_pd(fscal,dy22);
 435             tz               = _mm_mul_pd(fscal,dz22);
 436
 437             /* Update vectorial force */
 438             fix2             = _mm_add_pd(fix2,tx);
 439             fiy2             = _mm_add_pd(fiy2,ty);
 440             fiz2             = _mm_add_pd(fiz2,tz);
 441
 442             fjx2             = _mm_add_pd(fjx2,tx);
 443             fjy2             = _mm_add_pd(fjy2,ty);
 444             fjz2             = _mm_add_pd(fjz2,tz);
 445
 446             /**************************
 447              * CALCULATE INTERACTIONS *
 448              **************************/
 449
 450             /* COULOMB ELECTROSTATICS */
 451             velec            = _mm_mul_pd(qq23,rinv23);
 452             felec            = _mm_mul_pd(velec,rinvsq23);
 453
 454             /* Update potential sum for this i atom from the interaction with this j atom. */
 455             velecsum         = _mm_add_pd(velecsum,velec);
 456
 457             fscal            = felec;
 458
 459             /* Calculate temporary vectorial force */
 460             tx               = _mm_mul_pd(fscal,dx23);
 461             ty               = _mm_mul_pd(fscal,dy23);
 462             tz               = _mm_mul_pd(fscal,dz23);
 463
 464             /* Update vectorial force */
 465             fix2             = _mm_add_pd(fix2,tx);
 466             fiy2             = _mm_add_pd(fiy2,ty);
 467             fiz2             = _mm_add_pd(fiz2,tz);
 468
 469             fjx3             = _mm_add_pd(fjx3,tx);
 470             fjy3             = _mm_add_pd(fjy3,ty);
 471             fjz3             = _mm_add_pd(fjz3,tz);
 472
 473             /**************************
 474              * CALCULATE INTERACTIONS *
 475              **************************/
 476
 477             /* COULOMB ELECTROSTATICS */
 478             velec            = _mm_mul_pd(qq31,rinv31);
 479             felec            = _mm_mul_pd(velec,rinvsq31);
 480
 481             /* Update potential sum for this i atom from the interaction with this j atom. */
 482             velecsum         = _mm_add_pd(velecsum,velec);
 483
 484             fscal            = felec;
 485
 486             /* Calculate temporary vectorial force */
 487             tx               = _mm_mul_pd(fscal,dx31);
 488             ty               = _mm_mul_pd(fscal,dy31);
 489             tz               = _mm_mul_pd(fscal,dz31);
 490
 491             /* Update vectorial force */
 492             fix3             = _mm_add_pd(fix3,tx);
 493             fiy3             = _mm_add_pd(fiy3,ty);
 494             fiz3             = _mm_add_pd(fiz3,tz);
 495
 496             fjx1             = _mm_add_pd(fjx1,tx);
 497             fjy1             = _mm_add_pd(fjy1,ty);
 498             fjz1             = _mm_add_pd(fjz1,tz);
 499
 500             /**************************
 501              * CALCULATE INTERACTIONS *
 502              **************************/
 503
 504             /* COULOMB ELECTROSTATICS */
 505             velec            = _mm_mul_pd(qq32,rinv32);
 506             felec            = _mm_mul_pd(velec,rinvsq32);
 507
 508             /* Update potential sum for this i atom from the interaction with this j atom. */
 509             velecsum         = _mm_add_pd(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             /* Calculate temporary vectorial force */
 514             tx               = _mm_mul_pd(fscal,dx32);
 515             ty               = _mm_mul_pd(fscal,dy32);
 516             tz               = _mm_mul_pd(fscal,dz32);
 517
 518             /* Update vectorial force */
 519             fix3             = _mm_add_pd(fix3,tx);
 520             fiy3             = _mm_add_pd(fiy3,ty);
 521             fiz3             = _mm_add_pd(fiz3,tz);
 522
 523             fjx2             = _mm_add_pd(fjx2,tx);
 524             fjy2             = _mm_add_pd(fjy2,ty);
 525             fjz2             = _mm_add_pd(fjz2,tz);
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             /* COULOMB ELECTROSTATICS */
 532             velec            = _mm_mul_pd(qq33,rinv33);
 533             felec            = _mm_mul_pd(velec,rinvsq33);
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velecsum         = _mm_add_pd(velecsum,velec);
 537
 538             fscal            = felec;
 539
 540             /* Calculate temporary vectorial force */
 541             tx               = _mm_mul_pd(fscal,dx33);
 542             ty               = _mm_mul_pd(fscal,dy33);
 543             tz               = _mm_mul_pd(fscal,dz33);
 544
 545             /* Update vectorial force */
 546             fix3             = _mm_add_pd(fix3,tx);
 547             fiy3             = _mm_add_pd(fiy3,ty);
 548             fiz3             = _mm_add_pd(fiz3,tz);
 549
 550             fjx3             = _mm_add_pd(fjx3,tx);
 551             fjy3             = _mm_add_pd(fjy3,ty);
 552             fjz3             = _mm_add_pd(fjz3,tz);
 553
 554             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 555
 556             /* Inner loop uses 287 flops */
 557         }
 558
 559         if(jidx<j_index_end)
 560         {
 561
 562             jnrA             = jjnr[jidx];
 563             j_coord_offsetA  = DIM*jnrA;
 564
 565             /* load j atom coordinates */
 566             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 567                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 568                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 569
 570             /* Calculate displacement vector */
 571             dx00             = _mm_sub_pd(ix0,jx0);
 572             dy00             = _mm_sub_pd(iy0,jy0);
 573             dz00             = _mm_sub_pd(iz0,jz0);
 574             dx11             = _mm_sub_pd(ix1,jx1);
 575             dy11             = _mm_sub_pd(iy1,jy1);
 576             dz11             = _mm_sub_pd(iz1,jz1);
 577             dx12             = _mm_sub_pd(ix1,jx2);
 578             dy12             = _mm_sub_pd(iy1,jy2);
 579             dz12             = _mm_sub_pd(iz1,jz2);
 580             dx13             = _mm_sub_pd(ix1,jx3);
 581             dy13             = _mm_sub_pd(iy1,jy3);
 582             dz13             = _mm_sub_pd(iz1,jz3);
 583             dx21             = _mm_sub_pd(ix2,jx1);
 584             dy21             = _mm_sub_pd(iy2,jy1);
 585             dz21             = _mm_sub_pd(iz2,jz1);
 586             dx22             = _mm_sub_pd(ix2,jx2);
 587             dy22             = _mm_sub_pd(iy2,jy2);
 588             dz22             = _mm_sub_pd(iz2,jz2);
 589             dx23             = _mm_sub_pd(ix2,jx3);
 590             dy23             = _mm_sub_pd(iy2,jy3);
 591             dz23             = _mm_sub_pd(iz2,jz3);
 592             dx31             = _mm_sub_pd(ix3,jx1);
 593             dy31             = _mm_sub_pd(iy3,jy1);
 594             dz31             = _mm_sub_pd(iz3,jz1);
 595             dx32             = _mm_sub_pd(ix3,jx2);
 596             dy32             = _mm_sub_pd(iy3,jy2);
 597             dz32             = _mm_sub_pd(iz3,jz2);
 598             dx33             = _mm_sub_pd(ix3,jx3);
 599             dy33             = _mm_sub_pd(iy3,jy3);
 600             dz33             = _mm_sub_pd(iz3,jz3);
 601
 602             /* Calculate squared distance and things based on it */
 603             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 604             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 605             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 606             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 607             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 608             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 609             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 610             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 611             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 612             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 613
 614             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 615             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 616             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 617             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 618             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 619             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 620             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 621             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 622             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 623
 624             rinvsq00         = gmx_mm_inv_pd(rsq00);
 625             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 626             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 627             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 628             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 629             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 630             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 631             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 632             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 633             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 634
 635             fjx0             = _mm_setzero_pd();
 636             fjy0             = _mm_setzero_pd();
 637             fjz0             = _mm_setzero_pd();
 638             fjx1             = _mm_setzero_pd();
 639             fjy1             = _mm_setzero_pd();
 640             fjz1             = _mm_setzero_pd();
 641             fjx2             = _mm_setzero_pd();
 642             fjy2             = _mm_setzero_pd();
 643             fjz2             = _mm_setzero_pd();
 644             fjx3             = _mm_setzero_pd();
 645             fjy3             = _mm_setzero_pd();
 646             fjz3             = _mm_setzero_pd();
 647
 648             /**************************
 649              * CALCULATE INTERACTIONS *
 650              **************************/
 651
 652             /* LENNARD-JONES DISPERSION/REPULSION */
 653
 654             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 655             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 656             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 657             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 658             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 659
 660             /* Update potential sum for this i atom from the interaction with this j atom. */
 661             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 662             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 663
 664             fscal            = fvdw;
 665
 666             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 667
 668             /* Calculate temporary vectorial force */
 669             tx               = _mm_mul_pd(fscal,dx00);
 670             ty               = _mm_mul_pd(fscal,dy00);
 671             tz               = _mm_mul_pd(fscal,dz00);
 672
 673             /* Update vectorial force */
 674             fix0             = _mm_add_pd(fix0,tx);
 675             fiy0             = _mm_add_pd(fiy0,ty);
 676             fiz0             = _mm_add_pd(fiz0,tz);
 677
 678             fjx0             = _mm_add_pd(fjx0,tx);
 679             fjy0             = _mm_add_pd(fjy0,ty);
 680             fjz0             = _mm_add_pd(fjz0,tz);
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             /* COULOMB ELECTROSTATICS */
 687             velec            = _mm_mul_pd(qq11,rinv11);
 688             felec            = _mm_mul_pd(velec,rinvsq11);
 689
 690             /* Update potential sum for this i atom from the interaction with this j atom. */
 691             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 692             velecsum         = _mm_add_pd(velecsum,velec);
 693
 694             fscal            = felec;
 695
 696             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 697
 698             /* Calculate temporary vectorial force */
 699             tx               = _mm_mul_pd(fscal,dx11);
 700             ty               = _mm_mul_pd(fscal,dy11);
 701             tz               = _mm_mul_pd(fscal,dz11);
 702
 703             /* Update vectorial force */
 704             fix1             = _mm_add_pd(fix1,tx);
 705             fiy1             = _mm_add_pd(fiy1,ty);
 706             fiz1             = _mm_add_pd(fiz1,tz);
 707
 708             fjx1             = _mm_add_pd(fjx1,tx);
 709             fjy1             = _mm_add_pd(fjy1,ty);
 710             fjz1             = _mm_add_pd(fjz1,tz);
 711
 712             /**************************
 713              * CALCULATE INTERACTIONS *
 714              **************************/
 715
 716             /* COULOMB ELECTROSTATICS */
 717             velec            = _mm_mul_pd(qq12,rinv12);
 718             felec            = _mm_mul_pd(velec,rinvsq12);
 719
 720             /* Update potential sum for this i atom from the interaction with this j atom. */
 721             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 722             velecsum         = _mm_add_pd(velecsum,velec);
 723
 724             fscal            = felec;
 725
 726             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 727
 728             /* Calculate temporary vectorial force */
 729             tx               = _mm_mul_pd(fscal,dx12);
 730             ty               = _mm_mul_pd(fscal,dy12);
 731             tz               = _mm_mul_pd(fscal,dz12);
 732
 733             /* Update vectorial force */
 734             fix1             = _mm_add_pd(fix1,tx);
 735             fiy1             = _mm_add_pd(fiy1,ty);
 736             fiz1             = _mm_add_pd(fiz1,tz);
 737
 738             fjx2             = _mm_add_pd(fjx2,tx);
 739             fjy2             = _mm_add_pd(fjy2,ty);
 740             fjz2             = _mm_add_pd(fjz2,tz);
 741
 742             /**************************
 743              * CALCULATE INTERACTIONS *
 744              **************************/
 745
 746             /* COULOMB ELECTROSTATICS */
 747             velec            = _mm_mul_pd(qq13,rinv13);
 748             felec            = _mm_mul_pd(velec,rinvsq13);
 749
 750             /* Update potential sum for this i atom from the interaction with this j atom. */
 751             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 752             velecsum         = _mm_add_pd(velecsum,velec);
 753
 754             fscal            = felec;
 755
 756             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 757
 758             /* Calculate temporary vectorial force */
 759             tx               = _mm_mul_pd(fscal,dx13);
 760             ty               = _mm_mul_pd(fscal,dy13);
 761             tz               = _mm_mul_pd(fscal,dz13);
 762
 763             /* Update vectorial force */
 764             fix1             = _mm_add_pd(fix1,tx);
 765             fiy1             = _mm_add_pd(fiy1,ty);
 766             fiz1             = _mm_add_pd(fiz1,tz);
 767
 768             fjx3             = _mm_add_pd(fjx3,tx);
 769             fjy3             = _mm_add_pd(fjy3,ty);
 770             fjz3             = _mm_add_pd(fjz3,tz);
 771
 772             /**************************
 773              * CALCULATE INTERACTIONS *
 774              **************************/
 775
 776             /* COULOMB ELECTROSTATICS */
 777             velec            = _mm_mul_pd(qq21,rinv21);
 778             felec            = _mm_mul_pd(velec,rinvsq21);
 779
 780             /* Update potential sum for this i atom from the interaction with this j atom. */
 781             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 782             velecsum         = _mm_add_pd(velecsum,velec);
 783
 784             fscal            = felec;
 785
 786             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 787
 788             /* Calculate temporary vectorial force */
 789             tx               = _mm_mul_pd(fscal,dx21);
 790             ty               = _mm_mul_pd(fscal,dy21);
 791             tz               = _mm_mul_pd(fscal,dz21);
 792
 793             /* Update vectorial force */
 794             fix2             = _mm_add_pd(fix2,tx);
 795             fiy2             = _mm_add_pd(fiy2,ty);
 796             fiz2             = _mm_add_pd(fiz2,tz);
 797
 798             fjx1             = _mm_add_pd(fjx1,tx);
 799             fjy1             = _mm_add_pd(fjy1,ty);
 800             fjz1             = _mm_add_pd(fjz1,tz);
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             /* COULOMB ELECTROSTATICS */
 807             velec            = _mm_mul_pd(qq22,rinv22);
 808             felec            = _mm_mul_pd(velec,rinvsq22);
 809
 810             /* Update potential sum for this i atom from the interaction with this j atom. */
 811             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 812             velecsum         = _mm_add_pd(velecsum,velec);
 813
 814             fscal            = felec;
 815
 816             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 817
 818             /* Calculate temporary vectorial force */
 819             tx               = _mm_mul_pd(fscal,dx22);
 820             ty               = _mm_mul_pd(fscal,dy22);
 821             tz               = _mm_mul_pd(fscal,dz22);
 822
 823             /* Update vectorial force */
 824             fix2             = _mm_add_pd(fix2,tx);
 825             fiy2             = _mm_add_pd(fiy2,ty);
 826             fiz2             = _mm_add_pd(fiz2,tz);
 827
 828             fjx2             = _mm_add_pd(fjx2,tx);
 829             fjy2             = _mm_add_pd(fjy2,ty);
 830             fjz2             = _mm_add_pd(fjz2,tz);
 831
 832             /**************************
 833              * CALCULATE INTERACTIONS *
 834              **************************/
 835
 836             /* COULOMB ELECTROSTATICS */
 837             velec            = _mm_mul_pd(qq23,rinv23);
 838             felec            = _mm_mul_pd(velec,rinvsq23);
 839
 840             /* Update potential sum for this i atom from the interaction with this j atom. */
 841             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 842             velecsum         = _mm_add_pd(velecsum,velec);
 843
 844             fscal            = felec;
 845
 846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 847
 848             /* Calculate temporary vectorial force */
 849             tx               = _mm_mul_pd(fscal,dx23);
 850             ty               = _mm_mul_pd(fscal,dy23);
 851             tz               = _mm_mul_pd(fscal,dz23);
 852
 853             /* Update vectorial force */
 854             fix2             = _mm_add_pd(fix2,tx);
 855             fiy2             = _mm_add_pd(fiy2,ty);
 856             fiz2             = _mm_add_pd(fiz2,tz);
 857
 858             fjx3             = _mm_add_pd(fjx3,tx);
 859             fjy3             = _mm_add_pd(fjy3,ty);
 860             fjz3             = _mm_add_pd(fjz3,tz);
 861
 862             /**************************
 863              * CALCULATE INTERACTIONS *
 864              **************************/
 865
 866             /* COULOMB ELECTROSTATICS */
 867             velec            = _mm_mul_pd(qq31,rinv31);
 868             felec            = _mm_mul_pd(velec,rinvsq31);
 869
 870             /* Update potential sum for this i atom from the interaction with this j atom. */
 871             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 872             velecsum         = _mm_add_pd(velecsum,velec);
 873
 874             fscal            = felec;
 875
 876             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 877
 878             /* Calculate temporary vectorial force */
 879             tx               = _mm_mul_pd(fscal,dx31);
 880             ty               = _mm_mul_pd(fscal,dy31);
 881             tz               = _mm_mul_pd(fscal,dz31);
 882
 883             /* Update vectorial force */
 884             fix3             = _mm_add_pd(fix3,tx);
 885             fiy3             = _mm_add_pd(fiy3,ty);
 886             fiz3             = _mm_add_pd(fiz3,tz);
 887
 888             fjx1             = _mm_add_pd(fjx1,tx);
 889             fjy1             = _mm_add_pd(fjy1,ty);
 890             fjz1             = _mm_add_pd(fjz1,tz);
 891
 892             /**************************
 893              * CALCULATE INTERACTIONS *
 894              **************************/
 895
 896             /* COULOMB ELECTROSTATICS */
 897             velec            = _mm_mul_pd(qq32,rinv32);
 898             felec            = _mm_mul_pd(velec,rinvsq32);
 899
 900             /* Update potential sum for this i atom from the interaction with this j atom. */
 901             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 902             velecsum         = _mm_add_pd(velecsum,velec);
 903
 904             fscal            = felec;
 905
 906             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 907
 908             /* Calculate temporary vectorial force */
 909             tx               = _mm_mul_pd(fscal,dx32);
 910             ty               = _mm_mul_pd(fscal,dy32);
 911             tz               = _mm_mul_pd(fscal,dz32);
 912
 913             /* Update vectorial force */
 914             fix3             = _mm_add_pd(fix3,tx);
 915             fiy3             = _mm_add_pd(fiy3,ty);
 916             fiz3             = _mm_add_pd(fiz3,tz);
 917
 918             fjx2             = _mm_add_pd(fjx2,tx);
 919             fjy2             = _mm_add_pd(fjy2,ty);
 920             fjz2             = _mm_add_pd(fjz2,tz);
 921
 922             /**************************
 923              * CALCULATE INTERACTIONS *
 924              **************************/
 925
 926             /* COULOMB ELECTROSTATICS */
 927             velec            = _mm_mul_pd(qq33,rinv33);
 928             felec            = _mm_mul_pd(velec,rinvsq33);
 929
 930             /* Update potential sum for this i atom from the interaction with this j atom. */
 931             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 932             velecsum         = _mm_add_pd(velecsum,velec);
 933
 934             fscal            = felec;
 935
 936             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 937
 938             /* Calculate temporary vectorial force */
 939             tx               = _mm_mul_pd(fscal,dx33);
 940             ty               = _mm_mul_pd(fscal,dy33);
 941             tz               = _mm_mul_pd(fscal,dz33);
 942
 943             /* Update vectorial force */
 944             fix3             = _mm_add_pd(fix3,tx);
 945             fiy3             = _mm_add_pd(fiy3,ty);
 946             fiz3             = _mm_add_pd(fiz3,tz);
 947
 948             fjx3             = _mm_add_pd(fjx3,tx);
 949             fjy3             = _mm_add_pd(fjy3,ty);
 950             fjz3             = _mm_add_pd(fjz3,tz);
 951
 952             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 953
 954             /* Inner loop uses 287 flops */
 955         }
 956
 957         /* End of innermost loop */
 958
 959         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 960                                               f+i_coord_offset,fshift+i_shift_offset);
 961
 962         ggid                        = gid[iidx];
 963         /* Update potential energies */
 964         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 965         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 966
 967         /* Increment number of inner iterations */
 968         inneriter                  += j_index_end - j_index_start;
 969
 970         /* Outer loop uses 26 flops */
 971     }
 972
 973     /* Increment number of outer iterations */
 974     outeriter        += nri;
 975
 976     /* Update outer/inner flops */
 977
 978     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*287);
 979 }
 980 /*
 981  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_double
 982  * Electrostatics interaction: Coulomb
 983  * VdW interaction:            LennardJones
 984  * Geometry:                   Water4-Water4
 985  * Calculate force/pot:        Force
 986  */
 987 void
 988 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_double
 989                     (t_nblist * gmx_restrict                nlist,
 990                      rvec * gmx_restrict                    xx,
 991                      rvec * gmx_restrict                    ff,
 992                      t_forcerec * gmx_restrict              fr,
 993                      t_mdatoms * gmx_restrict               mdatoms,
 994                      nb_kernel_data_t * gmx_restrict        kernel_data,
 995                      t_nrnb * gmx_restrict                  nrnb)
 996 {
 997     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 998      * just 0 for non-waters.
 999      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1000      * jnr indices corresponding to data put in the four positions in the SIMD register.
1001      */
1002     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1003     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1004     int              jnrA,jnrB;
1005     int              j_coord_offsetA,j_coord_offsetB;
1006     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1007     real             rcutoff_scalar;
1008     real             *shiftvec,*fshift,*x,*f;
1009     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1010     int              vdwioffset0;
1011     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1012     int              vdwioffset1;
1013     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1014     int              vdwioffset2;
1015     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1016     int              vdwioffset3;
1017     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1018     int              vdwjidx0A,vdwjidx0B;
1019     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1020     int              vdwjidx1A,vdwjidx1B;
1021     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1022     int              vdwjidx2A,vdwjidx2B;
1023     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1024     int              vdwjidx3A,vdwjidx3B;
1025     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1026     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1027     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1028     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1029     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1030     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1031     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1032     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1033     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1034     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1035     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1036     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1037     real             *charge;
1038     int              nvdwtype;
1039     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1040     int              *vdwtype;
1041     real             *vdwparam;
1042     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1043     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1044     __m128d          dummy_mask,cutoff_mask;
1045     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1046     __m128d          one     = _mm_set1_pd(1.0);
1047     __m128d          two     = _mm_set1_pd(2.0);
1048     x                = xx[0];
1049     f                = ff[0];
1050
1051     nri              = nlist->nri;
1052     iinr             = nlist->iinr;
1053     jindex           = nlist->jindex;
1054     jjnr             = nlist->jjnr;
1055     shiftidx         = nlist->shift;
1056     gid              = nlist->gid;
1057     shiftvec         = fr->shift_vec[0];
1058     fshift           = fr->fshift[0];
1059     facel            = _mm_set1_pd(fr->epsfac);
1060     charge           = mdatoms->chargeA;
1061     nvdwtype         = fr->ntype;
1062     vdwparam         = fr->nbfp;
1063     vdwtype          = mdatoms->typeA;
1064
1065     /* Setup water-specific parameters */
1066     inr              = nlist->iinr[0];
1067     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1068     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1069     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1070     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1071
1072     jq1              = _mm_set1_pd(charge[inr+1]);
1073     jq2              = _mm_set1_pd(charge[inr+2]);
1074     jq3              = _mm_set1_pd(charge[inr+3]);
1075     vdwjidx0A        = 2*vdwtype[inr+0];
1076     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1077     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1078     qq11             = _mm_mul_pd(iq1,jq1);
1079     qq12             = _mm_mul_pd(iq1,jq2);
1080     qq13             = _mm_mul_pd(iq1,jq3);
1081     qq21             = _mm_mul_pd(iq2,jq1);
1082     qq22             = _mm_mul_pd(iq2,jq2);
1083     qq23             = _mm_mul_pd(iq2,jq3);
1084     qq31             = _mm_mul_pd(iq3,jq1);
1085     qq32             = _mm_mul_pd(iq3,jq2);
1086     qq33             = _mm_mul_pd(iq3,jq3);
1087
1088     /* Avoid stupid compiler warnings */
1089     jnrA = jnrB = 0;
1090     j_coord_offsetA = 0;
1091     j_coord_offsetB = 0;
1092
1093     outeriter        = 0;
1094     inneriter        = 0;
1095
1096     /* Start outer loop over neighborlists */
1097     for(iidx=0; iidx<nri; iidx++)
1098     {
1099         /* Load shift vector for this list */
1100         i_shift_offset   = DIM*shiftidx[iidx];
1101
1102         /* Load limits for loop over neighbors */
1103         j_index_start    = jindex[iidx];
1104         j_index_end      = jindex[iidx+1];
1105
1106         /* Get outer coordinate index */
1107         inr              = iinr[iidx];
1108         i_coord_offset   = DIM*inr;
1109
1110         /* Load i particle coords and add shift vector */
1111         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1112                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1113
1114         fix0             = _mm_setzero_pd();
1115         fiy0             = _mm_setzero_pd();
1116         fiz0             = _mm_setzero_pd();
1117         fix1             = _mm_setzero_pd();
1118         fiy1             = _mm_setzero_pd();
1119         fiz1             = _mm_setzero_pd();
1120         fix2             = _mm_setzero_pd();
1121         fiy2             = _mm_setzero_pd();
1122         fiz2             = _mm_setzero_pd();
1123         fix3             = _mm_setzero_pd();
1124         fiy3             = _mm_setzero_pd();
1125         fiz3             = _mm_setzero_pd();
1126
1127         /* Start inner kernel loop */
1128         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1129         {
1130
1131             /* Get j neighbor index, and coordinate index */
1132             jnrA             = jjnr[jidx];
1133             jnrB             = jjnr[jidx+1];
1134             j_coord_offsetA  = DIM*jnrA;
1135             j_coord_offsetB  = DIM*jnrB;
1136
1137             /* load j atom coordinates */
1138             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1139                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1140                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1141
1142             /* Calculate displacement vector */
1143             dx00             = _mm_sub_pd(ix0,jx0);
1144             dy00             = _mm_sub_pd(iy0,jy0);
1145             dz00             = _mm_sub_pd(iz0,jz0);
1146             dx11             = _mm_sub_pd(ix1,jx1);
1147             dy11             = _mm_sub_pd(iy1,jy1);
1148             dz11             = _mm_sub_pd(iz1,jz1);
1149             dx12             = _mm_sub_pd(ix1,jx2);
1150             dy12             = _mm_sub_pd(iy1,jy2);
1151             dz12             = _mm_sub_pd(iz1,jz2);
1152             dx13             = _mm_sub_pd(ix1,jx3);
1153             dy13             = _mm_sub_pd(iy1,jy3);
1154             dz13             = _mm_sub_pd(iz1,jz3);
1155             dx21             = _mm_sub_pd(ix2,jx1);
1156             dy21             = _mm_sub_pd(iy2,jy1);
1157             dz21             = _mm_sub_pd(iz2,jz1);
1158             dx22             = _mm_sub_pd(ix2,jx2);
1159             dy22             = _mm_sub_pd(iy2,jy2);
1160             dz22             = _mm_sub_pd(iz2,jz2);
1161             dx23             = _mm_sub_pd(ix2,jx3);
1162             dy23             = _mm_sub_pd(iy2,jy3);
1163             dz23             = _mm_sub_pd(iz2,jz3);
1164             dx31             = _mm_sub_pd(ix3,jx1);
1165             dy31             = _mm_sub_pd(iy3,jy1);
1166             dz31             = _mm_sub_pd(iz3,jz1);
1167             dx32             = _mm_sub_pd(ix3,jx2);
1168             dy32             = _mm_sub_pd(iy3,jy2);
1169             dz32             = _mm_sub_pd(iz3,jz2);
1170             dx33             = _mm_sub_pd(ix3,jx3);
1171             dy33             = _mm_sub_pd(iy3,jy3);
1172             dz33             = _mm_sub_pd(iz3,jz3);
1173
1174             /* Calculate squared distance and things based on it */
1175             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1176             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1177             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1178             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1179             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1180             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1181             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1182             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1183             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1184             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1185
1186             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1187             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1188             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1189             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1190             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1191             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1192             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1193             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1194             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1195
1196             rinvsq00         = gmx_mm_inv_pd(rsq00);
1197             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1198             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1199             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1200             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1201             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1202             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1203             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1204             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1205             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1206
1207             fjx0             = _mm_setzero_pd();
1208             fjy0             = _mm_setzero_pd();
1209             fjz0             = _mm_setzero_pd();
1210             fjx1             = _mm_setzero_pd();
1211             fjy1             = _mm_setzero_pd();
1212             fjz1             = _mm_setzero_pd();
1213             fjx2             = _mm_setzero_pd();
1214             fjy2             = _mm_setzero_pd();
1215             fjz2             = _mm_setzero_pd();
1216             fjx3             = _mm_setzero_pd();
1217             fjy3             = _mm_setzero_pd();
1218             fjz3             = _mm_setzero_pd();
1219
1220             /**************************
1221              * CALCULATE INTERACTIONS *
1222              **************************/
1223
1224             /* LENNARD-JONES DISPERSION/REPULSION */
1225
1226             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1227             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1228
1229             fscal            = fvdw;
1230
1231             /* Calculate temporary vectorial force */
1232             tx               = _mm_mul_pd(fscal,dx00);
1233             ty               = _mm_mul_pd(fscal,dy00);
1234             tz               = _mm_mul_pd(fscal,dz00);
1235
1236             /* Update vectorial force */
1237             fix0             = _mm_add_pd(fix0,tx);
1238             fiy0             = _mm_add_pd(fiy0,ty);
1239             fiz0             = _mm_add_pd(fiz0,tz);
1240
1241             fjx0             = _mm_add_pd(fjx0,tx);
1242             fjy0             = _mm_add_pd(fjy0,ty);
1243             fjz0             = _mm_add_pd(fjz0,tz);
1244
1245             /**************************
1246              * CALCULATE INTERACTIONS *
1247              **************************/
1248
1249             /* COULOMB ELECTROSTATICS */
1250             velec            = _mm_mul_pd(qq11,rinv11);
1251             felec            = _mm_mul_pd(velec,rinvsq11);
1252
1253             fscal            = felec;
1254
1255             /* Calculate temporary vectorial force */
1256             tx               = _mm_mul_pd(fscal,dx11);
1257             ty               = _mm_mul_pd(fscal,dy11);
1258             tz               = _mm_mul_pd(fscal,dz11);
1259
1260             /* Update vectorial force */
1261             fix1             = _mm_add_pd(fix1,tx);
1262             fiy1             = _mm_add_pd(fiy1,ty);
1263             fiz1             = _mm_add_pd(fiz1,tz);
1264
1265             fjx1             = _mm_add_pd(fjx1,tx);
1266             fjy1             = _mm_add_pd(fjy1,ty);
1267             fjz1             = _mm_add_pd(fjz1,tz);
1268
1269             /**************************
1270              * CALCULATE INTERACTIONS *
1271              **************************/
1272
1273             /* COULOMB ELECTROSTATICS */
1274             velec            = _mm_mul_pd(qq12,rinv12);
1275             felec            = _mm_mul_pd(velec,rinvsq12);
1276
1277             fscal            = felec;
1278
1279             /* Calculate temporary vectorial force */
1280             tx               = _mm_mul_pd(fscal,dx12);
1281             ty               = _mm_mul_pd(fscal,dy12);
1282             tz               = _mm_mul_pd(fscal,dz12);
1283
1284             /* Update vectorial force */
1285             fix1             = _mm_add_pd(fix1,tx);
1286             fiy1             = _mm_add_pd(fiy1,ty);
1287             fiz1             = _mm_add_pd(fiz1,tz);
1288
1289             fjx2             = _mm_add_pd(fjx2,tx);
1290             fjy2             = _mm_add_pd(fjy2,ty);
1291             fjz2             = _mm_add_pd(fjz2,tz);
1292
1293             /**************************
1294              * CALCULATE INTERACTIONS *
1295              **************************/
1296
1297             /* COULOMB ELECTROSTATICS */
1298             velec            = _mm_mul_pd(qq13,rinv13);
1299             felec            = _mm_mul_pd(velec,rinvsq13);
1300
1301             fscal            = felec;
1302
1303             /* Calculate temporary vectorial force */
1304             tx               = _mm_mul_pd(fscal,dx13);
1305             ty               = _mm_mul_pd(fscal,dy13);
1306             tz               = _mm_mul_pd(fscal,dz13);
1307
1308             /* Update vectorial force */
1309             fix1             = _mm_add_pd(fix1,tx);
1310             fiy1             = _mm_add_pd(fiy1,ty);
1311             fiz1             = _mm_add_pd(fiz1,tz);
1312
1313             fjx3             = _mm_add_pd(fjx3,tx);
1314             fjy3             = _mm_add_pd(fjy3,ty);
1315             fjz3             = _mm_add_pd(fjz3,tz);
1316
1317             /**************************
1318              * CALCULATE INTERACTIONS *
1319              **************************/
1320
1321             /* COULOMB ELECTROSTATICS */
1322             velec            = _mm_mul_pd(qq21,rinv21);
1323             felec            = _mm_mul_pd(velec,rinvsq21);
1324
1325             fscal            = felec;
1326
1327             /* Calculate temporary vectorial force */
1328             tx               = _mm_mul_pd(fscal,dx21);
1329             ty               = _mm_mul_pd(fscal,dy21);
1330             tz               = _mm_mul_pd(fscal,dz21);
1331
1332             /* Update vectorial force */
1333             fix2             = _mm_add_pd(fix2,tx);
1334             fiy2             = _mm_add_pd(fiy2,ty);
1335             fiz2             = _mm_add_pd(fiz2,tz);
1336
1337             fjx1             = _mm_add_pd(fjx1,tx);
1338             fjy1             = _mm_add_pd(fjy1,ty);
1339             fjz1             = _mm_add_pd(fjz1,tz);
1340
1341             /**************************
1342              * CALCULATE INTERACTIONS *
1343              **************************/
1344
1345             /* COULOMB ELECTROSTATICS */
1346             velec            = _mm_mul_pd(qq22,rinv22);
1347             felec            = _mm_mul_pd(velec,rinvsq22);
1348
1349             fscal            = felec;
1350
1351             /* Calculate temporary vectorial force */
1352             tx               = _mm_mul_pd(fscal,dx22);
1353             ty               = _mm_mul_pd(fscal,dy22);
1354             tz               = _mm_mul_pd(fscal,dz22);
1355
1356             /* Update vectorial force */
1357             fix2             = _mm_add_pd(fix2,tx);
1358             fiy2             = _mm_add_pd(fiy2,ty);
1359             fiz2             = _mm_add_pd(fiz2,tz);
1360
1361             fjx2             = _mm_add_pd(fjx2,tx);
1362             fjy2             = _mm_add_pd(fjy2,ty);
1363             fjz2             = _mm_add_pd(fjz2,tz);
1364
1365             /**************************
1366              * CALCULATE INTERACTIONS *
1367              **************************/
1368
1369             /* COULOMB ELECTROSTATICS */
1370             velec            = _mm_mul_pd(qq23,rinv23);
1371             felec            = _mm_mul_pd(velec,rinvsq23);
1372
1373             fscal            = felec;
1374
1375             /* Calculate temporary vectorial force */
1376             tx               = _mm_mul_pd(fscal,dx23);
1377             ty               = _mm_mul_pd(fscal,dy23);
1378             tz               = _mm_mul_pd(fscal,dz23);
1379
1380             /* Update vectorial force */
1381             fix2             = _mm_add_pd(fix2,tx);
1382             fiy2             = _mm_add_pd(fiy2,ty);
1383             fiz2             = _mm_add_pd(fiz2,tz);
1384
1385             fjx3             = _mm_add_pd(fjx3,tx);
1386             fjy3             = _mm_add_pd(fjy3,ty);
1387             fjz3             = _mm_add_pd(fjz3,tz);
1388
1389             /**************************
1390              * CALCULATE INTERACTIONS *
1391              **************************/
1392
1393             /* COULOMB ELECTROSTATICS */
1394             velec            = _mm_mul_pd(qq31,rinv31);
1395             felec            = _mm_mul_pd(velec,rinvsq31);
1396
1397             fscal            = felec;
1398
1399             /* Calculate temporary vectorial force */
1400             tx               = _mm_mul_pd(fscal,dx31);
1401             ty               = _mm_mul_pd(fscal,dy31);
1402             tz               = _mm_mul_pd(fscal,dz31);
1403
1404             /* Update vectorial force */
1405             fix3             = _mm_add_pd(fix3,tx);
1406             fiy3             = _mm_add_pd(fiy3,ty);
1407             fiz3             = _mm_add_pd(fiz3,tz);
1408
1409             fjx1             = _mm_add_pd(fjx1,tx);
1410             fjy1             = _mm_add_pd(fjy1,ty);
1411             fjz1             = _mm_add_pd(fjz1,tz);
1412
1413             /**************************
1414              * CALCULATE INTERACTIONS *
1415              **************************/
1416
1417             /* COULOMB ELECTROSTATICS */
1418             velec            = _mm_mul_pd(qq32,rinv32);
1419             felec            = _mm_mul_pd(velec,rinvsq32);
1420
1421             fscal            = felec;
1422
1423             /* Calculate temporary vectorial force */
1424             tx               = _mm_mul_pd(fscal,dx32);
1425             ty               = _mm_mul_pd(fscal,dy32);
1426             tz               = _mm_mul_pd(fscal,dz32);
1427
1428             /* Update vectorial force */
1429             fix3             = _mm_add_pd(fix3,tx);
1430             fiy3             = _mm_add_pd(fiy3,ty);
1431             fiz3             = _mm_add_pd(fiz3,tz);
1432
1433             fjx2             = _mm_add_pd(fjx2,tx);
1434             fjy2             = _mm_add_pd(fjy2,ty);
1435             fjz2             = _mm_add_pd(fjz2,tz);
1436
1437             /**************************
1438              * CALCULATE INTERACTIONS *
1439              **************************/
1440
1441             /* COULOMB ELECTROSTATICS */
1442             velec            = _mm_mul_pd(qq33,rinv33);
1443             felec            = _mm_mul_pd(velec,rinvsq33);
1444
1445             fscal            = felec;
1446
1447             /* Calculate temporary vectorial force */
1448             tx               = _mm_mul_pd(fscal,dx33);
1449             ty               = _mm_mul_pd(fscal,dy33);
1450             tz               = _mm_mul_pd(fscal,dz33);
1451
1452             /* Update vectorial force */
1453             fix3             = _mm_add_pd(fix3,tx);
1454             fiy3             = _mm_add_pd(fiy3,ty);
1455             fiz3             = _mm_add_pd(fiz3,tz);
1456
1457             fjx3             = _mm_add_pd(fjx3,tx);
1458             fjy3             = _mm_add_pd(fjy3,ty);
1459             fjz3             = _mm_add_pd(fjz3,tz);
1460
1461             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1462
1463             /* Inner loop uses 273 flops */
1464         }
1465
1466         if(jidx<j_index_end)
1467         {
1468
1469             jnrA             = jjnr[jidx];
1470             j_coord_offsetA  = DIM*jnrA;
1471
1472             /* load j atom coordinates */
1473             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1474                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1475                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1476
1477             /* Calculate displacement vector */
1478             dx00             = _mm_sub_pd(ix0,jx0);
1479             dy00             = _mm_sub_pd(iy0,jy0);
1480             dz00             = _mm_sub_pd(iz0,jz0);
1481             dx11             = _mm_sub_pd(ix1,jx1);
1482             dy11             = _mm_sub_pd(iy1,jy1);
1483             dz11             = _mm_sub_pd(iz1,jz1);
1484             dx12             = _mm_sub_pd(ix1,jx2);
1485             dy12             = _mm_sub_pd(iy1,jy2);
1486             dz12             = _mm_sub_pd(iz1,jz2);
1487             dx13             = _mm_sub_pd(ix1,jx3);
1488             dy13             = _mm_sub_pd(iy1,jy3);
1489             dz13             = _mm_sub_pd(iz1,jz3);
1490             dx21             = _mm_sub_pd(ix2,jx1);
1491             dy21             = _mm_sub_pd(iy2,jy1);
1492             dz21             = _mm_sub_pd(iz2,jz1);
1493             dx22             = _mm_sub_pd(ix2,jx2);
1494             dy22             = _mm_sub_pd(iy2,jy2);
1495             dz22             = _mm_sub_pd(iz2,jz2);
1496             dx23             = _mm_sub_pd(ix2,jx3);
1497             dy23             = _mm_sub_pd(iy2,jy3);
1498             dz23             = _mm_sub_pd(iz2,jz3);
1499             dx31             = _mm_sub_pd(ix3,jx1);
1500             dy31             = _mm_sub_pd(iy3,jy1);
1501             dz31             = _mm_sub_pd(iz3,jz1);
1502             dx32             = _mm_sub_pd(ix3,jx2);
1503             dy32             = _mm_sub_pd(iy3,jy2);
1504             dz32             = _mm_sub_pd(iz3,jz2);
1505             dx33             = _mm_sub_pd(ix3,jx3);
1506             dy33             = _mm_sub_pd(iy3,jy3);
1507             dz33             = _mm_sub_pd(iz3,jz3);
1508
1509             /* Calculate squared distance and things based on it */
1510             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1511             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1512             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1513             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1514             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1515             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1516             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1517             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1518             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1519             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1520
1521             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1522             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1523             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1524             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1525             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1526             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1527             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1528             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1529             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1530
1531             rinvsq00         = gmx_mm_inv_pd(rsq00);
1532             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1533             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1534             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1535             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1536             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1537             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1538             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1539             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1540             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1541
1542             fjx0             = _mm_setzero_pd();
1543             fjy0             = _mm_setzero_pd();
1544             fjz0             = _mm_setzero_pd();
1545             fjx1             = _mm_setzero_pd();
1546             fjy1             = _mm_setzero_pd();
1547             fjz1             = _mm_setzero_pd();
1548             fjx2             = _mm_setzero_pd();
1549             fjy2             = _mm_setzero_pd();
1550             fjz2             = _mm_setzero_pd();
1551             fjx3             = _mm_setzero_pd();
1552             fjy3             = _mm_setzero_pd();
1553             fjz3             = _mm_setzero_pd();
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             /* LENNARD-JONES DISPERSION/REPULSION */
1560
1561             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1562             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1563
1564             fscal            = fvdw;
1565
1566             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1567
1568             /* Calculate temporary vectorial force */
1569             tx               = _mm_mul_pd(fscal,dx00);
1570             ty               = _mm_mul_pd(fscal,dy00);
1571             tz               = _mm_mul_pd(fscal,dz00);
1572
1573             /* Update vectorial force */
1574             fix0             = _mm_add_pd(fix0,tx);
1575             fiy0             = _mm_add_pd(fiy0,ty);
1576             fiz0             = _mm_add_pd(fiz0,tz);
1577
1578             fjx0             = _mm_add_pd(fjx0,tx);
1579             fjy0             = _mm_add_pd(fjy0,ty);
1580             fjz0             = _mm_add_pd(fjz0,tz);
1581
1582             /**************************
1583              * CALCULATE INTERACTIONS *
1584              **************************/
1585
1586             /* COULOMB ELECTROSTATICS */
1587             velec            = _mm_mul_pd(qq11,rinv11);
1588             felec            = _mm_mul_pd(velec,rinvsq11);
1589
1590             fscal            = felec;
1591
1592             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1593
1594             /* Calculate temporary vectorial force */
1595             tx               = _mm_mul_pd(fscal,dx11);
1596             ty               = _mm_mul_pd(fscal,dy11);
1597             tz               = _mm_mul_pd(fscal,dz11);
1598
1599             /* Update vectorial force */
1600             fix1             = _mm_add_pd(fix1,tx);
1601             fiy1             = _mm_add_pd(fiy1,ty);
1602             fiz1             = _mm_add_pd(fiz1,tz);
1603
1604             fjx1             = _mm_add_pd(fjx1,tx);
1605             fjy1             = _mm_add_pd(fjy1,ty);
1606             fjz1             = _mm_add_pd(fjz1,tz);
1607
1608             /**************************
1609              * CALCULATE INTERACTIONS *
1610              **************************/
1611
1612             /* COULOMB ELECTROSTATICS */
1613             velec            = _mm_mul_pd(qq12,rinv12);
1614             felec            = _mm_mul_pd(velec,rinvsq12);
1615
1616             fscal            = felec;
1617
1618             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1619
1620             /* Calculate temporary vectorial force */
1621             tx               = _mm_mul_pd(fscal,dx12);
1622             ty               = _mm_mul_pd(fscal,dy12);
1623             tz               = _mm_mul_pd(fscal,dz12);
1624
1625             /* Update vectorial force */
1626             fix1             = _mm_add_pd(fix1,tx);
1627             fiy1             = _mm_add_pd(fiy1,ty);
1628             fiz1             = _mm_add_pd(fiz1,tz);
1629
1630             fjx2             = _mm_add_pd(fjx2,tx);
1631             fjy2             = _mm_add_pd(fjy2,ty);
1632             fjz2             = _mm_add_pd(fjz2,tz);
1633
1634             /**************************
1635              * CALCULATE INTERACTIONS *
1636              **************************/
1637
1638             /* COULOMB ELECTROSTATICS */
1639             velec            = _mm_mul_pd(qq13,rinv13);
1640             felec            = _mm_mul_pd(velec,rinvsq13);
1641
1642             fscal            = felec;
1643
1644             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1645
1646             /* Calculate temporary vectorial force */
1647             tx               = _mm_mul_pd(fscal,dx13);
1648             ty               = _mm_mul_pd(fscal,dy13);
1649             tz               = _mm_mul_pd(fscal,dz13);
1650
1651             /* Update vectorial force */
1652             fix1             = _mm_add_pd(fix1,tx);
1653             fiy1             = _mm_add_pd(fiy1,ty);
1654             fiz1             = _mm_add_pd(fiz1,tz);
1655
1656             fjx3             = _mm_add_pd(fjx3,tx);
1657             fjy3             = _mm_add_pd(fjy3,ty);
1658             fjz3             = _mm_add_pd(fjz3,tz);
1659
1660             /**************************
1661              * CALCULATE INTERACTIONS *
1662              **************************/
1663
1664             /* COULOMB ELECTROSTATICS */
1665             velec            = _mm_mul_pd(qq21,rinv21);
1666             felec            = _mm_mul_pd(velec,rinvsq21);
1667
1668             fscal            = felec;
1669
1670             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1671
1672             /* Calculate temporary vectorial force */
1673             tx               = _mm_mul_pd(fscal,dx21);
1674             ty               = _mm_mul_pd(fscal,dy21);
1675             tz               = _mm_mul_pd(fscal,dz21);
1676
1677             /* Update vectorial force */
1678             fix2             = _mm_add_pd(fix2,tx);
1679             fiy2             = _mm_add_pd(fiy2,ty);
1680             fiz2             = _mm_add_pd(fiz2,tz);
1681
1682             fjx1             = _mm_add_pd(fjx1,tx);
1683             fjy1             = _mm_add_pd(fjy1,ty);
1684             fjz1             = _mm_add_pd(fjz1,tz);
1685
1686             /**************************
1687              * CALCULATE INTERACTIONS *
1688              **************************/
1689
1690             /* COULOMB ELECTROSTATICS */
1691             velec            = _mm_mul_pd(qq22,rinv22);
1692             felec            = _mm_mul_pd(velec,rinvsq22);
1693
1694             fscal            = felec;
1695
1696             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1697
1698             /* Calculate temporary vectorial force */
1699             tx               = _mm_mul_pd(fscal,dx22);
1700             ty               = _mm_mul_pd(fscal,dy22);
1701             tz               = _mm_mul_pd(fscal,dz22);
1702
1703             /* Update vectorial force */
1704             fix2             = _mm_add_pd(fix2,tx);
1705             fiy2             = _mm_add_pd(fiy2,ty);
1706             fiz2             = _mm_add_pd(fiz2,tz);
1707
1708             fjx2             = _mm_add_pd(fjx2,tx);
1709             fjy2             = _mm_add_pd(fjy2,ty);
1710             fjz2             = _mm_add_pd(fjz2,tz);
1711
1712             /**************************
1713              * CALCULATE INTERACTIONS *
1714              **************************/
1715
1716             /* COULOMB ELECTROSTATICS */
1717             velec            = _mm_mul_pd(qq23,rinv23);
1718             felec            = _mm_mul_pd(velec,rinvsq23);
1719
1720             fscal            = felec;
1721
1722             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1723
1724             /* Calculate temporary vectorial force */
1725             tx               = _mm_mul_pd(fscal,dx23);
1726             ty               = _mm_mul_pd(fscal,dy23);
1727             tz               = _mm_mul_pd(fscal,dz23);
1728
1729             /* Update vectorial force */
1730             fix2             = _mm_add_pd(fix2,tx);
1731             fiy2             = _mm_add_pd(fiy2,ty);
1732             fiz2             = _mm_add_pd(fiz2,tz);
1733
1734             fjx3             = _mm_add_pd(fjx3,tx);
1735             fjy3             = _mm_add_pd(fjy3,ty);
1736             fjz3             = _mm_add_pd(fjz3,tz);
1737
1738             /**************************
1739              * CALCULATE INTERACTIONS *
1740              **************************/
1741
1742             /* COULOMB ELECTROSTATICS */
1743             velec            = _mm_mul_pd(qq31,rinv31);
1744             felec            = _mm_mul_pd(velec,rinvsq31);
1745
1746             fscal            = felec;
1747
1748             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1749
1750             /* Calculate temporary vectorial force */
1751             tx               = _mm_mul_pd(fscal,dx31);
1752             ty               = _mm_mul_pd(fscal,dy31);
1753             tz               = _mm_mul_pd(fscal,dz31);
1754
1755             /* Update vectorial force */
1756             fix3             = _mm_add_pd(fix3,tx);
1757             fiy3             = _mm_add_pd(fiy3,ty);
1758             fiz3             = _mm_add_pd(fiz3,tz);
1759
1760             fjx1             = _mm_add_pd(fjx1,tx);
1761             fjy1             = _mm_add_pd(fjy1,ty);
1762             fjz1             = _mm_add_pd(fjz1,tz);
1763
1764             /**************************
1765              * CALCULATE INTERACTIONS *
1766              **************************/
1767
1768             /* COULOMB ELECTROSTATICS */
1769             velec            = _mm_mul_pd(qq32,rinv32);
1770             felec            = _mm_mul_pd(velec,rinvsq32);
1771
1772             fscal            = felec;
1773
1774             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1775
1776             /* Calculate temporary vectorial force */
1777             tx               = _mm_mul_pd(fscal,dx32);
1778             ty               = _mm_mul_pd(fscal,dy32);
1779             tz               = _mm_mul_pd(fscal,dz32);
1780
1781             /* Update vectorial force */
1782             fix3             = _mm_add_pd(fix3,tx);
1783             fiy3             = _mm_add_pd(fiy3,ty);
1784             fiz3             = _mm_add_pd(fiz3,tz);
1785
1786             fjx2             = _mm_add_pd(fjx2,tx);
1787             fjy2             = _mm_add_pd(fjy2,ty);
1788             fjz2             = _mm_add_pd(fjz2,tz);
1789
1790             /**************************
1791              * CALCULATE INTERACTIONS *
1792              **************************/
1793
1794             /* COULOMB ELECTROSTATICS */
1795             velec            = _mm_mul_pd(qq33,rinv33);
1796             felec            = _mm_mul_pd(velec,rinvsq33);
1797
1798             fscal            = felec;
1799
1800             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1801
1802             /* Calculate temporary vectorial force */
1803             tx               = _mm_mul_pd(fscal,dx33);
1804             ty               = _mm_mul_pd(fscal,dy33);
1805             tz               = _mm_mul_pd(fscal,dz33);
1806
1807             /* Update vectorial force */
1808             fix3             = _mm_add_pd(fix3,tx);
1809             fiy3             = _mm_add_pd(fiy3,ty);
1810             fiz3             = _mm_add_pd(fiz3,tz);
1811
1812             fjx3             = _mm_add_pd(fjx3,tx);
1813             fjy3             = _mm_add_pd(fjy3,ty);
1814             fjz3             = _mm_add_pd(fjz3,tz);
1815
1816             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1817
1818             /* Inner loop uses 273 flops */
1819         }
1820
1821         /* End of innermost loop */
1822
1823         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1824                                               f+i_coord_offset,fshift+i_shift_offset);
1825
1826         /* Increment number of inner iterations */
1827         inneriter                  += j_index_end - j_index_start;
1828
1829         /* Outer loop uses 24 flops */
1830     }
1831
1832     /* Increment number of outer iterations */
1833     outeriter        += nri;
1834
1835     /* Update outer/inner flops */
1836
1837     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1838 }