src/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRF_VdwLJ_GeomW4W4_sse2_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_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_sse2_double.h"
  34 #include "kernelutil_x86_sse2_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_sse2_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_sse2_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     krf              = _mm_set1_pd(fr->ic->k_rf);
 118     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 119     crf              = _mm_set1_pd(fr->ic->c_rf);
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 129     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 130
 131     jq1              = _mm_set1_pd(charge[inr+1]);
 132     jq2              = _mm_set1_pd(charge[inr+2]);
 133     jq3              = _mm_set1_pd(charge[inr+3]);
 134     vdwjidx0A        = 2*vdwtype[inr+0];
 135     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 136     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 137     qq11             = _mm_mul_pd(iq1,jq1);
 138     qq12             = _mm_mul_pd(iq1,jq2);
 139     qq13             = _mm_mul_pd(iq1,jq3);
 140     qq21             = _mm_mul_pd(iq2,jq1);
 141     qq22             = _mm_mul_pd(iq2,jq2);
 142     qq23             = _mm_mul_pd(iq2,jq3);
 143     qq31             = _mm_mul_pd(iq3,jq1);
 144     qq32             = _mm_mul_pd(iq3,jq2);
 145     qq33             = _mm_mul_pd(iq3,jq3);
 146
 147     /* Avoid stupid compiler warnings */
 148     jnrA = jnrB = 0;
 149     j_coord_offsetA = 0;
 150     j_coord_offsetB = 0;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     /* Start outer loop over neighborlists */
 156     for(iidx=0; iidx<nri; iidx++)
 157     {
 158         /* Load shift vector for this list */
 159         i_shift_offset   = DIM*shiftidx[iidx];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 171                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 172
 173         fix0             = _mm_setzero_pd();
 174         fiy0             = _mm_setzero_pd();
 175         fiz0             = _mm_setzero_pd();
 176         fix1             = _mm_setzero_pd();
 177         fiy1             = _mm_setzero_pd();
 178         fiz1             = _mm_setzero_pd();
 179         fix2             = _mm_setzero_pd();
 180         fiy2             = _mm_setzero_pd();
 181         fiz2             = _mm_setzero_pd();
 182         fix3             = _mm_setzero_pd();
 183         fiy3             = _mm_setzero_pd();
 184         fiz3             = _mm_setzero_pd();
 185
 186         /* Reset potential sums */
 187         velecsum         = _mm_setzero_pd();
 188         vvdwsum          = _mm_setzero_pd();
 189
 190         /* Start inner kernel loop */
 191         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 192         {
 193
 194             /* Get j neighbor index, and coordinate index */
 195             jnrA             = jjnr[jidx];
 196             jnrB             = jjnr[jidx+1];
 197             j_coord_offsetA  = DIM*jnrA;
 198             j_coord_offsetB  = DIM*jnrB;
 199
 200             /* load j atom coordinates */
 201             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 202                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 203                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 204
 205             /* Calculate displacement vector */
 206             dx00             = _mm_sub_pd(ix0,jx0);
 207             dy00             = _mm_sub_pd(iy0,jy0);
 208             dz00             = _mm_sub_pd(iz0,jz0);
 209             dx11             = _mm_sub_pd(ix1,jx1);
 210             dy11             = _mm_sub_pd(iy1,jy1);
 211             dz11             = _mm_sub_pd(iz1,jz1);
 212             dx12             = _mm_sub_pd(ix1,jx2);
 213             dy12             = _mm_sub_pd(iy1,jy2);
 214             dz12             = _mm_sub_pd(iz1,jz2);
 215             dx13             = _mm_sub_pd(ix1,jx3);
 216             dy13             = _mm_sub_pd(iy1,jy3);
 217             dz13             = _mm_sub_pd(iz1,jz3);
 218             dx21             = _mm_sub_pd(ix2,jx1);
 219             dy21             = _mm_sub_pd(iy2,jy1);
 220             dz21             = _mm_sub_pd(iz2,jz1);
 221             dx22             = _mm_sub_pd(ix2,jx2);
 222             dy22             = _mm_sub_pd(iy2,jy2);
 223             dz22             = _mm_sub_pd(iz2,jz2);
 224             dx23             = _mm_sub_pd(ix2,jx3);
 225             dy23             = _mm_sub_pd(iy2,jy3);
 226             dz23             = _mm_sub_pd(iz2,jz3);
 227             dx31             = _mm_sub_pd(ix3,jx1);
 228             dy31             = _mm_sub_pd(iy3,jy1);
 229             dz31             = _mm_sub_pd(iz3,jz1);
 230             dx32             = _mm_sub_pd(ix3,jx2);
 231             dy32             = _mm_sub_pd(iy3,jy2);
 232             dz32             = _mm_sub_pd(iz3,jz2);
 233             dx33             = _mm_sub_pd(ix3,jx3);
 234             dy33             = _mm_sub_pd(iy3,jy3);
 235             dz33             = _mm_sub_pd(iz3,jz3);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 239             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 240             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 241             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 242             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 243             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 244             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 245             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 246             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 247             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 248
 249             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 250             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 251             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 252             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 253             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 254             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 255             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 256             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 257             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 258
 259             rinvsq00         = gmx_mm_inv_pd(rsq00);
 260             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 261             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 262             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 263             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 264             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 265             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 266             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 267             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 268             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 269
 270             fjx0             = _mm_setzero_pd();
 271             fjy0             = _mm_setzero_pd();
 272             fjz0             = _mm_setzero_pd();
 273             fjx1             = _mm_setzero_pd();
 274             fjy1             = _mm_setzero_pd();
 275             fjz1             = _mm_setzero_pd();
 276             fjx2             = _mm_setzero_pd();
 277             fjy2             = _mm_setzero_pd();
 278             fjz2             = _mm_setzero_pd();
 279             fjx3             = _mm_setzero_pd();
 280             fjy3             = _mm_setzero_pd();
 281             fjz3             = _mm_setzero_pd();
 282
 283             /**************************
 284              * CALCULATE INTERACTIONS *
 285              **************************/
 286
 287             /* LENNARD-JONES DISPERSION/REPULSION */
 288
 289             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 290             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 291             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 292             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 293             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 294
 295             /* Update potential sum for this i atom from the interaction with this j atom. */
 296             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 297
 298             fscal            = fvdw;
 299
 300             /* Calculate temporary vectorial force */
 301             tx               = _mm_mul_pd(fscal,dx00);
 302             ty               = _mm_mul_pd(fscal,dy00);
 303             tz               = _mm_mul_pd(fscal,dz00);
 304
 305             /* Update vectorial force */
 306             fix0             = _mm_add_pd(fix0,tx);
 307             fiy0             = _mm_add_pd(fiy0,ty);
 308             fiz0             = _mm_add_pd(fiz0,tz);
 309
 310             fjx0             = _mm_add_pd(fjx0,tx);
 311             fjy0             = _mm_add_pd(fjy0,ty);
 312             fjz0             = _mm_add_pd(fjz0,tz);
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             /* REACTION-FIELD ELECTROSTATICS */
 319             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 320             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             velecsum         = _mm_add_pd(velecsum,velec);
 324
 325             fscal            = felec;
 326
 327             /* Calculate temporary vectorial force */
 328             tx               = _mm_mul_pd(fscal,dx11);
 329             ty               = _mm_mul_pd(fscal,dy11);
 330             tz               = _mm_mul_pd(fscal,dz11);
 331
 332             /* Update vectorial force */
 333             fix1             = _mm_add_pd(fix1,tx);
 334             fiy1             = _mm_add_pd(fiy1,ty);
 335             fiz1             = _mm_add_pd(fiz1,tz);
 336
 337             fjx1             = _mm_add_pd(fjx1,tx);
 338             fjy1             = _mm_add_pd(fjy1,ty);
 339             fjz1             = _mm_add_pd(fjz1,tz);
 340
 341             /**************************
 342              * CALCULATE INTERACTIONS *
 343              **************************/
 344
 345             /* REACTION-FIELD ELECTROSTATICS */
 346             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 347             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             velecsum         = _mm_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = _mm_mul_pd(fscal,dx12);
 356             ty               = _mm_mul_pd(fscal,dy12);
 357             tz               = _mm_mul_pd(fscal,dz12);
 358
 359             /* Update vectorial force */
 360             fix1             = _mm_add_pd(fix1,tx);
 361             fiy1             = _mm_add_pd(fiy1,ty);
 362             fiz1             = _mm_add_pd(fiz1,tz);
 363
 364             fjx2             = _mm_add_pd(fjx2,tx);
 365             fjy2             = _mm_add_pd(fjy2,ty);
 366             fjz2             = _mm_add_pd(fjz2,tz);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 374             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velecsum         = _mm_add_pd(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm_mul_pd(fscal,dx13);
 383             ty               = _mm_mul_pd(fscal,dy13);
 384             tz               = _mm_mul_pd(fscal,dz13);
 385
 386             /* Update vectorial force */
 387             fix1             = _mm_add_pd(fix1,tx);
 388             fiy1             = _mm_add_pd(fiy1,ty);
 389             fiz1             = _mm_add_pd(fiz1,tz);
 390
 391             fjx3             = _mm_add_pd(fjx3,tx);
 392             fjy3             = _mm_add_pd(fjy3,ty);
 393             fjz3             = _mm_add_pd(fjz3,tz);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* REACTION-FIELD ELECTROSTATICS */
 400             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 401             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 402
 403             /* Update potential sum for this i atom from the interaction with this j atom. */
 404             velecsum         = _mm_add_pd(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             /* Calculate temporary vectorial force */
 409             tx               = _mm_mul_pd(fscal,dx21);
 410             ty               = _mm_mul_pd(fscal,dy21);
 411             tz               = _mm_mul_pd(fscal,dz21);
 412
 413             /* Update vectorial force */
 414             fix2             = _mm_add_pd(fix2,tx);
 415             fiy2             = _mm_add_pd(fiy2,ty);
 416             fiz2             = _mm_add_pd(fiz2,tz);
 417
 418             fjx1             = _mm_add_pd(fjx1,tx);
 419             fjy1             = _mm_add_pd(fjy1,ty);
 420             fjz1             = _mm_add_pd(fjz1,tz);
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 428             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velecsum         = _mm_add_pd(velecsum,velec);
 432
 433             fscal            = felec;
 434
 435             /* Calculate temporary vectorial force */
 436             tx               = _mm_mul_pd(fscal,dx22);
 437             ty               = _mm_mul_pd(fscal,dy22);
 438             tz               = _mm_mul_pd(fscal,dz22);
 439
 440             /* Update vectorial force */
 441             fix2             = _mm_add_pd(fix2,tx);
 442             fiy2             = _mm_add_pd(fiy2,ty);
 443             fiz2             = _mm_add_pd(fiz2,tz);
 444
 445             fjx2             = _mm_add_pd(fjx2,tx);
 446             fjy2             = _mm_add_pd(fjy2,ty);
 447             fjz2             = _mm_add_pd(fjz2,tz);
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             /* REACTION-FIELD ELECTROSTATICS */
 454             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 455             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velecsum         = _mm_add_pd(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             /* Calculate temporary vectorial force */
 463             tx               = _mm_mul_pd(fscal,dx23);
 464             ty               = _mm_mul_pd(fscal,dy23);
 465             tz               = _mm_mul_pd(fscal,dz23);
 466
 467             /* Update vectorial force */
 468             fix2             = _mm_add_pd(fix2,tx);
 469             fiy2             = _mm_add_pd(fiy2,ty);
 470             fiz2             = _mm_add_pd(fiz2,tz);
 471
 472             fjx3             = _mm_add_pd(fjx3,tx);
 473             fjy3             = _mm_add_pd(fjy3,ty);
 474             fjz3             = _mm_add_pd(fjz3,tz);
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             /* REACTION-FIELD ELECTROSTATICS */
 481             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 482             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velecsum         = _mm_add_pd(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             /* Calculate temporary vectorial force */
 490             tx               = _mm_mul_pd(fscal,dx31);
 491             ty               = _mm_mul_pd(fscal,dy31);
 492             tz               = _mm_mul_pd(fscal,dz31);
 493
 494             /* Update vectorial force */
 495             fix3             = _mm_add_pd(fix3,tx);
 496             fiy3             = _mm_add_pd(fiy3,ty);
 497             fiz3             = _mm_add_pd(fiz3,tz);
 498
 499             fjx1             = _mm_add_pd(fjx1,tx);
 500             fjy1             = _mm_add_pd(fjy1,ty);
 501             fjz1             = _mm_add_pd(fjz1,tz);
 502
 503             /**************************
 504              * CALCULATE INTERACTIONS *
 505              **************************/
 506
 507             /* REACTION-FIELD ELECTROSTATICS */
 508             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 509             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 510
 511             /* Update potential sum for this i atom from the interaction with this j atom. */
 512             velecsum         = _mm_add_pd(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             /* Calculate temporary vectorial force */
 517             tx               = _mm_mul_pd(fscal,dx32);
 518             ty               = _mm_mul_pd(fscal,dy32);
 519             tz               = _mm_mul_pd(fscal,dz32);
 520
 521             /* Update vectorial force */
 522             fix3             = _mm_add_pd(fix3,tx);
 523             fiy3             = _mm_add_pd(fiy3,ty);
 524             fiz3             = _mm_add_pd(fiz3,tz);
 525
 526             fjx2             = _mm_add_pd(fjx2,tx);
 527             fjy2             = _mm_add_pd(fjy2,ty);
 528             fjz2             = _mm_add_pd(fjz2,tz);
 529
 530             /**************************
 531              * CALCULATE INTERACTIONS *
 532              **************************/
 533
 534             /* REACTION-FIELD ELECTROSTATICS */
 535             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 536             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velecsum         = _mm_add_pd(velecsum,velec);
 540
 541             fscal            = felec;
 542
 543             /* Calculate temporary vectorial force */
 544             tx               = _mm_mul_pd(fscal,dx33);
 545             ty               = _mm_mul_pd(fscal,dy33);
 546             tz               = _mm_mul_pd(fscal,dz33);
 547
 548             /* Update vectorial force */
 549             fix3             = _mm_add_pd(fix3,tx);
 550             fiy3             = _mm_add_pd(fiy3,ty);
 551             fiz3             = _mm_add_pd(fiz3,tz);
 552
 553             fjx3             = _mm_add_pd(fjx3,tx);
 554             fjy3             = _mm_add_pd(fjy3,ty);
 555             fjz3             = _mm_add_pd(fjz3,tz);
 556
 557             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);
 558
 559             /* Inner loop uses 323 flops */
 560         }
 561
 562         if(jidx<j_index_end)
 563         {
 564
 565             jnrA             = jjnr[jidx];
 566             j_coord_offsetA  = DIM*jnrA;
 567
 568             /* load j atom coordinates */
 569             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 570                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 571                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 572
 573             /* Calculate displacement vector */
 574             dx00             = _mm_sub_pd(ix0,jx0);
 575             dy00             = _mm_sub_pd(iy0,jy0);
 576             dz00             = _mm_sub_pd(iz0,jz0);
 577             dx11             = _mm_sub_pd(ix1,jx1);
 578             dy11             = _mm_sub_pd(iy1,jy1);
 579             dz11             = _mm_sub_pd(iz1,jz1);
 580             dx12             = _mm_sub_pd(ix1,jx2);
 581             dy12             = _mm_sub_pd(iy1,jy2);
 582             dz12             = _mm_sub_pd(iz1,jz2);
 583             dx13             = _mm_sub_pd(ix1,jx3);
 584             dy13             = _mm_sub_pd(iy1,jy3);
 585             dz13             = _mm_sub_pd(iz1,jz3);
 586             dx21             = _mm_sub_pd(ix2,jx1);
 587             dy21             = _mm_sub_pd(iy2,jy1);
 588             dz21             = _mm_sub_pd(iz2,jz1);
 589             dx22             = _mm_sub_pd(ix2,jx2);
 590             dy22             = _mm_sub_pd(iy2,jy2);
 591             dz22             = _mm_sub_pd(iz2,jz2);
 592             dx23             = _mm_sub_pd(ix2,jx3);
 593             dy23             = _mm_sub_pd(iy2,jy3);
 594             dz23             = _mm_sub_pd(iz2,jz3);
 595             dx31             = _mm_sub_pd(ix3,jx1);
 596             dy31             = _mm_sub_pd(iy3,jy1);
 597             dz31             = _mm_sub_pd(iz3,jz1);
 598             dx32             = _mm_sub_pd(ix3,jx2);
 599             dy32             = _mm_sub_pd(iy3,jy2);
 600             dz32             = _mm_sub_pd(iz3,jz2);
 601             dx33             = _mm_sub_pd(ix3,jx3);
 602             dy33             = _mm_sub_pd(iy3,jy3);
 603             dz33             = _mm_sub_pd(iz3,jz3);
 604
 605             /* Calculate squared distance and things based on it */
 606             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 607             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 608             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 609             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 610             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 611             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 612             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 613             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 614             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 615             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 616
 617             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 618             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 619             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 620             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 621             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 622             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 623             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 624             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 625             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 626
 627             rinvsq00         = gmx_mm_inv_pd(rsq00);
 628             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 629             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 630             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 631             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 632             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 633             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 634             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 635             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 636             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 637
 638             fjx0             = _mm_setzero_pd();
 639             fjy0             = _mm_setzero_pd();
 640             fjz0             = _mm_setzero_pd();
 641             fjx1             = _mm_setzero_pd();
 642             fjy1             = _mm_setzero_pd();
 643             fjz1             = _mm_setzero_pd();
 644             fjx2             = _mm_setzero_pd();
 645             fjy2             = _mm_setzero_pd();
 646             fjz2             = _mm_setzero_pd();
 647             fjx3             = _mm_setzero_pd();
 648             fjy3             = _mm_setzero_pd();
 649             fjz3             = _mm_setzero_pd();
 650
 651             /**************************
 652              * CALCULATE INTERACTIONS *
 653              **************************/
 654
 655             /* LENNARD-JONES DISPERSION/REPULSION */
 656
 657             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 658             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 659             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 660             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 661             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 662
 663             /* Update potential sum for this i atom from the interaction with this j atom. */
 664             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 665             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 666
 667             fscal            = fvdw;
 668
 669             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 670
 671             /* Calculate temporary vectorial force */
 672             tx               = _mm_mul_pd(fscal,dx00);
 673             ty               = _mm_mul_pd(fscal,dy00);
 674             tz               = _mm_mul_pd(fscal,dz00);
 675
 676             /* Update vectorial force */
 677             fix0             = _mm_add_pd(fix0,tx);
 678             fiy0             = _mm_add_pd(fiy0,ty);
 679             fiz0             = _mm_add_pd(fiz0,tz);
 680
 681             fjx0             = _mm_add_pd(fjx0,tx);
 682             fjy0             = _mm_add_pd(fjy0,ty);
 683             fjz0             = _mm_add_pd(fjz0,tz);
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             /* REACTION-FIELD ELECTROSTATICS */
 690             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 691             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 692
 693             /* Update potential sum for this i atom from the interaction with this j atom. */
 694             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 695             velecsum         = _mm_add_pd(velecsum,velec);
 696
 697             fscal            = felec;
 698
 699             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 700
 701             /* Calculate temporary vectorial force */
 702             tx               = _mm_mul_pd(fscal,dx11);
 703             ty               = _mm_mul_pd(fscal,dy11);
 704             tz               = _mm_mul_pd(fscal,dz11);
 705
 706             /* Update vectorial force */
 707             fix1             = _mm_add_pd(fix1,tx);
 708             fiy1             = _mm_add_pd(fiy1,ty);
 709             fiz1             = _mm_add_pd(fiz1,tz);
 710
 711             fjx1             = _mm_add_pd(fjx1,tx);
 712             fjy1             = _mm_add_pd(fjy1,ty);
 713             fjz1             = _mm_add_pd(fjz1,tz);
 714
 715             /**************************
 716              * CALCULATE INTERACTIONS *
 717              **************************/
 718
 719             /* REACTION-FIELD ELECTROSTATICS */
 720             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 721             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 722
 723             /* Update potential sum for this i atom from the interaction with this j atom. */
 724             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 725             velecsum         = _mm_add_pd(velecsum,velec);
 726
 727             fscal            = felec;
 728
 729             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 730
 731             /* Calculate temporary vectorial force */
 732             tx               = _mm_mul_pd(fscal,dx12);
 733             ty               = _mm_mul_pd(fscal,dy12);
 734             tz               = _mm_mul_pd(fscal,dz12);
 735
 736             /* Update vectorial force */
 737             fix1             = _mm_add_pd(fix1,tx);
 738             fiy1             = _mm_add_pd(fiy1,ty);
 739             fiz1             = _mm_add_pd(fiz1,tz);
 740
 741             fjx2             = _mm_add_pd(fjx2,tx);
 742             fjy2             = _mm_add_pd(fjy2,ty);
 743             fjz2             = _mm_add_pd(fjz2,tz);
 744
 745             /**************************
 746              * CALCULATE INTERACTIONS *
 747              **************************/
 748
 749             /* REACTION-FIELD ELECTROSTATICS */
 750             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 751             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 752
 753             /* Update potential sum for this i atom from the interaction with this j atom. */
 754             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 755             velecsum         = _mm_add_pd(velecsum,velec);
 756
 757             fscal            = felec;
 758
 759             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 760
 761             /* Calculate temporary vectorial force */
 762             tx               = _mm_mul_pd(fscal,dx13);
 763             ty               = _mm_mul_pd(fscal,dy13);
 764             tz               = _mm_mul_pd(fscal,dz13);
 765
 766             /* Update vectorial force */
 767             fix1             = _mm_add_pd(fix1,tx);
 768             fiy1             = _mm_add_pd(fiy1,ty);
 769             fiz1             = _mm_add_pd(fiz1,tz);
 770
 771             fjx3             = _mm_add_pd(fjx3,tx);
 772             fjy3             = _mm_add_pd(fjy3,ty);
 773             fjz3             = _mm_add_pd(fjz3,tz);
 774
 775             /**************************
 776              * CALCULATE INTERACTIONS *
 777              **************************/
 778
 779             /* REACTION-FIELD ELECTROSTATICS */
 780             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 781             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 782
 783             /* Update potential sum for this i atom from the interaction with this j atom. */
 784             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 785             velecsum         = _mm_add_pd(velecsum,velec);
 786
 787             fscal            = felec;
 788
 789             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 790
 791             /* Calculate temporary vectorial force */
 792             tx               = _mm_mul_pd(fscal,dx21);
 793             ty               = _mm_mul_pd(fscal,dy21);
 794             tz               = _mm_mul_pd(fscal,dz21);
 795
 796             /* Update vectorial force */
 797             fix2             = _mm_add_pd(fix2,tx);
 798             fiy2             = _mm_add_pd(fiy2,ty);
 799             fiz2             = _mm_add_pd(fiz2,tz);
 800
 801             fjx1             = _mm_add_pd(fjx1,tx);
 802             fjy1             = _mm_add_pd(fjy1,ty);
 803             fjz1             = _mm_add_pd(fjz1,tz);
 804
 805             /**************************
 806              * CALCULATE INTERACTIONS *
 807              **************************/
 808
 809             /* REACTION-FIELD ELECTROSTATICS */
 810             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 811             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 815             velecsum         = _mm_add_pd(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 820
 821             /* Calculate temporary vectorial force */
 822             tx               = _mm_mul_pd(fscal,dx22);
 823             ty               = _mm_mul_pd(fscal,dy22);
 824             tz               = _mm_mul_pd(fscal,dz22);
 825
 826             /* Update vectorial force */
 827             fix2             = _mm_add_pd(fix2,tx);
 828             fiy2             = _mm_add_pd(fiy2,ty);
 829             fiz2             = _mm_add_pd(fiz2,tz);
 830
 831             fjx2             = _mm_add_pd(fjx2,tx);
 832             fjy2             = _mm_add_pd(fjy2,ty);
 833             fjz2             = _mm_add_pd(fjz2,tz);
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             /* REACTION-FIELD ELECTROSTATICS */
 840             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 841             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 842
 843             /* Update potential sum for this i atom from the interaction with this j atom. */
 844             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 845             velecsum         = _mm_add_pd(velecsum,velec);
 846
 847             fscal            = felec;
 848
 849             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 850
 851             /* Calculate temporary vectorial force */
 852             tx               = _mm_mul_pd(fscal,dx23);
 853             ty               = _mm_mul_pd(fscal,dy23);
 854             tz               = _mm_mul_pd(fscal,dz23);
 855
 856             /* Update vectorial force */
 857             fix2             = _mm_add_pd(fix2,tx);
 858             fiy2             = _mm_add_pd(fiy2,ty);
 859             fiz2             = _mm_add_pd(fiz2,tz);
 860
 861             fjx3             = _mm_add_pd(fjx3,tx);
 862             fjy3             = _mm_add_pd(fjy3,ty);
 863             fjz3             = _mm_add_pd(fjz3,tz);
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             /* REACTION-FIELD ELECTROSTATICS */
 870             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 871             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 872
 873             /* Update potential sum for this i atom from the interaction with this j atom. */
 874             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 875             velecsum         = _mm_add_pd(velecsum,velec);
 876
 877             fscal            = felec;
 878
 879             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 880
 881             /* Calculate temporary vectorial force */
 882             tx               = _mm_mul_pd(fscal,dx31);
 883             ty               = _mm_mul_pd(fscal,dy31);
 884             tz               = _mm_mul_pd(fscal,dz31);
 885
 886             /* Update vectorial force */
 887             fix3             = _mm_add_pd(fix3,tx);
 888             fiy3             = _mm_add_pd(fiy3,ty);
 889             fiz3             = _mm_add_pd(fiz3,tz);
 890
 891             fjx1             = _mm_add_pd(fjx1,tx);
 892             fjy1             = _mm_add_pd(fjy1,ty);
 893             fjz1             = _mm_add_pd(fjz1,tz);
 894
 895             /**************************
 896              * CALCULATE INTERACTIONS *
 897              **************************/
 898
 899             /* REACTION-FIELD ELECTROSTATICS */
 900             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 901             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 902
 903             /* Update potential sum for this i atom from the interaction with this j atom. */
 904             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 905             velecsum         = _mm_add_pd(velecsum,velec);
 906
 907             fscal            = felec;
 908
 909             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 910
 911             /* Calculate temporary vectorial force */
 912             tx               = _mm_mul_pd(fscal,dx32);
 913             ty               = _mm_mul_pd(fscal,dy32);
 914             tz               = _mm_mul_pd(fscal,dz32);
 915
 916             /* Update vectorial force */
 917             fix3             = _mm_add_pd(fix3,tx);
 918             fiy3             = _mm_add_pd(fiy3,ty);
 919             fiz3             = _mm_add_pd(fiz3,tz);
 920
 921             fjx2             = _mm_add_pd(fjx2,tx);
 922             fjy2             = _mm_add_pd(fjy2,ty);
 923             fjz2             = _mm_add_pd(fjz2,tz);
 924
 925             /**************************
 926              * CALCULATE INTERACTIONS *
 927              **************************/
 928
 929             /* REACTION-FIELD ELECTROSTATICS */
 930             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 931             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 932
 933             /* Update potential sum for this i atom from the interaction with this j atom. */
 934             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 935             velecsum         = _mm_add_pd(velecsum,velec);
 936
 937             fscal            = felec;
 938
 939             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 940
 941             /* Calculate temporary vectorial force */
 942             tx               = _mm_mul_pd(fscal,dx33);
 943             ty               = _mm_mul_pd(fscal,dy33);
 944             tz               = _mm_mul_pd(fscal,dz33);
 945
 946             /* Update vectorial force */
 947             fix3             = _mm_add_pd(fix3,tx);
 948             fiy3             = _mm_add_pd(fiy3,ty);
 949             fiz3             = _mm_add_pd(fiz3,tz);
 950
 951             fjx3             = _mm_add_pd(fjx3,tx);
 952             fjy3             = _mm_add_pd(fjy3,ty);
 953             fjz3             = _mm_add_pd(fjz3,tz);
 954
 955             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 956
 957             /* Inner loop uses 323 flops */
 958         }
 959
 960         /* End of innermost loop */
 961
 962         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 963                                               f+i_coord_offset,fshift+i_shift_offset);
 964
 965         ggid                        = gid[iidx];
 966         /* Update potential energies */
 967         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 968         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 969
 970         /* Increment number of inner iterations */
 971         inneriter                  += j_index_end - j_index_start;
 972
 973         /* Outer loop uses 26 flops */
 974     }
 975
 976     /* Increment number of outer iterations */
 977     outeriter        += nri;
 978
 979     /* Update outer/inner flops */
 980
 981     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*323);
 982 }
 983 /*
 984  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_double
 985  * Electrostatics interaction: ReactionField
 986  * VdW interaction:            LennardJones
 987  * Geometry:                   Water4-Water4
 988  * Calculate force/pot:        Force
 989  */
 990 void
 991 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_sse2_double
 992                     (t_nblist * gmx_restrict                nlist,
 993                      rvec * gmx_restrict                    xx,
 994                      rvec * gmx_restrict                    ff,
 995                      t_forcerec * gmx_restrict              fr,
 996                      t_mdatoms * gmx_restrict               mdatoms,
 997                      nb_kernel_data_t * gmx_restrict        kernel_data,
 998                      t_nrnb * gmx_restrict                  nrnb)
 999 {
1000     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1001      * just 0 for non-waters.
1002      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1003      * jnr indices corresponding to data put in the four positions in the SIMD register.
1004      */
1005     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1006     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1007     int              jnrA,jnrB;
1008     int              j_coord_offsetA,j_coord_offsetB;
1009     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1010     real             rcutoff_scalar;
1011     real             *shiftvec,*fshift,*x,*f;
1012     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1013     int              vdwioffset0;
1014     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1015     int              vdwioffset1;
1016     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1017     int              vdwioffset2;
1018     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1019     int              vdwioffset3;
1020     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1021     int              vdwjidx0A,vdwjidx0B;
1022     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1023     int              vdwjidx1A,vdwjidx1B;
1024     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1025     int              vdwjidx2A,vdwjidx2B;
1026     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1027     int              vdwjidx3A,vdwjidx3B;
1028     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1029     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1030     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1031     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1032     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1033     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1034     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1035     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1036     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1037     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1038     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1039     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1040     real             *charge;
1041     int              nvdwtype;
1042     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1043     int              *vdwtype;
1044     real             *vdwparam;
1045     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1046     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1047     __m128d          dummy_mask,cutoff_mask;
1048     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1049     __m128d          one     = _mm_set1_pd(1.0);
1050     __m128d          two     = _mm_set1_pd(2.0);
1051     x                = xx[0];
1052     f                = ff[0];
1053
1054     nri              = nlist->nri;
1055     iinr             = nlist->iinr;
1056     jindex           = nlist->jindex;
1057     jjnr             = nlist->jjnr;
1058     shiftidx         = nlist->shift;
1059     gid              = nlist->gid;
1060     shiftvec         = fr->shift_vec[0];
1061     fshift           = fr->fshift[0];
1062     facel            = _mm_set1_pd(fr->epsfac);
1063     charge           = mdatoms->chargeA;
1064     krf              = _mm_set1_pd(fr->ic->k_rf);
1065     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1066     crf              = _mm_set1_pd(fr->ic->c_rf);
1067     nvdwtype         = fr->ntype;
1068     vdwparam         = fr->nbfp;
1069     vdwtype          = mdatoms->typeA;
1070
1071     /* Setup water-specific parameters */
1072     inr              = nlist->iinr[0];
1073     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1074     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1075     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1076     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1077
1078     jq1              = _mm_set1_pd(charge[inr+1]);
1079     jq2              = _mm_set1_pd(charge[inr+2]);
1080     jq3              = _mm_set1_pd(charge[inr+3]);
1081     vdwjidx0A        = 2*vdwtype[inr+0];
1082     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1083     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1084     qq11             = _mm_mul_pd(iq1,jq1);
1085     qq12             = _mm_mul_pd(iq1,jq2);
1086     qq13             = _mm_mul_pd(iq1,jq3);
1087     qq21             = _mm_mul_pd(iq2,jq1);
1088     qq22             = _mm_mul_pd(iq2,jq2);
1089     qq23             = _mm_mul_pd(iq2,jq3);
1090     qq31             = _mm_mul_pd(iq3,jq1);
1091     qq32             = _mm_mul_pd(iq3,jq2);
1092     qq33             = _mm_mul_pd(iq3,jq3);
1093
1094     /* Avoid stupid compiler warnings */
1095     jnrA = jnrB = 0;
1096     j_coord_offsetA = 0;
1097     j_coord_offsetB = 0;
1098
1099     outeriter        = 0;
1100     inneriter        = 0;
1101
1102     /* Start outer loop over neighborlists */
1103     for(iidx=0; iidx<nri; iidx++)
1104     {
1105         /* Load shift vector for this list */
1106         i_shift_offset   = DIM*shiftidx[iidx];
1107
1108         /* Load limits for loop over neighbors */
1109         j_index_start    = jindex[iidx];
1110         j_index_end      = jindex[iidx+1];
1111
1112         /* Get outer coordinate index */
1113         inr              = iinr[iidx];
1114         i_coord_offset   = DIM*inr;
1115
1116         /* Load i particle coords and add shift vector */
1117         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1118                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1119
1120         fix0             = _mm_setzero_pd();
1121         fiy0             = _mm_setzero_pd();
1122         fiz0             = _mm_setzero_pd();
1123         fix1             = _mm_setzero_pd();
1124         fiy1             = _mm_setzero_pd();
1125         fiz1             = _mm_setzero_pd();
1126         fix2             = _mm_setzero_pd();
1127         fiy2             = _mm_setzero_pd();
1128         fiz2             = _mm_setzero_pd();
1129         fix3             = _mm_setzero_pd();
1130         fiy3             = _mm_setzero_pd();
1131         fiz3             = _mm_setzero_pd();
1132
1133         /* Start inner kernel loop */
1134         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1135         {
1136
1137             /* Get j neighbor index, and coordinate index */
1138             jnrA             = jjnr[jidx];
1139             jnrB             = jjnr[jidx+1];
1140             j_coord_offsetA  = DIM*jnrA;
1141             j_coord_offsetB  = DIM*jnrB;
1142
1143             /* load j atom coordinates */
1144             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1145                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1146                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1147
1148             /* Calculate displacement vector */
1149             dx00             = _mm_sub_pd(ix0,jx0);
1150             dy00             = _mm_sub_pd(iy0,jy0);
1151             dz00             = _mm_sub_pd(iz0,jz0);
1152             dx11             = _mm_sub_pd(ix1,jx1);
1153             dy11             = _mm_sub_pd(iy1,jy1);
1154             dz11             = _mm_sub_pd(iz1,jz1);
1155             dx12             = _mm_sub_pd(ix1,jx2);
1156             dy12             = _mm_sub_pd(iy1,jy2);
1157             dz12             = _mm_sub_pd(iz1,jz2);
1158             dx13             = _mm_sub_pd(ix1,jx3);
1159             dy13             = _mm_sub_pd(iy1,jy3);
1160             dz13             = _mm_sub_pd(iz1,jz3);
1161             dx21             = _mm_sub_pd(ix2,jx1);
1162             dy21             = _mm_sub_pd(iy2,jy1);
1163             dz21             = _mm_sub_pd(iz2,jz1);
1164             dx22             = _mm_sub_pd(ix2,jx2);
1165             dy22             = _mm_sub_pd(iy2,jy2);
1166             dz22             = _mm_sub_pd(iz2,jz2);
1167             dx23             = _mm_sub_pd(ix2,jx3);
1168             dy23             = _mm_sub_pd(iy2,jy3);
1169             dz23             = _mm_sub_pd(iz2,jz3);
1170             dx31             = _mm_sub_pd(ix3,jx1);
1171             dy31             = _mm_sub_pd(iy3,jy1);
1172             dz31             = _mm_sub_pd(iz3,jz1);
1173             dx32             = _mm_sub_pd(ix3,jx2);
1174             dy32             = _mm_sub_pd(iy3,jy2);
1175             dz32             = _mm_sub_pd(iz3,jz2);
1176             dx33             = _mm_sub_pd(ix3,jx3);
1177             dy33             = _mm_sub_pd(iy3,jy3);
1178             dz33             = _mm_sub_pd(iz3,jz3);
1179
1180             /* Calculate squared distance and things based on it */
1181             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1182             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1183             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1184             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1185             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1186             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1187             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1188             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1189             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1190             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1191
1192             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1193             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1194             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1195             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1196             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1197             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1198             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1199             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1200             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1201
1202             rinvsq00         = gmx_mm_inv_pd(rsq00);
1203             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1204             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1205             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1206             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1207             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1208             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1209             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1210             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1211             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1212
1213             fjx0             = _mm_setzero_pd();
1214             fjy0             = _mm_setzero_pd();
1215             fjz0             = _mm_setzero_pd();
1216             fjx1             = _mm_setzero_pd();
1217             fjy1             = _mm_setzero_pd();
1218             fjz1             = _mm_setzero_pd();
1219             fjx2             = _mm_setzero_pd();
1220             fjy2             = _mm_setzero_pd();
1221             fjz2             = _mm_setzero_pd();
1222             fjx3             = _mm_setzero_pd();
1223             fjy3             = _mm_setzero_pd();
1224             fjz3             = _mm_setzero_pd();
1225
1226             /**************************
1227              * CALCULATE INTERACTIONS *
1228              **************************/
1229
1230             /* LENNARD-JONES DISPERSION/REPULSION */
1231
1232             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1233             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1234
1235             fscal            = fvdw;
1236
1237             /* Calculate temporary vectorial force */
1238             tx               = _mm_mul_pd(fscal,dx00);
1239             ty               = _mm_mul_pd(fscal,dy00);
1240             tz               = _mm_mul_pd(fscal,dz00);
1241
1242             /* Update vectorial force */
1243             fix0             = _mm_add_pd(fix0,tx);
1244             fiy0             = _mm_add_pd(fiy0,ty);
1245             fiz0             = _mm_add_pd(fiz0,tz);
1246
1247             fjx0             = _mm_add_pd(fjx0,tx);
1248             fjy0             = _mm_add_pd(fjy0,ty);
1249             fjz0             = _mm_add_pd(fjz0,tz);
1250
1251             /**************************
1252              * CALCULATE INTERACTIONS *
1253              **************************/
1254
1255             /* REACTION-FIELD ELECTROSTATICS */
1256             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1257
1258             fscal            = felec;
1259
1260             /* Calculate temporary vectorial force */
1261             tx               = _mm_mul_pd(fscal,dx11);
1262             ty               = _mm_mul_pd(fscal,dy11);
1263             tz               = _mm_mul_pd(fscal,dz11);
1264
1265             /* Update vectorial force */
1266             fix1             = _mm_add_pd(fix1,tx);
1267             fiy1             = _mm_add_pd(fiy1,ty);
1268             fiz1             = _mm_add_pd(fiz1,tz);
1269
1270             fjx1             = _mm_add_pd(fjx1,tx);
1271             fjy1             = _mm_add_pd(fjy1,ty);
1272             fjz1             = _mm_add_pd(fjz1,tz);
1273
1274             /**************************
1275              * CALCULATE INTERACTIONS *
1276              **************************/
1277
1278             /* REACTION-FIELD ELECTROSTATICS */
1279             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1280
1281             fscal            = felec;
1282
1283             /* Calculate temporary vectorial force */
1284             tx               = _mm_mul_pd(fscal,dx12);
1285             ty               = _mm_mul_pd(fscal,dy12);
1286             tz               = _mm_mul_pd(fscal,dz12);
1287
1288             /* Update vectorial force */
1289             fix1             = _mm_add_pd(fix1,tx);
1290             fiy1             = _mm_add_pd(fiy1,ty);
1291             fiz1             = _mm_add_pd(fiz1,tz);
1292
1293             fjx2             = _mm_add_pd(fjx2,tx);
1294             fjy2             = _mm_add_pd(fjy2,ty);
1295             fjz2             = _mm_add_pd(fjz2,tz);
1296
1297             /**************************
1298              * CALCULATE INTERACTIONS *
1299              **************************/
1300
1301             /* REACTION-FIELD ELECTROSTATICS */
1302             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1303
1304             fscal            = felec;
1305
1306             /* Calculate temporary vectorial force */
1307             tx               = _mm_mul_pd(fscal,dx13);
1308             ty               = _mm_mul_pd(fscal,dy13);
1309             tz               = _mm_mul_pd(fscal,dz13);
1310
1311             /* Update vectorial force */
1312             fix1             = _mm_add_pd(fix1,tx);
1313             fiy1             = _mm_add_pd(fiy1,ty);
1314             fiz1             = _mm_add_pd(fiz1,tz);
1315
1316             fjx3             = _mm_add_pd(fjx3,tx);
1317             fjy3             = _mm_add_pd(fjy3,ty);
1318             fjz3             = _mm_add_pd(fjz3,tz);
1319
1320             /**************************
1321              * CALCULATE INTERACTIONS *
1322              **************************/
1323
1324             /* REACTION-FIELD ELECTROSTATICS */
1325             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1326
1327             fscal            = felec;
1328
1329             /* Calculate temporary vectorial force */
1330             tx               = _mm_mul_pd(fscal,dx21);
1331             ty               = _mm_mul_pd(fscal,dy21);
1332             tz               = _mm_mul_pd(fscal,dz21);
1333
1334             /* Update vectorial force */
1335             fix2             = _mm_add_pd(fix2,tx);
1336             fiy2             = _mm_add_pd(fiy2,ty);
1337             fiz2             = _mm_add_pd(fiz2,tz);
1338
1339             fjx1             = _mm_add_pd(fjx1,tx);
1340             fjy1             = _mm_add_pd(fjy1,ty);
1341             fjz1             = _mm_add_pd(fjz1,tz);
1342
1343             /**************************
1344              * CALCULATE INTERACTIONS *
1345              **************************/
1346
1347             /* REACTION-FIELD ELECTROSTATICS */
1348             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1349
1350             fscal            = felec;
1351
1352             /* Calculate temporary vectorial force */
1353             tx               = _mm_mul_pd(fscal,dx22);
1354             ty               = _mm_mul_pd(fscal,dy22);
1355             tz               = _mm_mul_pd(fscal,dz22);
1356
1357             /* Update vectorial force */
1358             fix2             = _mm_add_pd(fix2,tx);
1359             fiy2             = _mm_add_pd(fiy2,ty);
1360             fiz2             = _mm_add_pd(fiz2,tz);
1361
1362             fjx2             = _mm_add_pd(fjx2,tx);
1363             fjy2             = _mm_add_pd(fjy2,ty);
1364             fjz2             = _mm_add_pd(fjz2,tz);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* REACTION-FIELD ELECTROSTATICS */
1371             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
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             /* REACTION-FIELD ELECTROSTATICS */
1394             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1395
1396             fscal            = felec;
1397
1398             /* Calculate temporary vectorial force */
1399             tx               = _mm_mul_pd(fscal,dx31);
1400             ty               = _mm_mul_pd(fscal,dy31);
1401             tz               = _mm_mul_pd(fscal,dz31);
1402
1403             /* Update vectorial force */
1404             fix3             = _mm_add_pd(fix3,tx);
1405             fiy3             = _mm_add_pd(fiy3,ty);
1406             fiz3             = _mm_add_pd(fiz3,tz);
1407
1408             fjx1             = _mm_add_pd(fjx1,tx);
1409             fjy1             = _mm_add_pd(fjy1,ty);
1410             fjz1             = _mm_add_pd(fjz1,tz);
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             /* REACTION-FIELD ELECTROSTATICS */
1417             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1418
1419             fscal            = felec;
1420
1421             /* Calculate temporary vectorial force */
1422             tx               = _mm_mul_pd(fscal,dx32);
1423             ty               = _mm_mul_pd(fscal,dy32);
1424             tz               = _mm_mul_pd(fscal,dz32);
1425
1426             /* Update vectorial force */
1427             fix3             = _mm_add_pd(fix3,tx);
1428             fiy3             = _mm_add_pd(fiy3,ty);
1429             fiz3             = _mm_add_pd(fiz3,tz);
1430
1431             fjx2             = _mm_add_pd(fjx2,tx);
1432             fjy2             = _mm_add_pd(fjy2,ty);
1433             fjz2             = _mm_add_pd(fjz2,tz);
1434
1435             /**************************
1436              * CALCULATE INTERACTIONS *
1437              **************************/
1438
1439             /* REACTION-FIELD ELECTROSTATICS */
1440             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1441
1442             fscal            = felec;
1443
1444             /* Calculate temporary vectorial force */
1445             tx               = _mm_mul_pd(fscal,dx33);
1446             ty               = _mm_mul_pd(fscal,dy33);
1447             tz               = _mm_mul_pd(fscal,dz33);
1448
1449             /* Update vectorial force */
1450             fix3             = _mm_add_pd(fix3,tx);
1451             fiy3             = _mm_add_pd(fiy3,ty);
1452             fiz3             = _mm_add_pd(fiz3,tz);
1453
1454             fjx3             = _mm_add_pd(fjx3,tx);
1455             fjy3             = _mm_add_pd(fjy3,ty);
1456             fjz3             = _mm_add_pd(fjz3,tz);
1457
1458             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);
1459
1460             /* Inner loop uses 273 flops */
1461         }
1462
1463         if(jidx<j_index_end)
1464         {
1465
1466             jnrA             = jjnr[jidx];
1467             j_coord_offsetA  = DIM*jnrA;
1468
1469             /* load j atom coordinates */
1470             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1471                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1472                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1473
1474             /* Calculate displacement vector */
1475             dx00             = _mm_sub_pd(ix0,jx0);
1476             dy00             = _mm_sub_pd(iy0,jy0);
1477             dz00             = _mm_sub_pd(iz0,jz0);
1478             dx11             = _mm_sub_pd(ix1,jx1);
1479             dy11             = _mm_sub_pd(iy1,jy1);
1480             dz11             = _mm_sub_pd(iz1,jz1);
1481             dx12             = _mm_sub_pd(ix1,jx2);
1482             dy12             = _mm_sub_pd(iy1,jy2);
1483             dz12             = _mm_sub_pd(iz1,jz2);
1484             dx13             = _mm_sub_pd(ix1,jx3);
1485             dy13             = _mm_sub_pd(iy1,jy3);
1486             dz13             = _mm_sub_pd(iz1,jz3);
1487             dx21             = _mm_sub_pd(ix2,jx1);
1488             dy21             = _mm_sub_pd(iy2,jy1);
1489             dz21             = _mm_sub_pd(iz2,jz1);
1490             dx22             = _mm_sub_pd(ix2,jx2);
1491             dy22             = _mm_sub_pd(iy2,jy2);
1492             dz22             = _mm_sub_pd(iz2,jz2);
1493             dx23             = _mm_sub_pd(ix2,jx3);
1494             dy23             = _mm_sub_pd(iy2,jy3);
1495             dz23             = _mm_sub_pd(iz2,jz3);
1496             dx31             = _mm_sub_pd(ix3,jx1);
1497             dy31             = _mm_sub_pd(iy3,jy1);
1498             dz31             = _mm_sub_pd(iz3,jz1);
1499             dx32             = _mm_sub_pd(ix3,jx2);
1500             dy32             = _mm_sub_pd(iy3,jy2);
1501             dz32             = _mm_sub_pd(iz3,jz2);
1502             dx33             = _mm_sub_pd(ix3,jx3);
1503             dy33             = _mm_sub_pd(iy3,jy3);
1504             dz33             = _mm_sub_pd(iz3,jz3);
1505
1506             /* Calculate squared distance and things based on it */
1507             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1508             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1509             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1510             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1511             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1512             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1513             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1514             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1515             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1516             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1517
1518             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1519             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1520             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1521             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1522             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1523             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1524             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1525             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1526             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1527
1528             rinvsq00         = gmx_mm_inv_pd(rsq00);
1529             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1530             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1531             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1532             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1533             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1534             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1535             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1536             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1537             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1538
1539             fjx0             = _mm_setzero_pd();
1540             fjy0             = _mm_setzero_pd();
1541             fjz0             = _mm_setzero_pd();
1542             fjx1             = _mm_setzero_pd();
1543             fjy1             = _mm_setzero_pd();
1544             fjz1             = _mm_setzero_pd();
1545             fjx2             = _mm_setzero_pd();
1546             fjy2             = _mm_setzero_pd();
1547             fjz2             = _mm_setzero_pd();
1548             fjx3             = _mm_setzero_pd();
1549             fjy3             = _mm_setzero_pd();
1550             fjz3             = _mm_setzero_pd();
1551
1552             /**************************
1553              * CALCULATE INTERACTIONS *
1554              **************************/
1555
1556             /* LENNARD-JONES DISPERSION/REPULSION */
1557
1558             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1559             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1560
1561             fscal            = fvdw;
1562
1563             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1564
1565             /* Calculate temporary vectorial force */
1566             tx               = _mm_mul_pd(fscal,dx00);
1567             ty               = _mm_mul_pd(fscal,dy00);
1568             tz               = _mm_mul_pd(fscal,dz00);
1569
1570             /* Update vectorial force */
1571             fix0             = _mm_add_pd(fix0,tx);
1572             fiy0             = _mm_add_pd(fiy0,ty);
1573             fiz0             = _mm_add_pd(fiz0,tz);
1574
1575             fjx0             = _mm_add_pd(fjx0,tx);
1576             fjy0             = _mm_add_pd(fjy0,ty);
1577             fjz0             = _mm_add_pd(fjz0,tz);
1578
1579             /**************************
1580              * CALCULATE INTERACTIONS *
1581              **************************/
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1585
1586             fscal            = felec;
1587
1588             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1589
1590             /* Calculate temporary vectorial force */
1591             tx               = _mm_mul_pd(fscal,dx11);
1592             ty               = _mm_mul_pd(fscal,dy11);
1593             tz               = _mm_mul_pd(fscal,dz11);
1594
1595             /* Update vectorial force */
1596             fix1             = _mm_add_pd(fix1,tx);
1597             fiy1             = _mm_add_pd(fiy1,ty);
1598             fiz1             = _mm_add_pd(fiz1,tz);
1599
1600             fjx1             = _mm_add_pd(fjx1,tx);
1601             fjy1             = _mm_add_pd(fjy1,ty);
1602             fjz1             = _mm_add_pd(fjz1,tz);
1603
1604             /**************************
1605              * CALCULATE INTERACTIONS *
1606              **************************/
1607
1608             /* REACTION-FIELD ELECTROSTATICS */
1609             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1610
1611             fscal            = felec;
1612
1613             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1614
1615             /* Calculate temporary vectorial force */
1616             tx               = _mm_mul_pd(fscal,dx12);
1617             ty               = _mm_mul_pd(fscal,dy12);
1618             tz               = _mm_mul_pd(fscal,dz12);
1619
1620             /* Update vectorial force */
1621             fix1             = _mm_add_pd(fix1,tx);
1622             fiy1             = _mm_add_pd(fiy1,ty);
1623             fiz1             = _mm_add_pd(fiz1,tz);
1624
1625             fjx2             = _mm_add_pd(fjx2,tx);
1626             fjy2             = _mm_add_pd(fjy2,ty);
1627             fjz2             = _mm_add_pd(fjz2,tz);
1628
1629             /**************************
1630              * CALCULATE INTERACTIONS *
1631              **************************/
1632
1633             /* REACTION-FIELD ELECTROSTATICS */
1634             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1635
1636             fscal            = felec;
1637
1638             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1639
1640             /* Calculate temporary vectorial force */
1641             tx               = _mm_mul_pd(fscal,dx13);
1642             ty               = _mm_mul_pd(fscal,dy13);
1643             tz               = _mm_mul_pd(fscal,dz13);
1644
1645             /* Update vectorial force */
1646             fix1             = _mm_add_pd(fix1,tx);
1647             fiy1             = _mm_add_pd(fiy1,ty);
1648             fiz1             = _mm_add_pd(fiz1,tz);
1649
1650             fjx3             = _mm_add_pd(fjx3,tx);
1651             fjy3             = _mm_add_pd(fjy3,ty);
1652             fjz3             = _mm_add_pd(fjz3,tz);
1653
1654             /**************************
1655              * CALCULATE INTERACTIONS *
1656              **************************/
1657
1658             /* REACTION-FIELD ELECTROSTATICS */
1659             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1660
1661             fscal            = felec;
1662
1663             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1664
1665             /* Calculate temporary vectorial force */
1666             tx               = _mm_mul_pd(fscal,dx21);
1667             ty               = _mm_mul_pd(fscal,dy21);
1668             tz               = _mm_mul_pd(fscal,dz21);
1669
1670             /* Update vectorial force */
1671             fix2             = _mm_add_pd(fix2,tx);
1672             fiy2             = _mm_add_pd(fiy2,ty);
1673             fiz2             = _mm_add_pd(fiz2,tz);
1674
1675             fjx1             = _mm_add_pd(fjx1,tx);
1676             fjy1             = _mm_add_pd(fjy1,ty);
1677             fjz1             = _mm_add_pd(fjz1,tz);
1678
1679             /**************************
1680              * CALCULATE INTERACTIONS *
1681              **************************/
1682
1683             /* REACTION-FIELD ELECTROSTATICS */
1684             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1685
1686             fscal            = felec;
1687
1688             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1689
1690             /* Calculate temporary vectorial force */
1691             tx               = _mm_mul_pd(fscal,dx22);
1692             ty               = _mm_mul_pd(fscal,dy22);
1693             tz               = _mm_mul_pd(fscal,dz22);
1694
1695             /* Update vectorial force */
1696             fix2             = _mm_add_pd(fix2,tx);
1697             fiy2             = _mm_add_pd(fiy2,ty);
1698             fiz2             = _mm_add_pd(fiz2,tz);
1699
1700             fjx2             = _mm_add_pd(fjx2,tx);
1701             fjy2             = _mm_add_pd(fjy2,ty);
1702             fjz2             = _mm_add_pd(fjz2,tz);
1703
1704             /**************************
1705              * CALCULATE INTERACTIONS *
1706              **************************/
1707
1708             /* REACTION-FIELD ELECTROSTATICS */
1709             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1710
1711             fscal            = felec;
1712
1713             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1714
1715             /* Calculate temporary vectorial force */
1716             tx               = _mm_mul_pd(fscal,dx23);
1717             ty               = _mm_mul_pd(fscal,dy23);
1718             tz               = _mm_mul_pd(fscal,dz23);
1719
1720             /* Update vectorial force */
1721             fix2             = _mm_add_pd(fix2,tx);
1722             fiy2             = _mm_add_pd(fiy2,ty);
1723             fiz2             = _mm_add_pd(fiz2,tz);
1724
1725             fjx3             = _mm_add_pd(fjx3,tx);
1726             fjy3             = _mm_add_pd(fjy3,ty);
1727             fjz3             = _mm_add_pd(fjz3,tz);
1728
1729             /**************************
1730              * CALCULATE INTERACTIONS *
1731              **************************/
1732
1733             /* REACTION-FIELD ELECTROSTATICS */
1734             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1735
1736             fscal            = felec;
1737
1738             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1739
1740             /* Calculate temporary vectorial force */
1741             tx               = _mm_mul_pd(fscal,dx31);
1742             ty               = _mm_mul_pd(fscal,dy31);
1743             tz               = _mm_mul_pd(fscal,dz31);
1744
1745             /* Update vectorial force */
1746             fix3             = _mm_add_pd(fix3,tx);
1747             fiy3             = _mm_add_pd(fiy3,ty);
1748             fiz3             = _mm_add_pd(fiz3,tz);
1749
1750             fjx1             = _mm_add_pd(fjx1,tx);
1751             fjy1             = _mm_add_pd(fjy1,ty);
1752             fjz1             = _mm_add_pd(fjz1,tz);
1753
1754             /**************************
1755              * CALCULATE INTERACTIONS *
1756              **************************/
1757
1758             /* REACTION-FIELD ELECTROSTATICS */
1759             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1760
1761             fscal            = felec;
1762
1763             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1764
1765             /* Calculate temporary vectorial force */
1766             tx               = _mm_mul_pd(fscal,dx32);
1767             ty               = _mm_mul_pd(fscal,dy32);
1768             tz               = _mm_mul_pd(fscal,dz32);
1769
1770             /* Update vectorial force */
1771             fix3             = _mm_add_pd(fix3,tx);
1772             fiy3             = _mm_add_pd(fiy3,ty);
1773             fiz3             = _mm_add_pd(fiz3,tz);
1774
1775             fjx2             = _mm_add_pd(fjx2,tx);
1776             fjy2             = _mm_add_pd(fjy2,ty);
1777             fjz2             = _mm_add_pd(fjz2,tz);
1778
1779             /**************************
1780              * CALCULATE INTERACTIONS *
1781              **************************/
1782
1783             /* REACTION-FIELD ELECTROSTATICS */
1784             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1785
1786             fscal            = felec;
1787
1788             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1789
1790             /* Calculate temporary vectorial force */
1791             tx               = _mm_mul_pd(fscal,dx33);
1792             ty               = _mm_mul_pd(fscal,dy33);
1793             tz               = _mm_mul_pd(fscal,dz33);
1794
1795             /* Update vectorial force */
1796             fix3             = _mm_add_pd(fix3,tx);
1797             fiy3             = _mm_add_pd(fiy3,ty);
1798             fiz3             = _mm_add_pd(fiz3,tz);
1799
1800             fjx3             = _mm_add_pd(fjx3,tx);
1801             fjy3             = _mm_add_pd(fjy3,ty);
1802             fjz3             = _mm_add_pd(fjz3,tz);
1803
1804             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1805
1806             /* Inner loop uses 273 flops */
1807         }
1808
1809         /* End of innermost loop */
1810
1811         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1812                                               f+i_coord_offset,fshift+i_shift_offset);
1813
1814         /* Increment number of inner iterations */
1815         inneriter                  += j_index_end - j_index_start;
1816
1817         /* Outer loop uses 24 flops */
1818     }
1819
1820     /* Increment number of outer iterations */
1821     outeriter        += nri;
1822
1823     /* Update outer/inner flops */
1824
1825     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1826 }