src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_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_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_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_avx_128_fma_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_msub_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             /* Update vectorial force */
 298             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 299             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 300             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 301
 302             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 303             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 304             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             /* COULOMB ELECTROSTATICS */
 311             velec            = _mm_mul_pd(qq11,rinv11);
 312             felec            = _mm_mul_pd(velec,rinvsq11);
 313
 314             /* Update potential sum for this i atom from the interaction with this j atom. */
 315             velecsum         = _mm_add_pd(velecsum,velec);
 316
 317             fscal            = felec;
 318
 319             /* Update vectorial force */
 320             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 321             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 322             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 323
 324             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 325             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 326             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             /* COULOMB ELECTROSTATICS */
 333             velec            = _mm_mul_pd(qq12,rinv12);
 334             felec            = _mm_mul_pd(velec,rinvsq12);
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             velecsum         = _mm_add_pd(velecsum,velec);
 338
 339             fscal            = felec;
 340
 341             /* Update vectorial force */
 342             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 343             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 344             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 345
 346             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 347             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 348             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 349
 350             /**************************
 351              * CALCULATE INTERACTIONS *
 352              **************************/
 353
 354             /* COULOMB ELECTROSTATICS */
 355             velec            = _mm_mul_pd(qq13,rinv13);
 356             felec            = _mm_mul_pd(velec,rinvsq13);
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velecsum         = _mm_add_pd(velecsum,velec);
 360
 361             fscal            = felec;
 362
 363             /* Update vectorial force */
 364             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 365             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 366             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 367
 368             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 369             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 370             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             /* COULOMB ELECTROSTATICS */
 377             velec            = _mm_mul_pd(qq21,rinv21);
 378             felec            = _mm_mul_pd(velec,rinvsq21);
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             /* Update vectorial force */
 386             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 387             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 388             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 389
 390             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 391             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 392             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             /* COULOMB ELECTROSTATICS */
 399             velec            = _mm_mul_pd(qq22,rinv22);
 400             felec            = _mm_mul_pd(velec,rinvsq22);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velecsum         = _mm_add_pd(velecsum,velec);
 404
 405             fscal            = felec;
 406
 407             /* Update vectorial force */
 408             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 409             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 410             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 411
 412             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 413             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 414             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* COULOMB ELECTROSTATICS */
 421             velec            = _mm_mul_pd(qq23,rinv23);
 422             felec            = _mm_mul_pd(velec,rinvsq23);
 423
 424             /* Update potential sum for this i atom from the interaction with this j atom. */
 425             velecsum         = _mm_add_pd(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             /* Update vectorial force */
 430             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 431             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 432             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 433
 434             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 435             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 436             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             /* COULOMB ELECTROSTATICS */
 443             velec            = _mm_mul_pd(qq31,rinv31);
 444             felec            = _mm_mul_pd(velec,rinvsq31);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velecsum         = _mm_add_pd(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             /* Update vectorial force */
 452             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 453             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 454             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 455
 456             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 457             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 458             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             /* COULOMB ELECTROSTATICS */
 465             velec            = _mm_mul_pd(qq32,rinv32);
 466             felec            = _mm_mul_pd(velec,rinvsq32);
 467
 468             /* Update potential sum for this i atom from the interaction with this j atom. */
 469             velecsum         = _mm_add_pd(velecsum,velec);
 470
 471             fscal            = felec;
 472
 473             /* Update vectorial force */
 474             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 475             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 476             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 477
 478             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 479             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 480             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             /* COULOMB ELECTROSTATICS */
 487             velec            = _mm_mul_pd(qq33,rinv33);
 488             felec            = _mm_mul_pd(velec,rinvsq33);
 489
 490             /* Update potential sum for this i atom from the interaction with this j atom. */
 491             velecsum         = _mm_add_pd(velecsum,velec);
 492
 493             fscal            = felec;
 494
 495             /* Update vectorial force */
 496             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 497             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 498             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 499
 500             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 501             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 502             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 503
 504             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);
 505
 506             /* Inner loop uses 317 flops */
 507         }
 508
 509         if(jidx<j_index_end)
 510         {
 511
 512             jnrA             = jjnr[jidx];
 513             j_coord_offsetA  = DIM*jnrA;
 514
 515             /* load j atom coordinates */
 516             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 517                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 518                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 519
 520             /* Calculate displacement vector */
 521             dx00             = _mm_sub_pd(ix0,jx0);
 522             dy00             = _mm_sub_pd(iy0,jy0);
 523             dz00             = _mm_sub_pd(iz0,jz0);
 524             dx11             = _mm_sub_pd(ix1,jx1);
 525             dy11             = _mm_sub_pd(iy1,jy1);
 526             dz11             = _mm_sub_pd(iz1,jz1);
 527             dx12             = _mm_sub_pd(ix1,jx2);
 528             dy12             = _mm_sub_pd(iy1,jy2);
 529             dz12             = _mm_sub_pd(iz1,jz2);
 530             dx13             = _mm_sub_pd(ix1,jx3);
 531             dy13             = _mm_sub_pd(iy1,jy3);
 532             dz13             = _mm_sub_pd(iz1,jz3);
 533             dx21             = _mm_sub_pd(ix2,jx1);
 534             dy21             = _mm_sub_pd(iy2,jy1);
 535             dz21             = _mm_sub_pd(iz2,jz1);
 536             dx22             = _mm_sub_pd(ix2,jx2);
 537             dy22             = _mm_sub_pd(iy2,jy2);
 538             dz22             = _mm_sub_pd(iz2,jz2);
 539             dx23             = _mm_sub_pd(ix2,jx3);
 540             dy23             = _mm_sub_pd(iy2,jy3);
 541             dz23             = _mm_sub_pd(iz2,jz3);
 542             dx31             = _mm_sub_pd(ix3,jx1);
 543             dy31             = _mm_sub_pd(iy3,jy1);
 544             dz31             = _mm_sub_pd(iz3,jz1);
 545             dx32             = _mm_sub_pd(ix3,jx2);
 546             dy32             = _mm_sub_pd(iy3,jy2);
 547             dz32             = _mm_sub_pd(iz3,jz2);
 548             dx33             = _mm_sub_pd(ix3,jx3);
 549             dy33             = _mm_sub_pd(iy3,jy3);
 550             dz33             = _mm_sub_pd(iz3,jz3);
 551
 552             /* Calculate squared distance and things based on it */
 553             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 554             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 555             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 556             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 557             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 558             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 559             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 560             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 561             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 562             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 563
 564             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 565             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 566             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 567             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 568             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 569             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 570             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 571             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 572             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 573
 574             rinvsq00         = gmx_mm_inv_pd(rsq00);
 575             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 576             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 577             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 578             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 579             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 580             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 581             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 582             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 583             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 584
 585             fjx0             = _mm_setzero_pd();
 586             fjy0             = _mm_setzero_pd();
 587             fjz0             = _mm_setzero_pd();
 588             fjx1             = _mm_setzero_pd();
 589             fjy1             = _mm_setzero_pd();
 590             fjz1             = _mm_setzero_pd();
 591             fjx2             = _mm_setzero_pd();
 592             fjy2             = _mm_setzero_pd();
 593             fjz2             = _mm_setzero_pd();
 594             fjx3             = _mm_setzero_pd();
 595             fjy3             = _mm_setzero_pd();
 596             fjz3             = _mm_setzero_pd();
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             /* LENNARD-JONES DISPERSION/REPULSION */
 603
 604             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 605             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 606             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 607             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 608             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 609
 610             /* Update potential sum for this i atom from the interaction with this j atom. */
 611             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 612             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 613
 614             fscal            = fvdw;
 615
 616             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 617
 618             /* Update vectorial force */
 619             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 620             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 621             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 622
 623             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 624             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 625             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 626
 627             /**************************
 628              * CALCULATE INTERACTIONS *
 629              **************************/
 630
 631             /* COULOMB ELECTROSTATICS */
 632             velec            = _mm_mul_pd(qq11,rinv11);
 633             felec            = _mm_mul_pd(velec,rinvsq11);
 634
 635             /* Update potential sum for this i atom from the interaction with this j atom. */
 636             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 637             velecsum         = _mm_add_pd(velecsum,velec);
 638
 639             fscal            = felec;
 640
 641             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 642
 643             /* Update vectorial force */
 644             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 645             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 646             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 647
 648             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 649             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 650             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             /* COULOMB ELECTROSTATICS */
 657             velec            = _mm_mul_pd(qq12,rinv12);
 658             felec            = _mm_mul_pd(velec,rinvsq12);
 659
 660             /* Update potential sum for this i atom from the interaction with this j atom. */
 661             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 662             velecsum         = _mm_add_pd(velecsum,velec);
 663
 664             fscal            = felec;
 665
 666             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 667
 668             /* Update vectorial force */
 669             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 670             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 671             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 672
 673             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 674             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 675             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             /* COULOMB ELECTROSTATICS */
 682             velec            = _mm_mul_pd(qq13,rinv13);
 683             felec            = _mm_mul_pd(velec,rinvsq13);
 684
 685             /* Update potential sum for this i atom from the interaction with this j atom. */
 686             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 687             velecsum         = _mm_add_pd(velecsum,velec);
 688
 689             fscal            = felec;
 690
 691             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 692
 693             /* Update vectorial force */
 694             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 695             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 696             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 697
 698             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 699             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 700             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 701
 702             /**************************
 703              * CALCULATE INTERACTIONS *
 704              **************************/
 705
 706             /* COULOMB ELECTROSTATICS */
 707             velec            = _mm_mul_pd(qq21,rinv21);
 708             felec            = _mm_mul_pd(velec,rinvsq21);
 709
 710             /* Update potential sum for this i atom from the interaction with this j atom. */
 711             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 712             velecsum         = _mm_add_pd(velecsum,velec);
 713
 714             fscal            = felec;
 715
 716             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 717
 718             /* Update vectorial force */
 719             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 720             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 721             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 722
 723             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 724             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 725             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             /* COULOMB ELECTROSTATICS */
 732             velec            = _mm_mul_pd(qq22,rinv22);
 733             felec            = _mm_mul_pd(velec,rinvsq22);
 734
 735             /* Update potential sum for this i atom from the interaction with this j atom. */
 736             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 737             velecsum         = _mm_add_pd(velecsum,velec);
 738
 739             fscal            = felec;
 740
 741             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 742
 743             /* Update vectorial force */
 744             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 745             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 746             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 747
 748             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 749             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 750             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* COULOMB ELECTROSTATICS */
 757             velec            = _mm_mul_pd(qq23,rinv23);
 758             felec            = _mm_mul_pd(velec,rinvsq23);
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 762             velecsum         = _mm_add_pd(velecsum,velec);
 763
 764             fscal            = felec;
 765
 766             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 767
 768             /* Update vectorial force */
 769             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 770             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 771             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 772
 773             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 774             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 775             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 776
 777             /**************************
 778              * CALCULATE INTERACTIONS *
 779              **************************/
 780
 781             /* COULOMB ELECTROSTATICS */
 782             velec            = _mm_mul_pd(qq31,rinv31);
 783             felec            = _mm_mul_pd(velec,rinvsq31);
 784
 785             /* Update potential sum for this i atom from the interaction with this j atom. */
 786             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 787             velecsum         = _mm_add_pd(velecsum,velec);
 788
 789             fscal            = felec;
 790
 791             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 792
 793             /* Update vectorial force */
 794             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 795             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 796             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 797
 798             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 799             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 800             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             /* COULOMB ELECTROSTATICS */
 807             velec            = _mm_mul_pd(qq32,rinv32);
 808             felec            = _mm_mul_pd(velec,rinvsq32);
 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             /* Update vectorial force */
 819             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 820             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 821             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 822
 823             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 824             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 825             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             /* COULOMB ELECTROSTATICS */
 832             velec            = _mm_mul_pd(qq33,rinv33);
 833             felec            = _mm_mul_pd(velec,rinvsq33);
 834
 835             /* Update potential sum for this i atom from the interaction with this j atom. */
 836             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 837             velecsum         = _mm_add_pd(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 842
 843             /* Update vectorial force */
 844             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 845             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 846             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 847
 848             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 849             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 850             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 851
 852             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 853
 854             /* Inner loop uses 317 flops */
 855         }
 856
 857         /* End of innermost loop */
 858
 859         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 860                                               f+i_coord_offset,fshift+i_shift_offset);
 861
 862         ggid                        = gid[iidx];
 863         /* Update potential energies */
 864         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 865         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 866
 867         /* Increment number of inner iterations */
 868         inneriter                  += j_index_end - j_index_start;
 869
 870         /* Outer loop uses 26 flops */
 871     }
 872
 873     /* Increment number of outer iterations */
 874     outeriter        += nri;
 875
 876     /* Update outer/inner flops */
 877
 878     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*317);
 879 }
 880 /*
 881  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_double
 882  * Electrostatics interaction: Coulomb
 883  * VdW interaction:            LennardJones
 884  * Geometry:                   Water4-Water4
 885  * Calculate force/pot:        Force
 886  */
 887 void
 888 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_double
 889                     (t_nblist * gmx_restrict                nlist,
 890                      rvec * gmx_restrict                    xx,
 891                      rvec * gmx_restrict                    ff,
 892                      t_forcerec * gmx_restrict              fr,
 893                      t_mdatoms * gmx_restrict               mdatoms,
 894                      nb_kernel_data_t * gmx_restrict        kernel_data,
 895                      t_nrnb * gmx_restrict                  nrnb)
 896 {
 897     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 898      * just 0 for non-waters.
 899      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 900      * jnr indices corresponding to data put in the four positions in the SIMD register.
 901      */
 902     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 903     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 904     int              jnrA,jnrB;
 905     int              j_coord_offsetA,j_coord_offsetB;
 906     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 907     real             rcutoff_scalar;
 908     real             *shiftvec,*fshift,*x,*f;
 909     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 910     int              vdwioffset0;
 911     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 912     int              vdwioffset1;
 913     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 914     int              vdwioffset2;
 915     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 916     int              vdwioffset3;
 917     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 918     int              vdwjidx0A,vdwjidx0B;
 919     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 920     int              vdwjidx1A,vdwjidx1B;
 921     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 922     int              vdwjidx2A,vdwjidx2B;
 923     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 924     int              vdwjidx3A,vdwjidx3B;
 925     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 926     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 927     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 928     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 929     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 930     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 931     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 932     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 933     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 934     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 935     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 936     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 937     real             *charge;
 938     int              nvdwtype;
 939     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 940     int              *vdwtype;
 941     real             *vdwparam;
 942     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 943     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 944     __m128d          dummy_mask,cutoff_mask;
 945     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 946     __m128d          one     = _mm_set1_pd(1.0);
 947     __m128d          two     = _mm_set1_pd(2.0);
 948     x                = xx[0];
 949     f                = ff[0];
 950
 951     nri              = nlist->nri;
 952     iinr             = nlist->iinr;
 953     jindex           = nlist->jindex;
 954     jjnr             = nlist->jjnr;
 955     shiftidx         = nlist->shift;
 956     gid              = nlist->gid;
 957     shiftvec         = fr->shift_vec[0];
 958     fshift           = fr->fshift[0];
 959     facel            = _mm_set1_pd(fr->epsfac);
 960     charge           = mdatoms->chargeA;
 961     nvdwtype         = fr->ntype;
 962     vdwparam         = fr->nbfp;
 963     vdwtype          = mdatoms->typeA;
 964
 965     /* Setup water-specific parameters */
 966     inr              = nlist->iinr[0];
 967     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 968     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 969     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 970     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 971
 972     jq1              = _mm_set1_pd(charge[inr+1]);
 973     jq2              = _mm_set1_pd(charge[inr+2]);
 974     jq3              = _mm_set1_pd(charge[inr+3]);
 975     vdwjidx0A        = 2*vdwtype[inr+0];
 976     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 977     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 978     qq11             = _mm_mul_pd(iq1,jq1);
 979     qq12             = _mm_mul_pd(iq1,jq2);
 980     qq13             = _mm_mul_pd(iq1,jq3);
 981     qq21             = _mm_mul_pd(iq2,jq1);
 982     qq22             = _mm_mul_pd(iq2,jq2);
 983     qq23             = _mm_mul_pd(iq2,jq3);
 984     qq31             = _mm_mul_pd(iq3,jq1);
 985     qq32             = _mm_mul_pd(iq3,jq2);
 986     qq33             = _mm_mul_pd(iq3,jq3);
 987
 988     /* Avoid stupid compiler warnings */
 989     jnrA = jnrB = 0;
 990     j_coord_offsetA = 0;
 991     j_coord_offsetB = 0;
 992
 993     outeriter        = 0;
 994     inneriter        = 0;
 995
 996     /* Start outer loop over neighborlists */
 997     for(iidx=0; iidx<nri; iidx++)
 998     {
 999         /* Load shift vector for this list */
1000         i_shift_offset   = DIM*shiftidx[iidx];
1001
1002         /* Load limits for loop over neighbors */
1003         j_index_start    = jindex[iidx];
1004         j_index_end      = jindex[iidx+1];
1005
1006         /* Get outer coordinate index */
1007         inr              = iinr[iidx];
1008         i_coord_offset   = DIM*inr;
1009
1010         /* Load i particle coords and add shift vector */
1011         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1012                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1013
1014         fix0             = _mm_setzero_pd();
1015         fiy0             = _mm_setzero_pd();
1016         fiz0             = _mm_setzero_pd();
1017         fix1             = _mm_setzero_pd();
1018         fiy1             = _mm_setzero_pd();
1019         fiz1             = _mm_setzero_pd();
1020         fix2             = _mm_setzero_pd();
1021         fiy2             = _mm_setzero_pd();
1022         fiz2             = _mm_setzero_pd();
1023         fix3             = _mm_setzero_pd();
1024         fiy3             = _mm_setzero_pd();
1025         fiz3             = _mm_setzero_pd();
1026
1027         /* Start inner kernel loop */
1028         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1029         {
1030
1031             /* Get j neighbor index, and coordinate index */
1032             jnrA             = jjnr[jidx];
1033             jnrB             = jjnr[jidx+1];
1034             j_coord_offsetA  = DIM*jnrA;
1035             j_coord_offsetB  = DIM*jnrB;
1036
1037             /* load j atom coordinates */
1038             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1039                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1040                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1041
1042             /* Calculate displacement vector */
1043             dx00             = _mm_sub_pd(ix0,jx0);
1044             dy00             = _mm_sub_pd(iy0,jy0);
1045             dz00             = _mm_sub_pd(iz0,jz0);
1046             dx11             = _mm_sub_pd(ix1,jx1);
1047             dy11             = _mm_sub_pd(iy1,jy1);
1048             dz11             = _mm_sub_pd(iz1,jz1);
1049             dx12             = _mm_sub_pd(ix1,jx2);
1050             dy12             = _mm_sub_pd(iy1,jy2);
1051             dz12             = _mm_sub_pd(iz1,jz2);
1052             dx13             = _mm_sub_pd(ix1,jx3);
1053             dy13             = _mm_sub_pd(iy1,jy3);
1054             dz13             = _mm_sub_pd(iz1,jz3);
1055             dx21             = _mm_sub_pd(ix2,jx1);
1056             dy21             = _mm_sub_pd(iy2,jy1);
1057             dz21             = _mm_sub_pd(iz2,jz1);
1058             dx22             = _mm_sub_pd(ix2,jx2);
1059             dy22             = _mm_sub_pd(iy2,jy2);
1060             dz22             = _mm_sub_pd(iz2,jz2);
1061             dx23             = _mm_sub_pd(ix2,jx3);
1062             dy23             = _mm_sub_pd(iy2,jy3);
1063             dz23             = _mm_sub_pd(iz2,jz3);
1064             dx31             = _mm_sub_pd(ix3,jx1);
1065             dy31             = _mm_sub_pd(iy3,jy1);
1066             dz31             = _mm_sub_pd(iz3,jz1);
1067             dx32             = _mm_sub_pd(ix3,jx2);
1068             dy32             = _mm_sub_pd(iy3,jy2);
1069             dz32             = _mm_sub_pd(iz3,jz2);
1070             dx33             = _mm_sub_pd(ix3,jx3);
1071             dy33             = _mm_sub_pd(iy3,jy3);
1072             dz33             = _mm_sub_pd(iz3,jz3);
1073
1074             /* Calculate squared distance and things based on it */
1075             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1076             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1077             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1078             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1079             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1080             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1081             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1082             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1083             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1084             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1085
1086             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1087             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1088             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1089             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1090             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1091             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1092             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1093             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1094             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1095
1096             rinvsq00         = gmx_mm_inv_pd(rsq00);
1097             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1098             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1099             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1100             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1101             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1102             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1103             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1104             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1105             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1106
1107             fjx0             = _mm_setzero_pd();
1108             fjy0             = _mm_setzero_pd();
1109             fjz0             = _mm_setzero_pd();
1110             fjx1             = _mm_setzero_pd();
1111             fjy1             = _mm_setzero_pd();
1112             fjz1             = _mm_setzero_pd();
1113             fjx2             = _mm_setzero_pd();
1114             fjy2             = _mm_setzero_pd();
1115             fjz2             = _mm_setzero_pd();
1116             fjx3             = _mm_setzero_pd();
1117             fjy3             = _mm_setzero_pd();
1118             fjz3             = _mm_setzero_pd();
1119
1120             /**************************
1121              * CALCULATE INTERACTIONS *
1122              **************************/
1123
1124             /* LENNARD-JONES DISPERSION/REPULSION */
1125
1126             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1127             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1128
1129             fscal            = fvdw;
1130
1131             /* Update vectorial force */
1132             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1133             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1134             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1135
1136             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1137             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1138             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1139
1140             /**************************
1141              * CALCULATE INTERACTIONS *
1142              **************************/
1143
1144             /* COULOMB ELECTROSTATICS */
1145             velec            = _mm_mul_pd(qq11,rinv11);
1146             felec            = _mm_mul_pd(velec,rinvsq11);
1147
1148             fscal            = felec;
1149
1150             /* Update vectorial force */
1151             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1152             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1153             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1154
1155             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1156             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1157             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1158
1159             /**************************
1160              * CALCULATE INTERACTIONS *
1161              **************************/
1162
1163             /* COULOMB ELECTROSTATICS */
1164             velec            = _mm_mul_pd(qq12,rinv12);
1165             felec            = _mm_mul_pd(velec,rinvsq12);
1166
1167             fscal            = felec;
1168
1169             /* Update vectorial force */
1170             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1171             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1172             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1173
1174             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1175             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1176             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1177
1178             /**************************
1179              * CALCULATE INTERACTIONS *
1180              **************************/
1181
1182             /* COULOMB ELECTROSTATICS */
1183             velec            = _mm_mul_pd(qq13,rinv13);
1184             felec            = _mm_mul_pd(velec,rinvsq13);
1185
1186             fscal            = felec;
1187
1188             /* Update vectorial force */
1189             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1190             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1191             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1192
1193             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1194             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1195             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1196
1197             /**************************
1198              * CALCULATE INTERACTIONS *
1199              **************************/
1200
1201             /* COULOMB ELECTROSTATICS */
1202             velec            = _mm_mul_pd(qq21,rinv21);
1203             felec            = _mm_mul_pd(velec,rinvsq21);
1204
1205             fscal            = felec;
1206
1207             /* Update vectorial force */
1208             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1209             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1210             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1211
1212             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1213             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1214             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1215
1216             /**************************
1217              * CALCULATE INTERACTIONS *
1218              **************************/
1219
1220             /* COULOMB ELECTROSTATICS */
1221             velec            = _mm_mul_pd(qq22,rinv22);
1222             felec            = _mm_mul_pd(velec,rinvsq22);
1223
1224             fscal            = felec;
1225
1226             /* Update vectorial force */
1227             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1228             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1229             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1230
1231             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1232             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1233             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1234
1235             /**************************
1236              * CALCULATE INTERACTIONS *
1237              **************************/
1238
1239             /* COULOMB ELECTROSTATICS */
1240             velec            = _mm_mul_pd(qq23,rinv23);
1241             felec            = _mm_mul_pd(velec,rinvsq23);
1242
1243             fscal            = felec;
1244
1245             /* Update vectorial force */
1246             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1247             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1248             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1249
1250             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1251             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1252             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1253
1254             /**************************
1255              * CALCULATE INTERACTIONS *
1256              **************************/
1257
1258             /* COULOMB ELECTROSTATICS */
1259             velec            = _mm_mul_pd(qq31,rinv31);
1260             felec            = _mm_mul_pd(velec,rinvsq31);
1261
1262             fscal            = felec;
1263
1264             /* Update vectorial force */
1265             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1266             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1267             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1268
1269             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1270             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1271             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1272
1273             /**************************
1274              * CALCULATE INTERACTIONS *
1275              **************************/
1276
1277             /* COULOMB ELECTROSTATICS */
1278             velec            = _mm_mul_pd(qq32,rinv32);
1279             felec            = _mm_mul_pd(velec,rinvsq32);
1280
1281             fscal            = felec;
1282
1283             /* Update vectorial force */
1284             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1285             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1286             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1287
1288             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1289             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1290             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* COULOMB ELECTROSTATICS */
1297             velec            = _mm_mul_pd(qq33,rinv33);
1298             felec            = _mm_mul_pd(velec,rinvsq33);
1299
1300             fscal            = felec;
1301
1302             /* Update vectorial force */
1303             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1304             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1305             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1306
1307             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1308             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1309             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1310
1311             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);
1312
1313             /* Inner loop uses 303 flops */
1314         }
1315
1316         if(jidx<j_index_end)
1317         {
1318
1319             jnrA             = jjnr[jidx];
1320             j_coord_offsetA  = DIM*jnrA;
1321
1322             /* load j atom coordinates */
1323             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1324                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1325                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1326
1327             /* Calculate displacement vector */
1328             dx00             = _mm_sub_pd(ix0,jx0);
1329             dy00             = _mm_sub_pd(iy0,jy0);
1330             dz00             = _mm_sub_pd(iz0,jz0);
1331             dx11             = _mm_sub_pd(ix1,jx1);
1332             dy11             = _mm_sub_pd(iy1,jy1);
1333             dz11             = _mm_sub_pd(iz1,jz1);
1334             dx12             = _mm_sub_pd(ix1,jx2);
1335             dy12             = _mm_sub_pd(iy1,jy2);
1336             dz12             = _mm_sub_pd(iz1,jz2);
1337             dx13             = _mm_sub_pd(ix1,jx3);
1338             dy13             = _mm_sub_pd(iy1,jy3);
1339             dz13             = _mm_sub_pd(iz1,jz3);
1340             dx21             = _mm_sub_pd(ix2,jx1);
1341             dy21             = _mm_sub_pd(iy2,jy1);
1342             dz21             = _mm_sub_pd(iz2,jz1);
1343             dx22             = _mm_sub_pd(ix2,jx2);
1344             dy22             = _mm_sub_pd(iy2,jy2);
1345             dz22             = _mm_sub_pd(iz2,jz2);
1346             dx23             = _mm_sub_pd(ix2,jx3);
1347             dy23             = _mm_sub_pd(iy2,jy3);
1348             dz23             = _mm_sub_pd(iz2,jz3);
1349             dx31             = _mm_sub_pd(ix3,jx1);
1350             dy31             = _mm_sub_pd(iy3,jy1);
1351             dz31             = _mm_sub_pd(iz3,jz1);
1352             dx32             = _mm_sub_pd(ix3,jx2);
1353             dy32             = _mm_sub_pd(iy3,jy2);
1354             dz32             = _mm_sub_pd(iz3,jz2);
1355             dx33             = _mm_sub_pd(ix3,jx3);
1356             dy33             = _mm_sub_pd(iy3,jy3);
1357             dz33             = _mm_sub_pd(iz3,jz3);
1358
1359             /* Calculate squared distance and things based on it */
1360             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1361             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1362             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1363             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1364             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1365             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1366             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1367             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1368             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1369             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1370
1371             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1372             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1373             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1374             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1375             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1376             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1377             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1378             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1379             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1380
1381             rinvsq00         = gmx_mm_inv_pd(rsq00);
1382             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1383             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1384             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1385             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1386             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1387             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1388             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1389             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1390             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1391
1392             fjx0             = _mm_setzero_pd();
1393             fjy0             = _mm_setzero_pd();
1394             fjz0             = _mm_setzero_pd();
1395             fjx1             = _mm_setzero_pd();
1396             fjy1             = _mm_setzero_pd();
1397             fjz1             = _mm_setzero_pd();
1398             fjx2             = _mm_setzero_pd();
1399             fjy2             = _mm_setzero_pd();
1400             fjz2             = _mm_setzero_pd();
1401             fjx3             = _mm_setzero_pd();
1402             fjy3             = _mm_setzero_pd();
1403             fjz3             = _mm_setzero_pd();
1404
1405             /**************************
1406              * CALCULATE INTERACTIONS *
1407              **************************/
1408
1409             /* LENNARD-JONES DISPERSION/REPULSION */
1410
1411             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1412             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1413
1414             fscal            = fvdw;
1415
1416             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1417
1418             /* Update vectorial force */
1419             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1420             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1421             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1422
1423             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1424             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1425             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1426
1427             /**************************
1428              * CALCULATE INTERACTIONS *
1429              **************************/
1430
1431             /* COULOMB ELECTROSTATICS */
1432             velec            = _mm_mul_pd(qq11,rinv11);
1433             felec            = _mm_mul_pd(velec,rinvsq11);
1434
1435             fscal            = felec;
1436
1437             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1438
1439             /* Update vectorial force */
1440             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1441             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1442             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1443
1444             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1445             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1446             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             /* COULOMB ELECTROSTATICS */
1453             velec            = _mm_mul_pd(qq12,rinv12);
1454             felec            = _mm_mul_pd(velec,rinvsq12);
1455
1456             fscal            = felec;
1457
1458             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1459
1460             /* Update vectorial force */
1461             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1462             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1463             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1464
1465             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1466             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1467             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1468
1469             /**************************
1470              * CALCULATE INTERACTIONS *
1471              **************************/
1472
1473             /* COULOMB ELECTROSTATICS */
1474             velec            = _mm_mul_pd(qq13,rinv13);
1475             felec            = _mm_mul_pd(velec,rinvsq13);
1476
1477             fscal            = felec;
1478
1479             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1480
1481             /* Update vectorial force */
1482             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1483             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1484             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1485
1486             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1487             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1488             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1489
1490             /**************************
1491              * CALCULATE INTERACTIONS *
1492              **************************/
1493
1494             /* COULOMB ELECTROSTATICS */
1495             velec            = _mm_mul_pd(qq21,rinv21);
1496             felec            = _mm_mul_pd(velec,rinvsq21);
1497
1498             fscal            = felec;
1499
1500             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1501
1502             /* Update vectorial force */
1503             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1504             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1505             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1506
1507             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1508             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1509             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1510
1511             /**************************
1512              * CALCULATE INTERACTIONS *
1513              **************************/
1514
1515             /* COULOMB ELECTROSTATICS */
1516             velec            = _mm_mul_pd(qq22,rinv22);
1517             felec            = _mm_mul_pd(velec,rinvsq22);
1518
1519             fscal            = felec;
1520
1521             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1522
1523             /* Update vectorial force */
1524             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1525             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1526             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1527
1528             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1529             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1530             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             /* COULOMB ELECTROSTATICS */
1537             velec            = _mm_mul_pd(qq23,rinv23);
1538             felec            = _mm_mul_pd(velec,rinvsq23);
1539
1540             fscal            = felec;
1541
1542             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1543
1544             /* Update vectorial force */
1545             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1546             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1547             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1548
1549             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1550             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1551             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1552
1553             /**************************
1554              * CALCULATE INTERACTIONS *
1555              **************************/
1556
1557             /* COULOMB ELECTROSTATICS */
1558             velec            = _mm_mul_pd(qq31,rinv31);
1559             felec            = _mm_mul_pd(velec,rinvsq31);
1560
1561             fscal            = felec;
1562
1563             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1564
1565             /* Update vectorial force */
1566             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1567             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1568             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1569
1570             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1571             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1572             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1573
1574             /**************************
1575              * CALCULATE INTERACTIONS *
1576              **************************/
1577
1578             /* COULOMB ELECTROSTATICS */
1579             velec            = _mm_mul_pd(qq32,rinv32);
1580             felec            = _mm_mul_pd(velec,rinvsq32);
1581
1582             fscal            = felec;
1583
1584             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1585
1586             /* Update vectorial force */
1587             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1588             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1589             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1590
1591             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1592             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1593             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1594
1595             /**************************
1596              * CALCULATE INTERACTIONS *
1597              **************************/
1598
1599             /* COULOMB ELECTROSTATICS */
1600             velec            = _mm_mul_pd(qq33,rinv33);
1601             felec            = _mm_mul_pd(velec,rinvsq33);
1602
1603             fscal            = felec;
1604
1605             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1606
1607             /* Update vectorial force */
1608             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1609             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1610             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1611
1612             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1613             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1614             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1615
1616             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1617
1618             /* Inner loop uses 303 flops */
1619         }
1620
1621         /* End of innermost loop */
1622
1623         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1624                                               f+i_coord_offset,fshift+i_shift_offset);
1625
1626         /* Increment number of inner iterations */
1627         inneriter                  += j_index_end - j_index_start;
1628
1629         /* Outer loop uses 24 flops */
1630     }
1631
1632     /* Increment number of outer iterations */
1633     outeriter        += nri;
1634
1635     /* Update outer/inner flops */
1636
1637     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1638 }