src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwCSTab_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128i          vfitab;
 117     __m128i          ifour       = _mm_set1_epi32(4);
 118     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 119     real             *vftab;
 120     __m128d          dummy_mask,cutoff_mask;
 121     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 122     __m128d          one     = _mm_set1_pd(1.0);
 123     __m128d          two     = _mm_set1_pd(2.0);
 124     x                = xx[0];
 125     f                = ff[0];
 126
 127     nri              = nlist->nri;
 128     iinr             = nlist->iinr;
 129     jindex           = nlist->jindex;
 130     jjnr             = nlist->jjnr;
 131     shiftidx         = nlist->shift;
 132     gid              = nlist->gid;
 133     shiftvec         = fr->shift_vec[0];
 134     fshift           = fr->fshift[0];
 135     facel            = _mm_set1_pd(fr->epsfac);
 136     charge           = mdatoms->chargeA;
 137     krf              = _mm_set1_pd(fr->ic->k_rf);
 138     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 139     crf              = _mm_set1_pd(fr->ic->c_rf);
 140     nvdwtype         = fr->ntype;
 141     vdwparam         = fr->nbfp;
 142     vdwtype          = mdatoms->typeA;
 143
 144     vftab            = kernel_data->table_vdw->data;
 145     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 146
 147     /* Setup water-specific parameters */
 148     inr              = nlist->iinr[0];
 149     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 150     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 151     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 152     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 153
 154     jq1              = _mm_set1_pd(charge[inr+1]);
 155     jq2              = _mm_set1_pd(charge[inr+2]);
 156     jq3              = _mm_set1_pd(charge[inr+3]);
 157     vdwjidx0A        = 2*vdwtype[inr+0];
 158     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 159     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 160     qq11             = _mm_mul_pd(iq1,jq1);
 161     qq12             = _mm_mul_pd(iq1,jq2);
 162     qq13             = _mm_mul_pd(iq1,jq3);
 163     qq21             = _mm_mul_pd(iq2,jq1);
 164     qq22             = _mm_mul_pd(iq2,jq2);
 165     qq23             = _mm_mul_pd(iq2,jq3);
 166     qq31             = _mm_mul_pd(iq3,jq1);
 167     qq32             = _mm_mul_pd(iq3,jq2);
 168     qq33             = _mm_mul_pd(iq3,jq3);
 169
 170     /* Avoid stupid compiler warnings */
 171     jnrA = jnrB = 0;
 172     j_coord_offsetA = 0;
 173     j_coord_offsetB = 0;
 174
 175     outeriter        = 0;
 176     inneriter        = 0;
 177
 178     /* Start outer loop over neighborlists */
 179     for(iidx=0; iidx<nri; iidx++)
 180     {
 181         /* Load shift vector for this list */
 182         i_shift_offset   = DIM*shiftidx[iidx];
 183
 184         /* Load limits for loop over neighbors */
 185         j_index_start    = jindex[iidx];
 186         j_index_end      = jindex[iidx+1];
 187
 188         /* Get outer coordinate index */
 189         inr              = iinr[iidx];
 190         i_coord_offset   = DIM*inr;
 191
 192         /* Load i particle coords and add shift vector */
 193         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 194                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 195
 196         fix0             = _mm_setzero_pd();
 197         fiy0             = _mm_setzero_pd();
 198         fiz0             = _mm_setzero_pd();
 199         fix1             = _mm_setzero_pd();
 200         fiy1             = _mm_setzero_pd();
 201         fiz1             = _mm_setzero_pd();
 202         fix2             = _mm_setzero_pd();
 203         fiy2             = _mm_setzero_pd();
 204         fiz2             = _mm_setzero_pd();
 205         fix3             = _mm_setzero_pd();
 206         fiy3             = _mm_setzero_pd();
 207         fiz3             = _mm_setzero_pd();
 208
 209         /* Reset potential sums */
 210         velecsum         = _mm_setzero_pd();
 211         vvdwsum          = _mm_setzero_pd();
 212
 213         /* Start inner kernel loop */
 214         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 215         {
 216
 217             /* Get j neighbor index, and coordinate index */
 218             jnrA             = jjnr[jidx];
 219             jnrB             = jjnr[jidx+1];
 220             j_coord_offsetA  = DIM*jnrA;
 221             j_coord_offsetB  = DIM*jnrB;
 222
 223             /* load j atom coordinates */
 224             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 225                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 226                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 227
 228             /* Calculate displacement vector */
 229             dx00             = _mm_sub_pd(ix0,jx0);
 230             dy00             = _mm_sub_pd(iy0,jy0);
 231             dz00             = _mm_sub_pd(iz0,jz0);
 232             dx11             = _mm_sub_pd(ix1,jx1);
 233             dy11             = _mm_sub_pd(iy1,jy1);
 234             dz11             = _mm_sub_pd(iz1,jz1);
 235             dx12             = _mm_sub_pd(ix1,jx2);
 236             dy12             = _mm_sub_pd(iy1,jy2);
 237             dz12             = _mm_sub_pd(iz1,jz2);
 238             dx13             = _mm_sub_pd(ix1,jx3);
 239             dy13             = _mm_sub_pd(iy1,jy3);
 240             dz13             = _mm_sub_pd(iz1,jz3);
 241             dx21             = _mm_sub_pd(ix2,jx1);
 242             dy21             = _mm_sub_pd(iy2,jy1);
 243             dz21             = _mm_sub_pd(iz2,jz1);
 244             dx22             = _mm_sub_pd(ix2,jx2);
 245             dy22             = _mm_sub_pd(iy2,jy2);
 246             dz22             = _mm_sub_pd(iz2,jz2);
 247             dx23             = _mm_sub_pd(ix2,jx3);
 248             dy23             = _mm_sub_pd(iy2,jy3);
 249             dz23             = _mm_sub_pd(iz2,jz3);
 250             dx31             = _mm_sub_pd(ix3,jx1);
 251             dy31             = _mm_sub_pd(iy3,jy1);
 252             dz31             = _mm_sub_pd(iz3,jz1);
 253             dx32             = _mm_sub_pd(ix3,jx2);
 254             dy32             = _mm_sub_pd(iy3,jy2);
 255             dz32             = _mm_sub_pd(iz3,jz2);
 256             dx33             = _mm_sub_pd(ix3,jx3);
 257             dy33             = _mm_sub_pd(iy3,jy3);
 258             dz33             = _mm_sub_pd(iz3,jz3);
 259
 260             /* Calculate squared distance and things based on it */
 261             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 262             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 263             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 264             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 265             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 266             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 267             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 268             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 269             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 270             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 271
 272             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 273             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 274             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 275             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 276             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 277             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 278             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 279             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 280             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 281             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 282
 283             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 284             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 285             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 286             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 287             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 288             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 289             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 290             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 291             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 292
 293             fjx0             = _mm_setzero_pd();
 294             fjy0             = _mm_setzero_pd();
 295             fjz0             = _mm_setzero_pd();
 296             fjx1             = _mm_setzero_pd();
 297             fjy1             = _mm_setzero_pd();
 298             fjz1             = _mm_setzero_pd();
 299             fjx2             = _mm_setzero_pd();
 300             fjy2             = _mm_setzero_pd();
 301             fjz2             = _mm_setzero_pd();
 302             fjx3             = _mm_setzero_pd();
 303             fjy3             = _mm_setzero_pd();
 304             fjz3             = _mm_setzero_pd();
 305
 306             /**************************
 307              * CALCULATE INTERACTIONS *
 308              **************************/
 309
 310             r00              = _mm_mul_pd(rsq00,rinv00);
 311
 312             /* Calculate table index by multiplying r with table scale and truncate to integer */
 313             rt               = _mm_mul_pd(r00,vftabscale);
 314             vfitab           = _mm_cvttpd_epi32(rt);
 315 #ifdef __XOP__
 316             vfeps            = _mm_frcz_pd(rt);
 317 #else
 318             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 319 #endif
 320             twovfeps         = _mm_add_pd(vfeps,vfeps);
 321             vfitab           = _mm_slli_epi32(vfitab,3);
 322
 323             /* CUBIC SPLINE TABLE DISPERSION */
 324             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 325             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 326             GMX_MM_TRANSPOSE2_PD(Y,F);
 327             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 328             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 329             GMX_MM_TRANSPOSE2_PD(G,H);
 330             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 331             VV               = _mm_macc_pd(vfeps,Fp,Y);
 332             vvdw6            = _mm_mul_pd(c6_00,VV);
 333             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 334             fvdw6            = _mm_mul_pd(c6_00,FF);
 335
 336             /* CUBIC SPLINE TABLE REPULSION */
 337             vfitab           = _mm_add_epi32(vfitab,ifour);
 338             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 339             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 340             GMX_MM_TRANSPOSE2_PD(Y,F);
 341             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 342             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 343             GMX_MM_TRANSPOSE2_PD(G,H);
 344             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 345             VV               = _mm_macc_pd(vfeps,Fp,Y);
 346             vvdw12           = _mm_mul_pd(c12_00,VV);
 347             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 348             fvdw12           = _mm_mul_pd(c12_00,FF);
 349             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 350             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 354
 355             fscal            = fvdw;
 356
 357             /* Update vectorial force */
 358             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 359             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 360             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 361
 362             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 363             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 364             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             /* REACTION-FIELD ELECTROSTATICS */
 371             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 372             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_pd(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             /* Update vectorial force */
 380             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 381             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 382             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 383
 384             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 385             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 386             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* REACTION-FIELD ELECTROSTATICS */
 393             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 394             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velecsum         = _mm_add_pd(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401             /* Update vectorial force */
 402             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 403             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 404             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 405
 406             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 407             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 408             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 416             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velecsum         = _mm_add_pd(velecsum,velec);
 420
 421             fscal            = felec;
 422
 423             /* Update vectorial force */
 424             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 425             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 426             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 427
 428             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 429             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 430             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             /* REACTION-FIELD ELECTROSTATICS */
 437             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 438             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 439
 440             /* Update potential sum for this i atom from the interaction with this j atom. */
 441             velecsum         = _mm_add_pd(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             /* Update vectorial force */
 446             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 447             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 448             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 449
 450             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 451             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 452             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 453
 454             /**************************
 455              * CALCULATE INTERACTIONS *
 456              **************************/
 457
 458             /* REACTION-FIELD ELECTROSTATICS */
 459             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 460             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 461
 462             /* Update potential sum for this i atom from the interaction with this j atom. */
 463             velecsum         = _mm_add_pd(velecsum,velec);
 464
 465             fscal            = felec;
 466
 467             /* Update vectorial force */
 468             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 469             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 470             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 471
 472             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 473             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 474             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 475
 476             /**************************
 477              * CALCULATE INTERACTIONS *
 478              **************************/
 479
 480             /* REACTION-FIELD ELECTROSTATICS */
 481             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 482             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,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             /* Update vectorial force */
 490             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 491             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 492             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 493
 494             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 495             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 496             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             /* REACTION-FIELD ELECTROSTATICS */
 503             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 504             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velecsum         = _mm_add_pd(velecsum,velec);
 508
 509             fscal            = felec;
 510
 511             /* Update vectorial force */
 512             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 513             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 514             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 515
 516             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 517             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 518             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 519
 520             /**************************
 521              * CALCULATE INTERACTIONS *
 522              **************************/
 523
 524             /* REACTION-FIELD ELECTROSTATICS */
 525             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 526             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 527
 528             /* Update potential sum for this i atom from the interaction with this j atom. */
 529             velecsum         = _mm_add_pd(velecsum,velec);
 530
 531             fscal            = felec;
 532
 533             /* Update vectorial force */
 534             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 535             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 536             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 537
 538             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 539             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 540             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 541
 542             /**************************
 543              * CALCULATE INTERACTIONS *
 544              **************************/
 545
 546             /* REACTION-FIELD ELECTROSTATICS */
 547             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 548             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 549
 550             /* Update potential sum for this i atom from the interaction with this j atom. */
 551             velecsum         = _mm_add_pd(velecsum,velec);
 552
 553             fscal            = felec;
 554
 555             /* Update vectorial force */
 556             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 557             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 558             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 559
 560             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 561             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 562             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 563
 564             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);
 565
 566             /* Inner loop uses 377 flops */
 567         }
 568
 569         if(jidx<j_index_end)
 570         {
 571
 572             jnrA             = jjnr[jidx];
 573             j_coord_offsetA  = DIM*jnrA;
 574
 575             /* load j atom coordinates */
 576             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 577                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 578                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 579
 580             /* Calculate displacement vector */
 581             dx00             = _mm_sub_pd(ix0,jx0);
 582             dy00             = _mm_sub_pd(iy0,jy0);
 583             dz00             = _mm_sub_pd(iz0,jz0);
 584             dx11             = _mm_sub_pd(ix1,jx1);
 585             dy11             = _mm_sub_pd(iy1,jy1);
 586             dz11             = _mm_sub_pd(iz1,jz1);
 587             dx12             = _mm_sub_pd(ix1,jx2);
 588             dy12             = _mm_sub_pd(iy1,jy2);
 589             dz12             = _mm_sub_pd(iz1,jz2);
 590             dx13             = _mm_sub_pd(ix1,jx3);
 591             dy13             = _mm_sub_pd(iy1,jy3);
 592             dz13             = _mm_sub_pd(iz1,jz3);
 593             dx21             = _mm_sub_pd(ix2,jx1);
 594             dy21             = _mm_sub_pd(iy2,jy1);
 595             dz21             = _mm_sub_pd(iz2,jz1);
 596             dx22             = _mm_sub_pd(ix2,jx2);
 597             dy22             = _mm_sub_pd(iy2,jy2);
 598             dz22             = _mm_sub_pd(iz2,jz2);
 599             dx23             = _mm_sub_pd(ix2,jx3);
 600             dy23             = _mm_sub_pd(iy2,jy3);
 601             dz23             = _mm_sub_pd(iz2,jz3);
 602             dx31             = _mm_sub_pd(ix3,jx1);
 603             dy31             = _mm_sub_pd(iy3,jy1);
 604             dz31             = _mm_sub_pd(iz3,jz1);
 605             dx32             = _mm_sub_pd(ix3,jx2);
 606             dy32             = _mm_sub_pd(iy3,jy2);
 607             dz32             = _mm_sub_pd(iz3,jz2);
 608             dx33             = _mm_sub_pd(ix3,jx3);
 609             dy33             = _mm_sub_pd(iy3,jy3);
 610             dz33             = _mm_sub_pd(iz3,jz3);
 611
 612             /* Calculate squared distance and things based on it */
 613             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 614             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 615             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 616             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 617             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 618             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 619             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 620             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 621             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 622             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 623
 624             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 625             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 626             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 627             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 628             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 629             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 630             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 631             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 632             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 633             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 634
 635             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 636             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 637             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 638             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 639             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 640             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 641             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 642             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 643             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 644
 645             fjx0             = _mm_setzero_pd();
 646             fjy0             = _mm_setzero_pd();
 647             fjz0             = _mm_setzero_pd();
 648             fjx1             = _mm_setzero_pd();
 649             fjy1             = _mm_setzero_pd();
 650             fjz1             = _mm_setzero_pd();
 651             fjx2             = _mm_setzero_pd();
 652             fjy2             = _mm_setzero_pd();
 653             fjz2             = _mm_setzero_pd();
 654             fjx3             = _mm_setzero_pd();
 655             fjy3             = _mm_setzero_pd();
 656             fjz3             = _mm_setzero_pd();
 657
 658             /**************************
 659              * CALCULATE INTERACTIONS *
 660              **************************/
 661
 662             r00              = _mm_mul_pd(rsq00,rinv00);
 663
 664             /* Calculate table index by multiplying r with table scale and truncate to integer */
 665             rt               = _mm_mul_pd(r00,vftabscale);
 666             vfitab           = _mm_cvttpd_epi32(rt);
 667 #ifdef __XOP__
 668             vfeps            = _mm_frcz_pd(rt);
 669 #else
 670             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 671 #endif
 672             twovfeps         = _mm_add_pd(vfeps,vfeps);
 673             vfitab           = _mm_slli_epi32(vfitab,3);
 674
 675             /* CUBIC SPLINE TABLE DISPERSION */
 676             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 677             F                = _mm_setzero_pd();
 678             GMX_MM_TRANSPOSE2_PD(Y,F);
 679             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 680             H                = _mm_setzero_pd();
 681             GMX_MM_TRANSPOSE2_PD(G,H);
 682             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 683             VV               = _mm_macc_pd(vfeps,Fp,Y);
 684             vvdw6            = _mm_mul_pd(c6_00,VV);
 685             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 686             fvdw6            = _mm_mul_pd(c6_00,FF);
 687
 688             /* CUBIC SPLINE TABLE REPULSION */
 689             vfitab           = _mm_add_epi32(vfitab,ifour);
 690             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 691             F                = _mm_setzero_pd();
 692             GMX_MM_TRANSPOSE2_PD(Y,F);
 693             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 694             H                = _mm_setzero_pd();
 695             GMX_MM_TRANSPOSE2_PD(G,H);
 696             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 697             VV               = _mm_macc_pd(vfeps,Fp,Y);
 698             vvdw12           = _mm_mul_pd(c12_00,VV);
 699             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 700             fvdw12           = _mm_mul_pd(c12_00,FF);
 701             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 702             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 703
 704             /* Update potential sum for this i atom from the interaction with this j atom. */
 705             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 706             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 707
 708             fscal            = fvdw;
 709
 710             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 711
 712             /* Update vectorial force */
 713             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 714             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 715             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 716
 717             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 718             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 719             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* REACTION-FIELD ELECTROSTATICS */
 726             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 727             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 731             velecsum         = _mm_add_pd(velecsum,velec);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 736
 737             /* Update vectorial force */
 738             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 739             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 740             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 741
 742             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 743             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 744             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 745
 746             /**************************
 747              * CALCULATE INTERACTIONS *
 748              **************************/
 749
 750             /* REACTION-FIELD ELECTROSTATICS */
 751             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 752             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 753
 754             /* Update potential sum for this i atom from the interaction with this j atom. */
 755             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 756             velecsum         = _mm_add_pd(velecsum,velec);
 757
 758             fscal            = felec;
 759
 760             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 761
 762             /* Update vectorial force */
 763             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 764             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 765             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 766
 767             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 768             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 769             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 770
 771             /**************************
 772              * CALCULATE INTERACTIONS *
 773              **************************/
 774
 775             /* REACTION-FIELD ELECTROSTATICS */
 776             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 777             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 781             velecsum         = _mm_add_pd(velecsum,velec);
 782
 783             fscal            = felec;
 784
 785             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 786
 787             /* Update vectorial force */
 788             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 789             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 790             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 791
 792             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 793             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 794             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             /* REACTION-FIELD ELECTROSTATICS */
 801             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 802             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 803
 804             /* Update potential sum for this i atom from the interaction with this j atom. */
 805             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 806             velecsum         = _mm_add_pd(velecsum,velec);
 807
 808             fscal            = felec;
 809
 810             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 811
 812             /* Update vectorial force */
 813             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 814             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 815             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 816
 817             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 818             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 819             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 820
 821             /**************************
 822              * CALCULATE INTERACTIONS *
 823              **************************/
 824
 825             /* REACTION-FIELD ELECTROSTATICS */
 826             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 827             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 828
 829             /* Update potential sum for this i atom from the interaction with this j atom. */
 830             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 831             velecsum         = _mm_add_pd(velecsum,velec);
 832
 833             fscal            = felec;
 834
 835             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 836
 837             /* Update vectorial force */
 838             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 839             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 840             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 841
 842             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 843             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 844             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 845
 846             /**************************
 847              * CALCULATE INTERACTIONS *
 848              **************************/
 849
 850             /* REACTION-FIELD ELECTROSTATICS */
 851             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 852             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 853
 854             /* Update potential sum for this i atom from the interaction with this j atom. */
 855             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 856             velecsum         = _mm_add_pd(velecsum,velec);
 857
 858             fscal            = felec;
 859
 860             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 861
 862             /* Update vectorial force */
 863             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 864             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 865             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 866
 867             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 868             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 869             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 870
 871             /**************************
 872              * CALCULATE INTERACTIONS *
 873              **************************/
 874
 875             /* REACTION-FIELD ELECTROSTATICS */
 876             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 877             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 878
 879             /* Update potential sum for this i atom from the interaction with this j atom. */
 880             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 881             velecsum         = _mm_add_pd(velecsum,velec);
 882
 883             fscal            = felec;
 884
 885             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 886
 887             /* Update vectorial force */
 888             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 889             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 890             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 891
 892             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 893             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 894             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             /* REACTION-FIELD ELECTROSTATICS */
 901             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 902             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 903
 904             /* Update potential sum for this i atom from the interaction with this j atom. */
 905             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 906             velecsum         = _mm_add_pd(velecsum,velec);
 907
 908             fscal            = felec;
 909
 910             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 911
 912             /* Update vectorial force */
 913             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 914             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 915             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 916
 917             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 918             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 919             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 920
 921             /**************************
 922              * CALCULATE INTERACTIONS *
 923              **************************/
 924
 925             /* REACTION-FIELD ELECTROSTATICS */
 926             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 927             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 928
 929             /* Update potential sum for this i atom from the interaction with this j atom. */
 930             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 931             velecsum         = _mm_add_pd(velecsum,velec);
 932
 933             fscal            = felec;
 934
 935             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 936
 937             /* Update vectorial force */
 938             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 939             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 940             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 941
 942             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 943             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 944             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 945
 946             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 947
 948             /* Inner loop uses 377 flops */
 949         }
 950
 951         /* End of innermost loop */
 952
 953         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 954                                               f+i_coord_offset,fshift+i_shift_offset);
 955
 956         ggid                        = gid[iidx];
 957         /* Update potential energies */
 958         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 959         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 960
 961         /* Increment number of inner iterations */
 962         inneriter                  += j_index_end - j_index_start;
 963
 964         /* Outer loop uses 26 flops */
 965     }
 966
 967     /* Increment number of outer iterations */
 968     outeriter        += nri;
 969
 970     /* Update outer/inner flops */
 971
 972     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*377);
 973 }
 974 /*
 975  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 976  * Electrostatics interaction: ReactionField
 977  * VdW interaction:            CubicSplineTable
 978  * Geometry:                   Water4-Water4
 979  * Calculate force/pot:        Force
 980  */
 981 void
 982 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 983                     (t_nblist                    * gmx_restrict       nlist,
 984                      rvec                        * gmx_restrict          xx,
 985                      rvec                        * gmx_restrict          ff,
 986                      t_forcerec                  * gmx_restrict          fr,
 987                      t_mdatoms                   * gmx_restrict     mdatoms,
 988                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 989                      t_nrnb                      * gmx_restrict        nrnb)
 990 {
 991     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 992      * just 0 for non-waters.
 993      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 994      * jnr indices corresponding to data put in the four positions in the SIMD register.
 995      */
 996     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 997     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 998     int              jnrA,jnrB;
 999     int              j_coord_offsetA,j_coord_offsetB;
1000     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1001     real             rcutoff_scalar;
1002     real             *shiftvec,*fshift,*x,*f;
1003     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1004     int              vdwioffset0;
1005     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1006     int              vdwioffset1;
1007     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1008     int              vdwioffset2;
1009     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1010     int              vdwioffset3;
1011     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1012     int              vdwjidx0A,vdwjidx0B;
1013     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1014     int              vdwjidx1A,vdwjidx1B;
1015     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1016     int              vdwjidx2A,vdwjidx2B;
1017     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1018     int              vdwjidx3A,vdwjidx3B;
1019     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1020     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1021     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1022     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1023     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1024     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1025     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1026     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1027     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1028     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1029     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1030     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1031     real             *charge;
1032     int              nvdwtype;
1033     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1034     int              *vdwtype;
1035     real             *vdwparam;
1036     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1037     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1038     __m128i          vfitab;
1039     __m128i          ifour       = _mm_set1_epi32(4);
1040     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1041     real             *vftab;
1042     __m128d          dummy_mask,cutoff_mask;
1043     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1044     __m128d          one     = _mm_set1_pd(1.0);
1045     __m128d          two     = _mm_set1_pd(2.0);
1046     x                = xx[0];
1047     f                = ff[0];
1048
1049     nri              = nlist->nri;
1050     iinr             = nlist->iinr;
1051     jindex           = nlist->jindex;
1052     jjnr             = nlist->jjnr;
1053     shiftidx         = nlist->shift;
1054     gid              = nlist->gid;
1055     shiftvec         = fr->shift_vec[0];
1056     fshift           = fr->fshift[0];
1057     facel            = _mm_set1_pd(fr->epsfac);
1058     charge           = mdatoms->chargeA;
1059     krf              = _mm_set1_pd(fr->ic->k_rf);
1060     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1061     crf              = _mm_set1_pd(fr->ic->c_rf);
1062     nvdwtype         = fr->ntype;
1063     vdwparam         = fr->nbfp;
1064     vdwtype          = mdatoms->typeA;
1065
1066     vftab            = kernel_data->table_vdw->data;
1067     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1068
1069     /* Setup water-specific parameters */
1070     inr              = nlist->iinr[0];
1071     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1072     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1073     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1074     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1075
1076     jq1              = _mm_set1_pd(charge[inr+1]);
1077     jq2              = _mm_set1_pd(charge[inr+2]);
1078     jq3              = _mm_set1_pd(charge[inr+3]);
1079     vdwjidx0A        = 2*vdwtype[inr+0];
1080     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1081     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1082     qq11             = _mm_mul_pd(iq1,jq1);
1083     qq12             = _mm_mul_pd(iq1,jq2);
1084     qq13             = _mm_mul_pd(iq1,jq3);
1085     qq21             = _mm_mul_pd(iq2,jq1);
1086     qq22             = _mm_mul_pd(iq2,jq2);
1087     qq23             = _mm_mul_pd(iq2,jq3);
1088     qq31             = _mm_mul_pd(iq3,jq1);
1089     qq32             = _mm_mul_pd(iq3,jq2);
1090     qq33             = _mm_mul_pd(iq3,jq3);
1091
1092     /* Avoid stupid compiler warnings */
1093     jnrA = jnrB = 0;
1094     j_coord_offsetA = 0;
1095     j_coord_offsetB = 0;
1096
1097     outeriter        = 0;
1098     inneriter        = 0;
1099
1100     /* Start outer loop over neighborlists */
1101     for(iidx=0; iidx<nri; iidx++)
1102     {
1103         /* Load shift vector for this list */
1104         i_shift_offset   = DIM*shiftidx[iidx];
1105
1106         /* Load limits for loop over neighbors */
1107         j_index_start    = jindex[iidx];
1108         j_index_end      = jindex[iidx+1];
1109
1110         /* Get outer coordinate index */
1111         inr              = iinr[iidx];
1112         i_coord_offset   = DIM*inr;
1113
1114         /* Load i particle coords and add shift vector */
1115         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1116                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1117
1118         fix0             = _mm_setzero_pd();
1119         fiy0             = _mm_setzero_pd();
1120         fiz0             = _mm_setzero_pd();
1121         fix1             = _mm_setzero_pd();
1122         fiy1             = _mm_setzero_pd();
1123         fiz1             = _mm_setzero_pd();
1124         fix2             = _mm_setzero_pd();
1125         fiy2             = _mm_setzero_pd();
1126         fiz2             = _mm_setzero_pd();
1127         fix3             = _mm_setzero_pd();
1128         fiy3             = _mm_setzero_pd();
1129         fiz3             = _mm_setzero_pd();
1130
1131         /* Start inner kernel loop */
1132         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1133         {
1134
1135             /* Get j neighbor index, and coordinate index */
1136             jnrA             = jjnr[jidx];
1137             jnrB             = jjnr[jidx+1];
1138             j_coord_offsetA  = DIM*jnrA;
1139             j_coord_offsetB  = DIM*jnrB;
1140
1141             /* load j atom coordinates */
1142             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1143                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1144                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1145
1146             /* Calculate displacement vector */
1147             dx00             = _mm_sub_pd(ix0,jx0);
1148             dy00             = _mm_sub_pd(iy0,jy0);
1149             dz00             = _mm_sub_pd(iz0,jz0);
1150             dx11             = _mm_sub_pd(ix1,jx1);
1151             dy11             = _mm_sub_pd(iy1,jy1);
1152             dz11             = _mm_sub_pd(iz1,jz1);
1153             dx12             = _mm_sub_pd(ix1,jx2);
1154             dy12             = _mm_sub_pd(iy1,jy2);
1155             dz12             = _mm_sub_pd(iz1,jz2);
1156             dx13             = _mm_sub_pd(ix1,jx3);
1157             dy13             = _mm_sub_pd(iy1,jy3);
1158             dz13             = _mm_sub_pd(iz1,jz3);
1159             dx21             = _mm_sub_pd(ix2,jx1);
1160             dy21             = _mm_sub_pd(iy2,jy1);
1161             dz21             = _mm_sub_pd(iz2,jz1);
1162             dx22             = _mm_sub_pd(ix2,jx2);
1163             dy22             = _mm_sub_pd(iy2,jy2);
1164             dz22             = _mm_sub_pd(iz2,jz2);
1165             dx23             = _mm_sub_pd(ix2,jx3);
1166             dy23             = _mm_sub_pd(iy2,jy3);
1167             dz23             = _mm_sub_pd(iz2,jz3);
1168             dx31             = _mm_sub_pd(ix3,jx1);
1169             dy31             = _mm_sub_pd(iy3,jy1);
1170             dz31             = _mm_sub_pd(iz3,jz1);
1171             dx32             = _mm_sub_pd(ix3,jx2);
1172             dy32             = _mm_sub_pd(iy3,jy2);
1173             dz32             = _mm_sub_pd(iz3,jz2);
1174             dx33             = _mm_sub_pd(ix3,jx3);
1175             dy33             = _mm_sub_pd(iy3,jy3);
1176             dz33             = _mm_sub_pd(iz3,jz3);
1177
1178             /* Calculate squared distance and things based on it */
1179             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1180             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1181             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1182             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1183             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1184             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1185             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1186             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1187             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1188             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1189
1190             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1191             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1192             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1193             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1194             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1195             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1196             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1197             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1198             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1199             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1200
1201             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1202             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1203             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1204             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1205             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1206             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1207             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1208             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1209             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1210
1211             fjx0             = _mm_setzero_pd();
1212             fjy0             = _mm_setzero_pd();
1213             fjz0             = _mm_setzero_pd();
1214             fjx1             = _mm_setzero_pd();
1215             fjy1             = _mm_setzero_pd();
1216             fjz1             = _mm_setzero_pd();
1217             fjx2             = _mm_setzero_pd();
1218             fjy2             = _mm_setzero_pd();
1219             fjz2             = _mm_setzero_pd();
1220             fjx3             = _mm_setzero_pd();
1221             fjy3             = _mm_setzero_pd();
1222             fjz3             = _mm_setzero_pd();
1223
1224             /**************************
1225              * CALCULATE INTERACTIONS *
1226              **************************/
1227
1228             r00              = _mm_mul_pd(rsq00,rinv00);
1229
1230             /* Calculate table index by multiplying r with table scale and truncate to integer */
1231             rt               = _mm_mul_pd(r00,vftabscale);
1232             vfitab           = _mm_cvttpd_epi32(rt);
1233 #ifdef __XOP__
1234             vfeps            = _mm_frcz_pd(rt);
1235 #else
1236             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1237 #endif
1238             twovfeps         = _mm_add_pd(vfeps,vfeps);
1239             vfitab           = _mm_slli_epi32(vfitab,3);
1240
1241             /* CUBIC SPLINE TABLE DISPERSION */
1242             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1243             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1244             GMX_MM_TRANSPOSE2_PD(Y,F);
1245             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1246             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1247             GMX_MM_TRANSPOSE2_PD(G,H);
1248             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1249             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1250             fvdw6            = _mm_mul_pd(c6_00,FF);
1251
1252             /* CUBIC SPLINE TABLE REPULSION */
1253             vfitab           = _mm_add_epi32(vfitab,ifour);
1254             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1255             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1256             GMX_MM_TRANSPOSE2_PD(Y,F);
1257             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1258             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1259             GMX_MM_TRANSPOSE2_PD(G,H);
1260             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1261             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1262             fvdw12           = _mm_mul_pd(c12_00,FF);
1263             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1264
1265             fscal            = fvdw;
1266
1267             /* Update vectorial force */
1268             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1269             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1270             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1271
1272             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1273             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1274             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1275
1276             /**************************
1277              * CALCULATE INTERACTIONS *
1278              **************************/
1279
1280             /* REACTION-FIELD ELECTROSTATICS */
1281             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1282
1283             fscal            = felec;
1284
1285             /* Update vectorial force */
1286             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1287             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1288             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1289
1290             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1291             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1292             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1293
1294             /**************************
1295              * CALCULATE INTERACTIONS *
1296              **************************/
1297
1298             /* REACTION-FIELD ELECTROSTATICS */
1299             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1300
1301             fscal            = felec;
1302
1303             /* Update vectorial force */
1304             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1305             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1306             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1307
1308             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1309             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1310             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             /* REACTION-FIELD ELECTROSTATICS */
1317             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1318
1319             fscal            = felec;
1320
1321             /* Update vectorial force */
1322             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1323             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1324             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1325
1326             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1327             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1328             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             /* REACTION-FIELD ELECTROSTATICS */
1335             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1336
1337             fscal            = felec;
1338
1339             /* Update vectorial force */
1340             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1341             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1342             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1343
1344             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1345             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1346             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* REACTION-FIELD ELECTROSTATICS */
1353             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1354
1355             fscal            = felec;
1356
1357             /* Update vectorial force */
1358             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1359             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1360             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1361
1362             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1363             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1364             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* REACTION-FIELD ELECTROSTATICS */
1371             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1372
1373             fscal            = felec;
1374
1375             /* Update vectorial force */
1376             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1377             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1378             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1379
1380             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1381             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1382             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             /* REACTION-FIELD ELECTROSTATICS */
1389             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1390
1391             fscal            = felec;
1392
1393             /* Update vectorial force */
1394             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1395             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1396             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1397
1398             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1399             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1400             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1401
1402             /**************************
1403              * CALCULATE INTERACTIONS *
1404              **************************/
1405
1406             /* REACTION-FIELD ELECTROSTATICS */
1407             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1408
1409             fscal            = felec;
1410
1411             /* Update vectorial force */
1412             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1413             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1414             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1415
1416             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1417             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1418             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1419
1420             /**************************
1421              * CALCULATE INTERACTIONS *
1422              **************************/
1423
1424             /* REACTION-FIELD ELECTROSTATICS */
1425             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1426
1427             fscal            = felec;
1428
1429             /* Update vectorial force */
1430             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1431             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1432             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1433
1434             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1435             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1436             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1437
1438             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);
1439
1440             /* Inner loop uses 324 flops */
1441         }
1442
1443         if(jidx<j_index_end)
1444         {
1445
1446             jnrA             = jjnr[jidx];
1447             j_coord_offsetA  = DIM*jnrA;
1448
1449             /* load j atom coordinates */
1450             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1451                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1452                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1453
1454             /* Calculate displacement vector */
1455             dx00             = _mm_sub_pd(ix0,jx0);
1456             dy00             = _mm_sub_pd(iy0,jy0);
1457             dz00             = _mm_sub_pd(iz0,jz0);
1458             dx11             = _mm_sub_pd(ix1,jx1);
1459             dy11             = _mm_sub_pd(iy1,jy1);
1460             dz11             = _mm_sub_pd(iz1,jz1);
1461             dx12             = _mm_sub_pd(ix1,jx2);
1462             dy12             = _mm_sub_pd(iy1,jy2);
1463             dz12             = _mm_sub_pd(iz1,jz2);
1464             dx13             = _mm_sub_pd(ix1,jx3);
1465             dy13             = _mm_sub_pd(iy1,jy3);
1466             dz13             = _mm_sub_pd(iz1,jz3);
1467             dx21             = _mm_sub_pd(ix2,jx1);
1468             dy21             = _mm_sub_pd(iy2,jy1);
1469             dz21             = _mm_sub_pd(iz2,jz1);
1470             dx22             = _mm_sub_pd(ix2,jx2);
1471             dy22             = _mm_sub_pd(iy2,jy2);
1472             dz22             = _mm_sub_pd(iz2,jz2);
1473             dx23             = _mm_sub_pd(ix2,jx3);
1474             dy23             = _mm_sub_pd(iy2,jy3);
1475             dz23             = _mm_sub_pd(iz2,jz3);
1476             dx31             = _mm_sub_pd(ix3,jx1);
1477             dy31             = _mm_sub_pd(iy3,jy1);
1478             dz31             = _mm_sub_pd(iz3,jz1);
1479             dx32             = _mm_sub_pd(ix3,jx2);
1480             dy32             = _mm_sub_pd(iy3,jy2);
1481             dz32             = _mm_sub_pd(iz3,jz2);
1482             dx33             = _mm_sub_pd(ix3,jx3);
1483             dy33             = _mm_sub_pd(iy3,jy3);
1484             dz33             = _mm_sub_pd(iz3,jz3);
1485
1486             /* Calculate squared distance and things based on it */
1487             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1488             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1489             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1490             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1491             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1492             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1493             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1494             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1495             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1496             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1497
1498             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1499             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1500             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1501             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1502             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1503             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1504             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1505             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1506             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1507             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1508
1509             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1510             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1511             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1512             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1513             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1514             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1515             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1516             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1517             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1518
1519             fjx0             = _mm_setzero_pd();
1520             fjy0             = _mm_setzero_pd();
1521             fjz0             = _mm_setzero_pd();
1522             fjx1             = _mm_setzero_pd();
1523             fjy1             = _mm_setzero_pd();
1524             fjz1             = _mm_setzero_pd();
1525             fjx2             = _mm_setzero_pd();
1526             fjy2             = _mm_setzero_pd();
1527             fjz2             = _mm_setzero_pd();
1528             fjx3             = _mm_setzero_pd();
1529             fjy3             = _mm_setzero_pd();
1530             fjz3             = _mm_setzero_pd();
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             r00              = _mm_mul_pd(rsq00,rinv00);
1537
1538             /* Calculate table index by multiplying r with table scale and truncate to integer */
1539             rt               = _mm_mul_pd(r00,vftabscale);
1540             vfitab           = _mm_cvttpd_epi32(rt);
1541 #ifdef __XOP__
1542             vfeps            = _mm_frcz_pd(rt);
1543 #else
1544             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1545 #endif
1546             twovfeps         = _mm_add_pd(vfeps,vfeps);
1547             vfitab           = _mm_slli_epi32(vfitab,3);
1548
1549             /* CUBIC SPLINE TABLE DISPERSION */
1550             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1551             F                = _mm_setzero_pd();
1552             GMX_MM_TRANSPOSE2_PD(Y,F);
1553             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1554             H                = _mm_setzero_pd();
1555             GMX_MM_TRANSPOSE2_PD(G,H);
1556             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1557             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1558             fvdw6            = _mm_mul_pd(c6_00,FF);
1559
1560             /* CUBIC SPLINE TABLE REPULSION */
1561             vfitab           = _mm_add_epi32(vfitab,ifour);
1562             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1563             F                = _mm_setzero_pd();
1564             GMX_MM_TRANSPOSE2_PD(Y,F);
1565             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1566             H                = _mm_setzero_pd();
1567             GMX_MM_TRANSPOSE2_PD(G,H);
1568             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1569             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1570             fvdw12           = _mm_mul_pd(c12_00,FF);
1571             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1572
1573             fscal            = fvdw;
1574
1575             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1576
1577             /* Update vectorial force */
1578             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1579             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1580             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1581
1582             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1583             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1584             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1585
1586             /**************************
1587              * CALCULATE INTERACTIONS *
1588              **************************/
1589
1590             /* REACTION-FIELD ELECTROSTATICS */
1591             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1592
1593             fscal            = felec;
1594
1595             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1596
1597             /* Update vectorial force */
1598             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1599             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1600             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1601
1602             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1603             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1604             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1605
1606             /**************************
1607              * CALCULATE INTERACTIONS *
1608              **************************/
1609
1610             /* REACTION-FIELD ELECTROSTATICS */
1611             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1612
1613             fscal            = felec;
1614
1615             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1616
1617             /* Update vectorial force */
1618             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1619             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1620             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1621
1622             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1623             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1624             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1625
1626             /**************************
1627              * CALCULATE INTERACTIONS *
1628              **************************/
1629
1630             /* REACTION-FIELD ELECTROSTATICS */
1631             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1632
1633             fscal            = felec;
1634
1635             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1636
1637             /* Update vectorial force */
1638             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1639             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1640             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1641
1642             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1643             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1644             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1645
1646             /**************************
1647              * CALCULATE INTERACTIONS *
1648              **************************/
1649
1650             /* REACTION-FIELD ELECTROSTATICS */
1651             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1652
1653             fscal            = felec;
1654
1655             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1656
1657             /* Update vectorial force */
1658             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1659             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1660             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1661
1662             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1663             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1664             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1665
1666             /**************************
1667              * CALCULATE INTERACTIONS *
1668              **************************/
1669
1670             /* REACTION-FIELD ELECTROSTATICS */
1671             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1672
1673             fscal            = felec;
1674
1675             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1676
1677             /* Update vectorial force */
1678             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1679             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1680             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1681
1682             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1683             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1684             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1685
1686             /**************************
1687              * CALCULATE INTERACTIONS *
1688              **************************/
1689
1690             /* REACTION-FIELD ELECTROSTATICS */
1691             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1692
1693             fscal            = felec;
1694
1695             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1696
1697             /* Update vectorial force */
1698             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1699             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1700             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1701
1702             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1703             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1704             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1705
1706             /**************************
1707              * CALCULATE INTERACTIONS *
1708              **************************/
1709
1710             /* REACTION-FIELD ELECTROSTATICS */
1711             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1712
1713             fscal            = felec;
1714
1715             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1716
1717             /* Update vectorial force */
1718             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1719             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1720             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1721
1722             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1723             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1724             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1725
1726             /**************************
1727              * CALCULATE INTERACTIONS *
1728              **************************/
1729
1730             /* REACTION-FIELD ELECTROSTATICS */
1731             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1732
1733             fscal            = felec;
1734
1735             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1736
1737             /* Update vectorial force */
1738             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1739             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1740             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1741
1742             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1743             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1744             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1745
1746             /**************************
1747              * CALCULATE INTERACTIONS *
1748              **************************/
1749
1750             /* REACTION-FIELD ELECTROSTATICS */
1751             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1752
1753             fscal            = felec;
1754
1755             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1756
1757             /* Update vectorial force */
1758             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1759             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1760             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1761
1762             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1763             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1764             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1765
1766             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1767
1768             /* Inner loop uses 324 flops */
1769         }
1770
1771         /* End of innermost loop */
1772
1773         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1774                                               f+i_coord_offset,fshift+i_shift_offset);
1775
1776         /* Increment number of inner iterations */
1777         inneriter                  += j_index_end - j_index_start;
1778
1779         /* Outer loop uses 24 flops */
1780     }
1781
1782     /* Increment number of outer iterations */
1783     outeriter        += nri;
1784
1785     /* Update outer/inner flops */
1786
1787     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*324);
1788 }