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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     int              vdwjidx3A,vdwjidx3B;
  95     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  96     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  97     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 113     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 114     __m128i          vfitab;
 115     __m128i          ifour       = _mm_set1_epi32(4);
 116     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 117     real             *vftab;
 118     __m128d          dummy_mask,cutoff_mask;
 119     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 120     __m128d          one     = _mm_set1_pd(1.0);
 121     __m128d          two     = _mm_set1_pd(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_pd(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     krf              = _mm_set1_pd(fr->ic->k_rf);
 136     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 137     crf              = _mm_set1_pd(fr->ic->c_rf);
 138     nvdwtype         = fr->ntype;
 139     vdwparam         = fr->nbfp;
 140     vdwtype          = mdatoms->typeA;
 141
 142     vftab            = kernel_data->table_vdw->data;
 143     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 148     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 149     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     jq1              = _mm_set1_pd(charge[inr+1]);
 153     jq2              = _mm_set1_pd(charge[inr+2]);
 154     jq3              = _mm_set1_pd(charge[inr+3]);
 155     vdwjidx0A        = 2*vdwtype[inr+0];
 156     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 157     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 158     qq11             = _mm_mul_pd(iq1,jq1);
 159     qq12             = _mm_mul_pd(iq1,jq2);
 160     qq13             = _mm_mul_pd(iq1,jq3);
 161     qq21             = _mm_mul_pd(iq2,jq1);
 162     qq22             = _mm_mul_pd(iq2,jq2);
 163     qq23             = _mm_mul_pd(iq2,jq3);
 164     qq31             = _mm_mul_pd(iq3,jq1);
 165     qq32             = _mm_mul_pd(iq3,jq2);
 166     qq33             = _mm_mul_pd(iq3,jq3);
 167
 168     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172
 173     outeriter        = 0;
 174     inneriter        = 0;
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 193
 194         fix0             = _mm_setzero_pd();
 195         fiy0             = _mm_setzero_pd();
 196         fiz0             = _mm_setzero_pd();
 197         fix1             = _mm_setzero_pd();
 198         fiy1             = _mm_setzero_pd();
 199         fiz1             = _mm_setzero_pd();
 200         fix2             = _mm_setzero_pd();
 201         fiy2             = _mm_setzero_pd();
 202         fiz2             = _mm_setzero_pd();
 203         fix3             = _mm_setzero_pd();
 204         fiy3             = _mm_setzero_pd();
 205         fiz3             = _mm_setzero_pd();
 206
 207         /* Reset potential sums */
 208         velecsum         = _mm_setzero_pd();
 209         vvdwsum          = _mm_setzero_pd();
 210
 211         /* Start inner kernel loop */
 212         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 213         {
 214
 215             /* Get j neighbor index, and coordinate index */
 216             jnrA             = jjnr[jidx];
 217             jnrB             = jjnr[jidx+1];
 218             j_coord_offsetA  = DIM*jnrA;
 219             j_coord_offsetB  = DIM*jnrB;
 220
 221             /* load j atom coordinates */
 222             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 223                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 224                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 225
 226             /* Calculate displacement vector */
 227             dx00             = _mm_sub_pd(ix0,jx0);
 228             dy00             = _mm_sub_pd(iy0,jy0);
 229             dz00             = _mm_sub_pd(iz0,jz0);
 230             dx11             = _mm_sub_pd(ix1,jx1);
 231             dy11             = _mm_sub_pd(iy1,jy1);
 232             dz11             = _mm_sub_pd(iz1,jz1);
 233             dx12             = _mm_sub_pd(ix1,jx2);
 234             dy12             = _mm_sub_pd(iy1,jy2);
 235             dz12             = _mm_sub_pd(iz1,jz2);
 236             dx13             = _mm_sub_pd(ix1,jx3);
 237             dy13             = _mm_sub_pd(iy1,jy3);
 238             dz13             = _mm_sub_pd(iz1,jz3);
 239             dx21             = _mm_sub_pd(ix2,jx1);
 240             dy21             = _mm_sub_pd(iy2,jy1);
 241             dz21             = _mm_sub_pd(iz2,jz1);
 242             dx22             = _mm_sub_pd(ix2,jx2);
 243             dy22             = _mm_sub_pd(iy2,jy2);
 244             dz22             = _mm_sub_pd(iz2,jz2);
 245             dx23             = _mm_sub_pd(ix2,jx3);
 246             dy23             = _mm_sub_pd(iy2,jy3);
 247             dz23             = _mm_sub_pd(iz2,jz3);
 248             dx31             = _mm_sub_pd(ix3,jx1);
 249             dy31             = _mm_sub_pd(iy3,jy1);
 250             dz31             = _mm_sub_pd(iz3,jz1);
 251             dx32             = _mm_sub_pd(ix3,jx2);
 252             dy32             = _mm_sub_pd(iy3,jy2);
 253             dz32             = _mm_sub_pd(iz3,jz2);
 254             dx33             = _mm_sub_pd(ix3,jx3);
 255             dy33             = _mm_sub_pd(iy3,jy3);
 256             dz33             = _mm_sub_pd(iz3,jz3);
 257
 258             /* Calculate squared distance and things based on it */
 259             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 260             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 261             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 262             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 263             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 264             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 265             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 266             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 267             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 268             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 269
 270             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 271             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 272             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 273             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 274             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 275             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 276             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 277             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 278             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 279             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 280
 281             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 282             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 283             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 284             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 285             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 286             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 287             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 288             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 289             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 290
 291             fjx0             = _mm_setzero_pd();
 292             fjy0             = _mm_setzero_pd();
 293             fjz0             = _mm_setzero_pd();
 294             fjx1             = _mm_setzero_pd();
 295             fjy1             = _mm_setzero_pd();
 296             fjz1             = _mm_setzero_pd();
 297             fjx2             = _mm_setzero_pd();
 298             fjy2             = _mm_setzero_pd();
 299             fjz2             = _mm_setzero_pd();
 300             fjx3             = _mm_setzero_pd();
 301             fjy3             = _mm_setzero_pd();
 302             fjz3             = _mm_setzero_pd();
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             r00              = _mm_mul_pd(rsq00,rinv00);
 309
 310             /* Calculate table index by multiplying r with table scale and truncate to integer */
 311             rt               = _mm_mul_pd(r00,vftabscale);
 312             vfitab           = _mm_cvttpd_epi32(rt);
 313 #ifdef __XOP__
 314             vfeps            = _mm_frcz_pd(rt);
 315 #else
 316             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 317 #endif
 318             twovfeps         = _mm_add_pd(vfeps,vfeps);
 319             vfitab           = _mm_slli_epi32(vfitab,3);
 320
 321             /* CUBIC SPLINE TABLE DISPERSION */
 322             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 323             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 324             GMX_MM_TRANSPOSE2_PD(Y,F);
 325             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 326             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 327             GMX_MM_TRANSPOSE2_PD(G,H);
 328             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 329             VV               = _mm_macc_pd(vfeps,Fp,Y);
 330             vvdw6            = _mm_mul_pd(c6_00,VV);
 331             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 332             fvdw6            = _mm_mul_pd(c6_00,FF);
 333
 334             /* CUBIC SPLINE TABLE REPULSION */
 335             vfitab           = _mm_add_epi32(vfitab,ifour);
 336             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 337             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 338             GMX_MM_TRANSPOSE2_PD(Y,F);
 339             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 340             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 341             GMX_MM_TRANSPOSE2_PD(G,H);
 342             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 343             VV               = _mm_macc_pd(vfeps,Fp,Y);
 344             vvdw12           = _mm_mul_pd(c12_00,VV);
 345             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 346             fvdw12           = _mm_mul_pd(c12_00,FF);
 347             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 348             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 352
 353             fscal            = fvdw;
 354
 355             /* Update vectorial force */
 356             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 357             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 358             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 359
 360             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 361             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 362             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 363
 364             /**************************
 365              * CALCULATE INTERACTIONS *
 366              **************************/
 367
 368             /* REACTION-FIELD ELECTROSTATICS */
 369             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 370             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velecsum         = _mm_add_pd(velecsum,velec);
 374
 375             fscal            = felec;
 376
 377             /* Update vectorial force */
 378             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 379             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 380             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 381
 382             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 383             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 384             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             /* REACTION-FIELD ELECTROSTATICS */
 391             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 392             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 393
 394             /* Update potential sum for this i atom from the interaction with this j atom. */
 395             velecsum         = _mm_add_pd(velecsum,velec);
 396
 397             fscal            = felec;
 398
 399             /* Update vectorial force */
 400             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 401             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 402             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 403
 404             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 405             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 406             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             /* REACTION-FIELD ELECTROSTATICS */
 413             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 414             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 415
 416             /* Update potential sum for this i atom from the interaction with this j atom. */
 417             velecsum         = _mm_add_pd(velecsum,velec);
 418
 419             fscal            = felec;
 420
 421             /* Update vectorial force */
 422             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 423             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 424             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 425
 426             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 427             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 428             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             /* REACTION-FIELD ELECTROSTATICS */
 435             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 436             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 437
 438             /* Update potential sum for this i atom from the interaction with this j atom. */
 439             velecsum         = _mm_add_pd(velecsum,velec);
 440
 441             fscal            = felec;
 442
 443             /* Update vectorial force */
 444             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 445             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 446             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 447
 448             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 449             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 450             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             /* REACTION-FIELD ELECTROSTATICS */
 457             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 458             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velecsum         = _mm_add_pd(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             /* Update vectorial force */
 466             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 467             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 468             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 469
 470             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 471             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 472             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             /* REACTION-FIELD ELECTROSTATICS */
 479             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 480             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 481
 482             /* Update potential sum for this i atom from the interaction with this j atom. */
 483             velecsum         = _mm_add_pd(velecsum,velec);
 484
 485             fscal            = felec;
 486
 487             /* Update vectorial force */
 488             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 489             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 490             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 491
 492             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 493             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 494             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 502             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 503
 504             /* Update potential sum for this i atom from the interaction with this j atom. */
 505             velecsum         = _mm_add_pd(velecsum,velec);
 506
 507             fscal            = felec;
 508
 509             /* Update vectorial force */
 510             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 511             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 512             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 513
 514             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 515             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 516             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             /* REACTION-FIELD ELECTROSTATICS */
 523             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 524             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 525
 526             /* Update potential sum for this i atom from the interaction with this j atom. */
 527             velecsum         = _mm_add_pd(velecsum,velec);
 528
 529             fscal            = felec;
 530
 531             /* Update vectorial force */
 532             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 533             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 534             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 535
 536             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 537             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 538             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 539
 540             /**************************
 541              * CALCULATE INTERACTIONS *
 542              **************************/
 543
 544             /* REACTION-FIELD ELECTROSTATICS */
 545             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 546             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 547
 548             /* Update potential sum for this i atom from the interaction with this j atom. */
 549             velecsum         = _mm_add_pd(velecsum,velec);
 550
 551             fscal            = felec;
 552
 553             /* Update vectorial force */
 554             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 555             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 556             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 557
 558             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 559             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 560             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 561
 562             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);
 563
 564             /* Inner loop uses 377 flops */
 565         }
 566
 567         if(jidx<j_index_end)
 568         {
 569
 570             jnrA             = jjnr[jidx];
 571             j_coord_offsetA  = DIM*jnrA;
 572
 573             /* load j atom coordinates */
 574             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 575                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 576                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 577
 578             /* Calculate displacement vector */
 579             dx00             = _mm_sub_pd(ix0,jx0);
 580             dy00             = _mm_sub_pd(iy0,jy0);
 581             dz00             = _mm_sub_pd(iz0,jz0);
 582             dx11             = _mm_sub_pd(ix1,jx1);
 583             dy11             = _mm_sub_pd(iy1,jy1);
 584             dz11             = _mm_sub_pd(iz1,jz1);
 585             dx12             = _mm_sub_pd(ix1,jx2);
 586             dy12             = _mm_sub_pd(iy1,jy2);
 587             dz12             = _mm_sub_pd(iz1,jz2);
 588             dx13             = _mm_sub_pd(ix1,jx3);
 589             dy13             = _mm_sub_pd(iy1,jy3);
 590             dz13             = _mm_sub_pd(iz1,jz3);
 591             dx21             = _mm_sub_pd(ix2,jx1);
 592             dy21             = _mm_sub_pd(iy2,jy1);
 593             dz21             = _mm_sub_pd(iz2,jz1);
 594             dx22             = _mm_sub_pd(ix2,jx2);
 595             dy22             = _mm_sub_pd(iy2,jy2);
 596             dz22             = _mm_sub_pd(iz2,jz2);
 597             dx23             = _mm_sub_pd(ix2,jx3);
 598             dy23             = _mm_sub_pd(iy2,jy3);
 599             dz23             = _mm_sub_pd(iz2,jz3);
 600             dx31             = _mm_sub_pd(ix3,jx1);
 601             dy31             = _mm_sub_pd(iy3,jy1);
 602             dz31             = _mm_sub_pd(iz3,jz1);
 603             dx32             = _mm_sub_pd(ix3,jx2);
 604             dy32             = _mm_sub_pd(iy3,jy2);
 605             dz32             = _mm_sub_pd(iz3,jz2);
 606             dx33             = _mm_sub_pd(ix3,jx3);
 607             dy33             = _mm_sub_pd(iy3,jy3);
 608             dz33             = _mm_sub_pd(iz3,jz3);
 609
 610             /* Calculate squared distance and things based on it */
 611             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 612             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 613             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 614             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 615             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 616             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 617             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 618             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 619             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 620             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 621
 622             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 623             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 624             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 625             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 626             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 627             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 628             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 629             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 630             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 631             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 632
 633             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 634             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 635             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 636             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 637             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 638             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 639             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 640             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 641             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 642
 643             fjx0             = _mm_setzero_pd();
 644             fjy0             = _mm_setzero_pd();
 645             fjz0             = _mm_setzero_pd();
 646             fjx1             = _mm_setzero_pd();
 647             fjy1             = _mm_setzero_pd();
 648             fjz1             = _mm_setzero_pd();
 649             fjx2             = _mm_setzero_pd();
 650             fjy2             = _mm_setzero_pd();
 651             fjz2             = _mm_setzero_pd();
 652             fjx3             = _mm_setzero_pd();
 653             fjy3             = _mm_setzero_pd();
 654             fjz3             = _mm_setzero_pd();
 655
 656             /**************************
 657              * CALCULATE INTERACTIONS *
 658              **************************/
 659
 660             r00              = _mm_mul_pd(rsq00,rinv00);
 661
 662             /* Calculate table index by multiplying r with table scale and truncate to integer */
 663             rt               = _mm_mul_pd(r00,vftabscale);
 664             vfitab           = _mm_cvttpd_epi32(rt);
 665 #ifdef __XOP__
 666             vfeps            = _mm_frcz_pd(rt);
 667 #else
 668             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 669 #endif
 670             twovfeps         = _mm_add_pd(vfeps,vfeps);
 671             vfitab           = _mm_slli_epi32(vfitab,3);
 672
 673             /* CUBIC SPLINE TABLE DISPERSION */
 674             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 675             F                = _mm_setzero_pd();
 676             GMX_MM_TRANSPOSE2_PD(Y,F);
 677             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 678             H                = _mm_setzero_pd();
 679             GMX_MM_TRANSPOSE2_PD(G,H);
 680             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 681             VV               = _mm_macc_pd(vfeps,Fp,Y);
 682             vvdw6            = _mm_mul_pd(c6_00,VV);
 683             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 684             fvdw6            = _mm_mul_pd(c6_00,FF);
 685
 686             /* CUBIC SPLINE TABLE REPULSION */
 687             vfitab           = _mm_add_epi32(vfitab,ifour);
 688             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 689             F                = _mm_setzero_pd();
 690             GMX_MM_TRANSPOSE2_PD(Y,F);
 691             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 692             H                = _mm_setzero_pd();
 693             GMX_MM_TRANSPOSE2_PD(G,H);
 694             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 695             VV               = _mm_macc_pd(vfeps,Fp,Y);
 696             vvdw12           = _mm_mul_pd(c12_00,VV);
 697             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 698             fvdw12           = _mm_mul_pd(c12_00,FF);
 699             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 700             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 704             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 705
 706             fscal            = fvdw;
 707
 708             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 709
 710             /* Update vectorial force */
 711             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 712             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 713             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 714
 715             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 716             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 717             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 725             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 729             velecsum         = _mm_add_pd(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 734
 735             /* Update vectorial force */
 736             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 737             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 738             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 739
 740             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 741             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 742             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             /* REACTION-FIELD ELECTROSTATICS */
 749             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 750             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 751
 752             /* Update potential sum for this i atom from the interaction with this j atom. */
 753             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 754             velecsum         = _mm_add_pd(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 759
 760             /* Update vectorial force */
 761             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 762             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 763             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 764
 765             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 766             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 767             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* REACTION-FIELD ELECTROSTATICS */
 774             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 775             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 779             velecsum         = _mm_add_pd(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 784
 785             /* Update vectorial force */
 786             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 787             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 788             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 789
 790             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 791             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 792             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 800             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 804             velecsum         = _mm_add_pd(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 809
 810             /* Update vectorial force */
 811             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 812             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 813             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 814
 815             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 816             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 817             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             /* REACTION-FIELD ELECTROSTATICS */
 824             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 825             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 826
 827             /* Update potential sum for this i atom from the interaction with this j atom. */
 828             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 829             velecsum         = _mm_add_pd(velecsum,velec);
 830
 831             fscal            = felec;
 832
 833             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 834
 835             /* Update vectorial force */
 836             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 837             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 838             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 839
 840             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 841             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 842             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 850             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 854             velecsum         = _mm_add_pd(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 859
 860             /* Update vectorial force */
 861             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 862             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 863             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 864
 865             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 866             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 867             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 868
 869             /**************************
 870              * CALCULATE INTERACTIONS *
 871              **************************/
 872
 873             /* REACTION-FIELD ELECTROSTATICS */
 874             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 875             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 876
 877             /* Update potential sum for this i atom from the interaction with this j atom. */
 878             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 879             velecsum         = _mm_add_pd(velecsum,velec);
 880
 881             fscal            = felec;
 882
 883             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 884
 885             /* Update vectorial force */
 886             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 887             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 888             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 889
 890             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 891             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 892             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 900             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 901
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 904             velecsum         = _mm_add_pd(velecsum,velec);
 905
 906             fscal            = felec;
 907
 908             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 909
 910             /* Update vectorial force */
 911             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 912             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 913             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 914
 915             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 916             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 917             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             /* REACTION-FIELD ELECTROSTATICS */
 924             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 925             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 926
 927             /* Update potential sum for this i atom from the interaction with this j atom. */
 928             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 929             velecsum         = _mm_add_pd(velecsum,velec);
 930
 931             fscal            = felec;
 932
 933             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 934
 935             /* Update vectorial force */
 936             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 937             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 938             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 939
 940             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 941             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 942             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 943
 944             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 945
 946             /* Inner loop uses 377 flops */
 947         }
 948
 949         /* End of innermost loop */
 950
 951         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 952                                               f+i_coord_offset,fshift+i_shift_offset);
 953
 954         ggid                        = gid[iidx];
 955         /* Update potential energies */
 956         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 957         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 958
 959         /* Increment number of inner iterations */
 960         inneriter                  += j_index_end - j_index_start;
 961
 962         /* Outer loop uses 26 flops */
 963     }
 964
 965     /* Increment number of outer iterations */
 966     outeriter        += nri;
 967
 968     /* Update outer/inner flops */
 969
 970     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*377);
 971 }
 972 /*
 973  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 974  * Electrostatics interaction: ReactionField
 975  * VdW interaction:            CubicSplineTable
 976  * Geometry:                   Water4-Water4
 977  * Calculate force/pot:        Force
 978  */
 979 void
 980 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 981                     (t_nblist                    * gmx_restrict       nlist,
 982                      rvec                        * gmx_restrict          xx,
 983                      rvec                        * gmx_restrict          ff,
 984                      t_forcerec                  * gmx_restrict          fr,
 985                      t_mdatoms                   * gmx_restrict     mdatoms,
 986                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 987                      t_nrnb                      * gmx_restrict        nrnb)
 988 {
 989     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 990      * just 0 for non-waters.
 991      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 992      * jnr indices corresponding to data put in the four positions in the SIMD register.
 993      */
 994     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 995     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 996     int              jnrA,jnrB;
 997     int              j_coord_offsetA,j_coord_offsetB;
 998     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 999     real             rcutoff_scalar;
1000     real             *shiftvec,*fshift,*x,*f;
1001     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1002     int              vdwioffset0;
1003     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1004     int              vdwioffset1;
1005     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1006     int              vdwioffset2;
1007     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1008     int              vdwioffset3;
1009     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1010     int              vdwjidx0A,vdwjidx0B;
1011     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1012     int              vdwjidx1A,vdwjidx1B;
1013     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1014     int              vdwjidx2A,vdwjidx2B;
1015     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1016     int              vdwjidx3A,vdwjidx3B;
1017     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1018     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1019     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1020     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1021     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1022     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1023     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1024     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1025     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1026     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1027     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1028     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1029     real             *charge;
1030     int              nvdwtype;
1031     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1032     int              *vdwtype;
1033     real             *vdwparam;
1034     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1035     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1036     __m128i          vfitab;
1037     __m128i          ifour       = _mm_set1_epi32(4);
1038     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1039     real             *vftab;
1040     __m128d          dummy_mask,cutoff_mask;
1041     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1042     __m128d          one     = _mm_set1_pd(1.0);
1043     __m128d          two     = _mm_set1_pd(2.0);
1044     x                = xx[0];
1045     f                = ff[0];
1046
1047     nri              = nlist->nri;
1048     iinr             = nlist->iinr;
1049     jindex           = nlist->jindex;
1050     jjnr             = nlist->jjnr;
1051     shiftidx         = nlist->shift;
1052     gid              = nlist->gid;
1053     shiftvec         = fr->shift_vec[0];
1054     fshift           = fr->fshift[0];
1055     facel            = _mm_set1_pd(fr->epsfac);
1056     charge           = mdatoms->chargeA;
1057     krf              = _mm_set1_pd(fr->ic->k_rf);
1058     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1059     crf              = _mm_set1_pd(fr->ic->c_rf);
1060     nvdwtype         = fr->ntype;
1061     vdwparam         = fr->nbfp;
1062     vdwtype          = mdatoms->typeA;
1063
1064     vftab            = kernel_data->table_vdw->data;
1065     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1066
1067     /* Setup water-specific parameters */
1068     inr              = nlist->iinr[0];
1069     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1070     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1071     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1072     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1073
1074     jq1              = _mm_set1_pd(charge[inr+1]);
1075     jq2              = _mm_set1_pd(charge[inr+2]);
1076     jq3              = _mm_set1_pd(charge[inr+3]);
1077     vdwjidx0A        = 2*vdwtype[inr+0];
1078     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1079     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1080     qq11             = _mm_mul_pd(iq1,jq1);
1081     qq12             = _mm_mul_pd(iq1,jq2);
1082     qq13             = _mm_mul_pd(iq1,jq3);
1083     qq21             = _mm_mul_pd(iq2,jq1);
1084     qq22             = _mm_mul_pd(iq2,jq2);
1085     qq23             = _mm_mul_pd(iq2,jq3);
1086     qq31             = _mm_mul_pd(iq3,jq1);
1087     qq32             = _mm_mul_pd(iq3,jq2);
1088     qq33             = _mm_mul_pd(iq3,jq3);
1089
1090     /* Avoid stupid compiler warnings */
1091     jnrA = jnrB = 0;
1092     j_coord_offsetA = 0;
1093     j_coord_offsetB = 0;
1094
1095     outeriter        = 0;
1096     inneriter        = 0;
1097
1098     /* Start outer loop over neighborlists */
1099     for(iidx=0; iidx<nri; iidx++)
1100     {
1101         /* Load shift vector for this list */
1102         i_shift_offset   = DIM*shiftidx[iidx];
1103
1104         /* Load limits for loop over neighbors */
1105         j_index_start    = jindex[iidx];
1106         j_index_end      = jindex[iidx+1];
1107
1108         /* Get outer coordinate index */
1109         inr              = iinr[iidx];
1110         i_coord_offset   = DIM*inr;
1111
1112         /* Load i particle coords and add shift vector */
1113         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1114                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1115
1116         fix0             = _mm_setzero_pd();
1117         fiy0             = _mm_setzero_pd();
1118         fiz0             = _mm_setzero_pd();
1119         fix1             = _mm_setzero_pd();
1120         fiy1             = _mm_setzero_pd();
1121         fiz1             = _mm_setzero_pd();
1122         fix2             = _mm_setzero_pd();
1123         fiy2             = _mm_setzero_pd();
1124         fiz2             = _mm_setzero_pd();
1125         fix3             = _mm_setzero_pd();
1126         fiy3             = _mm_setzero_pd();
1127         fiz3             = _mm_setzero_pd();
1128
1129         /* Start inner kernel loop */
1130         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1131         {
1132
1133             /* Get j neighbor index, and coordinate index */
1134             jnrA             = jjnr[jidx];
1135             jnrB             = jjnr[jidx+1];
1136             j_coord_offsetA  = DIM*jnrA;
1137             j_coord_offsetB  = DIM*jnrB;
1138
1139             /* load j atom coordinates */
1140             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1141                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1142                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1143
1144             /* Calculate displacement vector */
1145             dx00             = _mm_sub_pd(ix0,jx0);
1146             dy00             = _mm_sub_pd(iy0,jy0);
1147             dz00             = _mm_sub_pd(iz0,jz0);
1148             dx11             = _mm_sub_pd(ix1,jx1);
1149             dy11             = _mm_sub_pd(iy1,jy1);
1150             dz11             = _mm_sub_pd(iz1,jz1);
1151             dx12             = _mm_sub_pd(ix1,jx2);
1152             dy12             = _mm_sub_pd(iy1,jy2);
1153             dz12             = _mm_sub_pd(iz1,jz2);
1154             dx13             = _mm_sub_pd(ix1,jx3);
1155             dy13             = _mm_sub_pd(iy1,jy3);
1156             dz13             = _mm_sub_pd(iz1,jz3);
1157             dx21             = _mm_sub_pd(ix2,jx1);
1158             dy21             = _mm_sub_pd(iy2,jy1);
1159             dz21             = _mm_sub_pd(iz2,jz1);
1160             dx22             = _mm_sub_pd(ix2,jx2);
1161             dy22             = _mm_sub_pd(iy2,jy2);
1162             dz22             = _mm_sub_pd(iz2,jz2);
1163             dx23             = _mm_sub_pd(ix2,jx3);
1164             dy23             = _mm_sub_pd(iy2,jy3);
1165             dz23             = _mm_sub_pd(iz2,jz3);
1166             dx31             = _mm_sub_pd(ix3,jx1);
1167             dy31             = _mm_sub_pd(iy3,jy1);
1168             dz31             = _mm_sub_pd(iz3,jz1);
1169             dx32             = _mm_sub_pd(ix3,jx2);
1170             dy32             = _mm_sub_pd(iy3,jy2);
1171             dz32             = _mm_sub_pd(iz3,jz2);
1172             dx33             = _mm_sub_pd(ix3,jx3);
1173             dy33             = _mm_sub_pd(iy3,jy3);
1174             dz33             = _mm_sub_pd(iz3,jz3);
1175
1176             /* Calculate squared distance and things based on it */
1177             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1178             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1179             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1180             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1181             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1182             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1183             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1184             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1185             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1186             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1187
1188             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1189             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1190             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1191             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1192             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1193             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1194             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1195             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1196             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1197             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1198
1199             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1200             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1201             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1202             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1203             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1204             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1205             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1206             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1207             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1208
1209             fjx0             = _mm_setzero_pd();
1210             fjy0             = _mm_setzero_pd();
1211             fjz0             = _mm_setzero_pd();
1212             fjx1             = _mm_setzero_pd();
1213             fjy1             = _mm_setzero_pd();
1214             fjz1             = _mm_setzero_pd();
1215             fjx2             = _mm_setzero_pd();
1216             fjy2             = _mm_setzero_pd();
1217             fjz2             = _mm_setzero_pd();
1218             fjx3             = _mm_setzero_pd();
1219             fjy3             = _mm_setzero_pd();
1220             fjz3             = _mm_setzero_pd();
1221
1222             /**************************
1223              * CALCULATE INTERACTIONS *
1224              **************************/
1225
1226             r00              = _mm_mul_pd(rsq00,rinv00);
1227
1228             /* Calculate table index by multiplying r with table scale and truncate to integer */
1229             rt               = _mm_mul_pd(r00,vftabscale);
1230             vfitab           = _mm_cvttpd_epi32(rt);
1231 #ifdef __XOP__
1232             vfeps            = _mm_frcz_pd(rt);
1233 #else
1234             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1235 #endif
1236             twovfeps         = _mm_add_pd(vfeps,vfeps);
1237             vfitab           = _mm_slli_epi32(vfitab,3);
1238
1239             /* CUBIC SPLINE TABLE DISPERSION */
1240             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1241             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1242             GMX_MM_TRANSPOSE2_PD(Y,F);
1243             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1244             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1245             GMX_MM_TRANSPOSE2_PD(G,H);
1246             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1247             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1248             fvdw6            = _mm_mul_pd(c6_00,FF);
1249
1250             /* CUBIC SPLINE TABLE REPULSION */
1251             vfitab           = _mm_add_epi32(vfitab,ifour);
1252             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1253             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1254             GMX_MM_TRANSPOSE2_PD(Y,F);
1255             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1256             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1257             GMX_MM_TRANSPOSE2_PD(G,H);
1258             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1259             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1260             fvdw12           = _mm_mul_pd(c12_00,FF);
1261             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1262
1263             fscal            = fvdw;
1264
1265             /* Update vectorial force */
1266             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1267             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1268             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1269
1270             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1271             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1272             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1273
1274             /**************************
1275              * CALCULATE INTERACTIONS *
1276              **************************/
1277
1278             /* REACTION-FIELD ELECTROSTATICS */
1279             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1280
1281             fscal            = felec;
1282
1283             /* Update vectorial force */
1284             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1285             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1286             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1287
1288             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1289             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1290             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* REACTION-FIELD ELECTROSTATICS */
1297             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1298
1299             fscal            = felec;
1300
1301             /* Update vectorial force */
1302             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1303             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1304             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1305
1306             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1307             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1308             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1309
1310             /**************************
1311              * CALCULATE INTERACTIONS *
1312              **************************/
1313
1314             /* REACTION-FIELD ELECTROSTATICS */
1315             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1316
1317             fscal            = felec;
1318
1319             /* Update vectorial force */
1320             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1321             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1322             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1323
1324             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1325             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1326             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1327
1328             /**************************
1329              * CALCULATE INTERACTIONS *
1330              **************************/
1331
1332             /* REACTION-FIELD ELECTROSTATICS */
1333             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1334
1335             fscal            = felec;
1336
1337             /* Update vectorial force */
1338             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1339             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1340             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1341
1342             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1343             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1344             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1345
1346             /**************************
1347              * CALCULATE INTERACTIONS *
1348              **************************/
1349
1350             /* REACTION-FIELD ELECTROSTATICS */
1351             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1352
1353             fscal            = felec;
1354
1355             /* Update vectorial force */
1356             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1357             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1358             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1359
1360             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1361             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1362             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1363
1364             /**************************
1365              * CALCULATE INTERACTIONS *
1366              **************************/
1367
1368             /* REACTION-FIELD ELECTROSTATICS */
1369             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1370
1371             fscal            = felec;
1372
1373             /* Update vectorial force */
1374             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1375             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1376             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1377
1378             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1379             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1380             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1381
1382             /**************************
1383              * CALCULATE INTERACTIONS *
1384              **************************/
1385
1386             /* REACTION-FIELD ELECTROSTATICS */
1387             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1388
1389             fscal            = felec;
1390
1391             /* Update vectorial force */
1392             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1393             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1394             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1395
1396             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1397             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1398             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1399
1400             /**************************
1401              * CALCULATE INTERACTIONS *
1402              **************************/
1403
1404             /* REACTION-FIELD ELECTROSTATICS */
1405             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1406
1407             fscal            = felec;
1408
1409             /* Update vectorial force */
1410             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1411             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1412             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1413
1414             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1415             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1416             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* REACTION-FIELD ELECTROSTATICS */
1423             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1424
1425             fscal            = felec;
1426
1427             /* Update vectorial force */
1428             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1429             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1430             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1431
1432             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1433             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1434             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1435
1436             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);
1437
1438             /* Inner loop uses 324 flops */
1439         }
1440
1441         if(jidx<j_index_end)
1442         {
1443
1444             jnrA             = jjnr[jidx];
1445             j_coord_offsetA  = DIM*jnrA;
1446
1447             /* load j atom coordinates */
1448             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1449                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1450                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1451
1452             /* Calculate displacement vector */
1453             dx00             = _mm_sub_pd(ix0,jx0);
1454             dy00             = _mm_sub_pd(iy0,jy0);
1455             dz00             = _mm_sub_pd(iz0,jz0);
1456             dx11             = _mm_sub_pd(ix1,jx1);
1457             dy11             = _mm_sub_pd(iy1,jy1);
1458             dz11             = _mm_sub_pd(iz1,jz1);
1459             dx12             = _mm_sub_pd(ix1,jx2);
1460             dy12             = _mm_sub_pd(iy1,jy2);
1461             dz12             = _mm_sub_pd(iz1,jz2);
1462             dx13             = _mm_sub_pd(ix1,jx3);
1463             dy13             = _mm_sub_pd(iy1,jy3);
1464             dz13             = _mm_sub_pd(iz1,jz3);
1465             dx21             = _mm_sub_pd(ix2,jx1);
1466             dy21             = _mm_sub_pd(iy2,jy1);
1467             dz21             = _mm_sub_pd(iz2,jz1);
1468             dx22             = _mm_sub_pd(ix2,jx2);
1469             dy22             = _mm_sub_pd(iy2,jy2);
1470             dz22             = _mm_sub_pd(iz2,jz2);
1471             dx23             = _mm_sub_pd(ix2,jx3);
1472             dy23             = _mm_sub_pd(iy2,jy3);
1473             dz23             = _mm_sub_pd(iz2,jz3);
1474             dx31             = _mm_sub_pd(ix3,jx1);
1475             dy31             = _mm_sub_pd(iy3,jy1);
1476             dz31             = _mm_sub_pd(iz3,jz1);
1477             dx32             = _mm_sub_pd(ix3,jx2);
1478             dy32             = _mm_sub_pd(iy3,jy2);
1479             dz32             = _mm_sub_pd(iz3,jz2);
1480             dx33             = _mm_sub_pd(ix3,jx3);
1481             dy33             = _mm_sub_pd(iy3,jy3);
1482             dz33             = _mm_sub_pd(iz3,jz3);
1483
1484             /* Calculate squared distance and things based on it */
1485             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1486             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1487             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1488             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1489             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1490             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1491             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1492             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1493             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1494             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1495
1496             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1497             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1498             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1499             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1500             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1501             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1502             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1503             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1504             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1505             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1506
1507             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1508             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1509             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1510             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1511             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1512             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1513             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1514             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1515             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1516
1517             fjx0             = _mm_setzero_pd();
1518             fjy0             = _mm_setzero_pd();
1519             fjz0             = _mm_setzero_pd();
1520             fjx1             = _mm_setzero_pd();
1521             fjy1             = _mm_setzero_pd();
1522             fjz1             = _mm_setzero_pd();
1523             fjx2             = _mm_setzero_pd();
1524             fjy2             = _mm_setzero_pd();
1525             fjz2             = _mm_setzero_pd();
1526             fjx3             = _mm_setzero_pd();
1527             fjy3             = _mm_setzero_pd();
1528             fjz3             = _mm_setzero_pd();
1529
1530             /**************************
1531              * CALCULATE INTERACTIONS *
1532              **************************/
1533
1534             r00              = _mm_mul_pd(rsq00,rinv00);
1535
1536             /* Calculate table index by multiplying r with table scale and truncate to integer */
1537             rt               = _mm_mul_pd(r00,vftabscale);
1538             vfitab           = _mm_cvttpd_epi32(rt);
1539 #ifdef __XOP__
1540             vfeps            = _mm_frcz_pd(rt);
1541 #else
1542             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1543 #endif
1544             twovfeps         = _mm_add_pd(vfeps,vfeps);
1545             vfitab           = _mm_slli_epi32(vfitab,3);
1546
1547             /* CUBIC SPLINE TABLE DISPERSION */
1548             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1549             F                = _mm_setzero_pd();
1550             GMX_MM_TRANSPOSE2_PD(Y,F);
1551             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1552             H                = _mm_setzero_pd();
1553             GMX_MM_TRANSPOSE2_PD(G,H);
1554             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1555             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1556             fvdw6            = _mm_mul_pd(c6_00,FF);
1557
1558             /* CUBIC SPLINE TABLE REPULSION */
1559             vfitab           = _mm_add_epi32(vfitab,ifour);
1560             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1561             F                = _mm_setzero_pd();
1562             GMX_MM_TRANSPOSE2_PD(Y,F);
1563             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1564             H                = _mm_setzero_pd();
1565             GMX_MM_TRANSPOSE2_PD(G,H);
1566             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1567             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1568             fvdw12           = _mm_mul_pd(c12_00,FF);
1569             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1570
1571             fscal            = fvdw;
1572
1573             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1574
1575             /* Update vectorial force */
1576             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1577             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1578             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1579
1580             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1581             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1582             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1583
1584             /**************************
1585              * CALCULATE INTERACTIONS *
1586              **************************/
1587
1588             /* REACTION-FIELD ELECTROSTATICS */
1589             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1590
1591             fscal            = felec;
1592
1593             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1594
1595             /* Update vectorial force */
1596             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1597             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1598             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1599
1600             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1601             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1602             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1603
1604             /**************************
1605              * CALCULATE INTERACTIONS *
1606              **************************/
1607
1608             /* REACTION-FIELD ELECTROSTATICS */
1609             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1610
1611             fscal            = felec;
1612
1613             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1614
1615             /* Update vectorial force */
1616             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1617             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1618             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1619
1620             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1621             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1622             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1623
1624             /**************************
1625              * CALCULATE INTERACTIONS *
1626              **************************/
1627
1628             /* REACTION-FIELD ELECTROSTATICS */
1629             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1630
1631             fscal            = felec;
1632
1633             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1634
1635             /* Update vectorial force */
1636             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1637             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1638             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1639
1640             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1641             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1642             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1643
1644             /**************************
1645              * CALCULATE INTERACTIONS *
1646              **************************/
1647
1648             /* REACTION-FIELD ELECTROSTATICS */
1649             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1650
1651             fscal            = felec;
1652
1653             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1654
1655             /* Update vectorial force */
1656             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1657             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1658             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1659
1660             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1661             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1662             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1663
1664             /**************************
1665              * CALCULATE INTERACTIONS *
1666              **************************/
1667
1668             /* REACTION-FIELD ELECTROSTATICS */
1669             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1670
1671             fscal            = felec;
1672
1673             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1674
1675             /* Update vectorial force */
1676             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1677             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1678             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1679
1680             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1681             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1682             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1683
1684             /**************************
1685              * CALCULATE INTERACTIONS *
1686              **************************/
1687
1688             /* REACTION-FIELD ELECTROSTATICS */
1689             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1690
1691             fscal            = felec;
1692
1693             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1694
1695             /* Update vectorial force */
1696             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1697             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1698             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1699
1700             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1701             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1702             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1703
1704             /**************************
1705              * CALCULATE INTERACTIONS *
1706              **************************/
1707
1708             /* REACTION-FIELD ELECTROSTATICS */
1709             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1710
1711             fscal            = felec;
1712
1713             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1714
1715             /* Update vectorial force */
1716             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1717             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1718             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1719
1720             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1721             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1722             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1723
1724             /**************************
1725              * CALCULATE INTERACTIONS *
1726              **************************/
1727
1728             /* REACTION-FIELD ELECTROSTATICS */
1729             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1730
1731             fscal            = felec;
1732
1733             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1734
1735             /* Update vectorial force */
1736             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1737             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1738             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1739
1740             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1741             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1742             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1743
1744             /**************************
1745              * CALCULATE INTERACTIONS *
1746              **************************/
1747
1748             /* REACTION-FIELD ELECTROSTATICS */
1749             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1750
1751             fscal            = felec;
1752
1753             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1754
1755             /* Update vectorial force */
1756             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1757             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1758             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1759
1760             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1761             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1762             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1763
1764             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1765
1766             /* Inner loop uses 324 flops */
1767         }
1768
1769         /* End of innermost loop */
1770
1771         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1772                                               f+i_coord_offset,fshift+i_shift_offset);
1773
1774         /* Increment number of inner iterations */
1775         inneriter                  += j_index_end - j_index_start;
1776
1777         /* Outer loop uses 24 flops */
1778     }
1779
1780     /* Increment number of outer iterations */
1781     outeriter        += nri;
1782
1783     /* Update outer/inner flops */
1784
1785     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*324);
1786 }