src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwCSTab_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwioffset3;
  73     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B;
  75     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B;
  77     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B;
  79     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B;
  81     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  99     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 103     real             *vftab;
 104     __m128d          dummy_mask,cutoff_mask;
 105     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 106     __m128d          one     = _mm_set1_pd(1.0);
 107     __m128d          two     = _mm_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     krf              = _mm_set1_pd(fr->ic->k_rf);
 122     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 123     crf              = _mm_set1_pd(fr->ic->c_rf);
 124     nvdwtype         = fr->ntype;
 125     vdwparam         = fr->nbfp;
 126     vdwtype          = mdatoms->typeA;
 127
 128     vftab            = kernel_data->table_vdw->data;
 129     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 134     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 135     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 136     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 137
 138     jq1              = _mm_set1_pd(charge[inr+1]);
 139     jq2              = _mm_set1_pd(charge[inr+2]);
 140     jq3              = _mm_set1_pd(charge[inr+3]);
 141     vdwjidx0A        = 2*vdwtype[inr+0];
 142     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 143     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 144     qq11             = _mm_mul_pd(iq1,jq1);
 145     qq12             = _mm_mul_pd(iq1,jq2);
 146     qq13             = _mm_mul_pd(iq1,jq3);
 147     qq21             = _mm_mul_pd(iq2,jq1);
 148     qq22             = _mm_mul_pd(iq2,jq2);
 149     qq23             = _mm_mul_pd(iq2,jq3);
 150     qq31             = _mm_mul_pd(iq3,jq1);
 151     qq32             = _mm_mul_pd(iq3,jq2);
 152     qq33             = _mm_mul_pd(iq3,jq3);
 153
 154     /* Avoid stupid compiler warnings */
 155     jnrA = jnrB = 0;
 156     j_coord_offsetA = 0;
 157     j_coord_offsetB = 0;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 178                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 179
 180         fix0             = _mm_setzero_pd();
 181         fiy0             = _mm_setzero_pd();
 182         fiz0             = _mm_setzero_pd();
 183         fix1             = _mm_setzero_pd();
 184         fiy1             = _mm_setzero_pd();
 185         fiz1             = _mm_setzero_pd();
 186         fix2             = _mm_setzero_pd();
 187         fiy2             = _mm_setzero_pd();
 188         fiz2             = _mm_setzero_pd();
 189         fix3             = _mm_setzero_pd();
 190         fiy3             = _mm_setzero_pd();
 191         fiz3             = _mm_setzero_pd();
 192
 193         /* Reset potential sums */
 194         velecsum         = _mm_setzero_pd();
 195         vvdwsum          = _mm_setzero_pd();
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 199         {
 200
 201             /* Get j neighbor index, and coordinate index */
 202             jnrA             = jjnr[jidx];
 203             jnrB             = jjnr[jidx+1];
 204             j_coord_offsetA  = DIM*jnrA;
 205             j_coord_offsetB  = DIM*jnrB;
 206
 207             /* load j atom coordinates */
 208             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 209                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 210                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 211
 212             /* Calculate displacement vector */
 213             dx00             = _mm_sub_pd(ix0,jx0);
 214             dy00             = _mm_sub_pd(iy0,jy0);
 215             dz00             = _mm_sub_pd(iz0,jz0);
 216             dx11             = _mm_sub_pd(ix1,jx1);
 217             dy11             = _mm_sub_pd(iy1,jy1);
 218             dz11             = _mm_sub_pd(iz1,jz1);
 219             dx12             = _mm_sub_pd(ix1,jx2);
 220             dy12             = _mm_sub_pd(iy1,jy2);
 221             dz12             = _mm_sub_pd(iz1,jz2);
 222             dx13             = _mm_sub_pd(ix1,jx3);
 223             dy13             = _mm_sub_pd(iy1,jy3);
 224             dz13             = _mm_sub_pd(iz1,jz3);
 225             dx21             = _mm_sub_pd(ix2,jx1);
 226             dy21             = _mm_sub_pd(iy2,jy1);
 227             dz21             = _mm_sub_pd(iz2,jz1);
 228             dx22             = _mm_sub_pd(ix2,jx2);
 229             dy22             = _mm_sub_pd(iy2,jy2);
 230             dz22             = _mm_sub_pd(iz2,jz2);
 231             dx23             = _mm_sub_pd(ix2,jx3);
 232             dy23             = _mm_sub_pd(iy2,jy3);
 233             dz23             = _mm_sub_pd(iz2,jz3);
 234             dx31             = _mm_sub_pd(ix3,jx1);
 235             dy31             = _mm_sub_pd(iy3,jy1);
 236             dz31             = _mm_sub_pd(iz3,jz1);
 237             dx32             = _mm_sub_pd(ix3,jx2);
 238             dy32             = _mm_sub_pd(iy3,jy2);
 239             dz32             = _mm_sub_pd(iz3,jz2);
 240             dx33             = _mm_sub_pd(ix3,jx3);
 241             dy33             = _mm_sub_pd(iy3,jy3);
 242             dz33             = _mm_sub_pd(iz3,jz3);
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 246             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 247             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 248             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 249             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 250             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 251             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 252             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 253             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 254             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 255
 256             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 257             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 258             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 259             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 260             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 261             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 262             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 263             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 264             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 265             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 266
 267             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 268             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 269             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 270             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 271             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 272             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 273             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 274             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 275             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 276
 277             fjx0             = _mm_setzero_pd();
 278             fjy0             = _mm_setzero_pd();
 279             fjz0             = _mm_setzero_pd();
 280             fjx1             = _mm_setzero_pd();
 281             fjy1             = _mm_setzero_pd();
 282             fjz1             = _mm_setzero_pd();
 283             fjx2             = _mm_setzero_pd();
 284             fjy2             = _mm_setzero_pd();
 285             fjz2             = _mm_setzero_pd();
 286             fjx3             = _mm_setzero_pd();
 287             fjy3             = _mm_setzero_pd();
 288             fjz3             = _mm_setzero_pd();
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             r00              = _mm_mul_pd(rsq00,rinv00);
 295
 296             /* Calculate table index by multiplying r with table scale and truncate to integer */
 297             rt               = _mm_mul_pd(r00,vftabscale);
 298             vfitab           = _mm_cvttpd_epi32(rt);
 299 #ifdef __XOP__
 300             vfeps            = _mm_frcz_pd(rt);
 301 #else
 302             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 303 #endif
 304             twovfeps         = _mm_add_pd(vfeps,vfeps);
 305             vfitab           = _mm_slli_epi32(vfitab,3);
 306
 307             /* CUBIC SPLINE TABLE DISPERSION */
 308             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 309             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 310             GMX_MM_TRANSPOSE2_PD(Y,F);
 311             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 312             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 313             GMX_MM_TRANSPOSE2_PD(G,H);
 314             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 315             VV               = _mm_macc_pd(vfeps,Fp,Y);
 316             vvdw6            = _mm_mul_pd(c6_00,VV);
 317             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 318             fvdw6            = _mm_mul_pd(c6_00,FF);
 319
 320             /* CUBIC SPLINE TABLE REPULSION */
 321             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm_mul_pd(c12_00,VV);
 331             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 332             fvdw12           = _mm_mul_pd(c12_00,FF);
 333             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 334             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 338
 339             fscal            = fvdw;
 340
 341             /* Update vectorial force */
 342             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 343             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 344             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 345
 346             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 347             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 348             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 349
 350             /**************************
 351              * CALCULATE INTERACTIONS *
 352              **************************/
 353
 354             /* REACTION-FIELD ELECTROSTATICS */
 355             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 356             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velecsum         = _mm_add_pd(velecsum,velec);
 360
 361             fscal            = felec;
 362
 363             /* Update vectorial force */
 364             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 365             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 366             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 367
 368             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 369             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 370             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 371
 372             /**************************
 373              * CALCULATE INTERACTIONS *
 374              **************************/
 375
 376             /* REACTION-FIELD ELECTROSTATICS */
 377             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 378             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             /* Update vectorial force */
 386             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 387             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 388             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 389
 390             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 391             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 392             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             /* REACTION-FIELD ELECTROSTATICS */
 399             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 400             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velecsum         = _mm_add_pd(velecsum,velec);
 404
 405             fscal            = felec;
 406
 407             /* Update vectorial force */
 408             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 409             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 410             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 411
 412             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 413             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 414             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* REACTION-FIELD ELECTROSTATICS */
 421             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 422             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 423
 424             /* Update potential sum for this i atom from the interaction with this j atom. */
 425             velecsum         = _mm_add_pd(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             /* Update vectorial force */
 430             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 431             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 432             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 433
 434             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 435             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 436             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 437
 438             /**************************
 439              * CALCULATE INTERACTIONS *
 440              **************************/
 441
 442             /* REACTION-FIELD ELECTROSTATICS */
 443             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 444             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velecsum         = _mm_add_pd(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             /* Update vectorial force */
 452             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 453             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 454             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 455
 456             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 457             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 458             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 459
 460             /**************************
 461              * CALCULATE INTERACTIONS *
 462              **************************/
 463
 464             /* REACTION-FIELD ELECTROSTATICS */
 465             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 466             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 467
 468             /* Update potential sum for this i atom from the interaction with this j atom. */
 469             velecsum         = _mm_add_pd(velecsum,velec);
 470
 471             fscal            = felec;
 472
 473             /* Update vectorial force */
 474             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 475             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 476             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 477
 478             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 479             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 480             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             /* REACTION-FIELD ELECTROSTATICS */
 487             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 488             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 489
 490             /* Update potential sum for this i atom from the interaction with this j atom. */
 491             velecsum         = _mm_add_pd(velecsum,velec);
 492
 493             fscal            = felec;
 494
 495             /* Update vectorial force */
 496             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 497             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 498             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 499
 500             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 501             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 502             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 503
 504             /**************************
 505              * CALCULATE INTERACTIONS *
 506              **************************/
 507
 508             /* REACTION-FIELD ELECTROSTATICS */
 509             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 510             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 511
 512             /* Update potential sum for this i atom from the interaction with this j atom. */
 513             velecsum         = _mm_add_pd(velecsum,velec);
 514
 515             fscal            = felec;
 516
 517             /* Update vectorial force */
 518             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 519             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 520             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 521
 522             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 523             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 524             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 525
 526             /**************************
 527              * CALCULATE INTERACTIONS *
 528              **************************/
 529
 530             /* REACTION-FIELD ELECTROSTATICS */
 531             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 532             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velecsum         = _mm_add_pd(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539             /* Update vectorial force */
 540             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 541             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 542             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 543
 544             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 545             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 546             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 547
 548             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);
 549
 550             /* Inner loop uses 377 flops */
 551         }
 552
 553         if(jidx<j_index_end)
 554         {
 555
 556             jnrA             = jjnr[jidx];
 557             j_coord_offsetA  = DIM*jnrA;
 558
 559             /* load j atom coordinates */
 560             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 561                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 562                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 563
 564             /* Calculate displacement vector */
 565             dx00             = _mm_sub_pd(ix0,jx0);
 566             dy00             = _mm_sub_pd(iy0,jy0);
 567             dz00             = _mm_sub_pd(iz0,jz0);
 568             dx11             = _mm_sub_pd(ix1,jx1);
 569             dy11             = _mm_sub_pd(iy1,jy1);
 570             dz11             = _mm_sub_pd(iz1,jz1);
 571             dx12             = _mm_sub_pd(ix1,jx2);
 572             dy12             = _mm_sub_pd(iy1,jy2);
 573             dz12             = _mm_sub_pd(iz1,jz2);
 574             dx13             = _mm_sub_pd(ix1,jx3);
 575             dy13             = _mm_sub_pd(iy1,jy3);
 576             dz13             = _mm_sub_pd(iz1,jz3);
 577             dx21             = _mm_sub_pd(ix2,jx1);
 578             dy21             = _mm_sub_pd(iy2,jy1);
 579             dz21             = _mm_sub_pd(iz2,jz1);
 580             dx22             = _mm_sub_pd(ix2,jx2);
 581             dy22             = _mm_sub_pd(iy2,jy2);
 582             dz22             = _mm_sub_pd(iz2,jz2);
 583             dx23             = _mm_sub_pd(ix2,jx3);
 584             dy23             = _mm_sub_pd(iy2,jy3);
 585             dz23             = _mm_sub_pd(iz2,jz3);
 586             dx31             = _mm_sub_pd(ix3,jx1);
 587             dy31             = _mm_sub_pd(iy3,jy1);
 588             dz31             = _mm_sub_pd(iz3,jz1);
 589             dx32             = _mm_sub_pd(ix3,jx2);
 590             dy32             = _mm_sub_pd(iy3,jy2);
 591             dz32             = _mm_sub_pd(iz3,jz2);
 592             dx33             = _mm_sub_pd(ix3,jx3);
 593             dy33             = _mm_sub_pd(iy3,jy3);
 594             dz33             = _mm_sub_pd(iz3,jz3);
 595
 596             /* Calculate squared distance and things based on it */
 597             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 598             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 599             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 600             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 601             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 602             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 603             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 604             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 605             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 606             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 607
 608             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 609             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 610             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 611             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 612             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 613             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 614             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 615             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 616             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 617             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 618
 619             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 620             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 621             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 622             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 623             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 624             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 625             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 626             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 627             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 628
 629             fjx0             = _mm_setzero_pd();
 630             fjy0             = _mm_setzero_pd();
 631             fjz0             = _mm_setzero_pd();
 632             fjx1             = _mm_setzero_pd();
 633             fjy1             = _mm_setzero_pd();
 634             fjz1             = _mm_setzero_pd();
 635             fjx2             = _mm_setzero_pd();
 636             fjy2             = _mm_setzero_pd();
 637             fjz2             = _mm_setzero_pd();
 638             fjx3             = _mm_setzero_pd();
 639             fjy3             = _mm_setzero_pd();
 640             fjz3             = _mm_setzero_pd();
 641
 642             /**************************
 643              * CALCULATE INTERACTIONS *
 644              **************************/
 645
 646             r00              = _mm_mul_pd(rsq00,rinv00);
 647
 648             /* Calculate table index by multiplying r with table scale and truncate to integer */
 649             rt               = _mm_mul_pd(r00,vftabscale);
 650             vfitab           = _mm_cvttpd_epi32(rt);
 651 #ifdef __XOP__
 652             vfeps            = _mm_frcz_pd(rt);
 653 #else
 654             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 655 #endif
 656             twovfeps         = _mm_add_pd(vfeps,vfeps);
 657             vfitab           = _mm_slli_epi32(vfitab,3);
 658
 659             /* CUBIC SPLINE TABLE DISPERSION */
 660             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 661             F                = _mm_setzero_pd();
 662             GMX_MM_TRANSPOSE2_PD(Y,F);
 663             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 664             H                = _mm_setzero_pd();
 665             GMX_MM_TRANSPOSE2_PD(G,H);
 666             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 667             VV               = _mm_macc_pd(vfeps,Fp,Y);
 668             vvdw6            = _mm_mul_pd(c6_00,VV);
 669             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 670             fvdw6            = _mm_mul_pd(c6_00,FF);
 671
 672             /* CUBIC SPLINE TABLE REPULSION */
 673             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm_mul_pd(c12_00,VV);
 683             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 684             fvdw12           = _mm_mul_pd(c12_00,FF);
 685             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 686             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 687
 688             /* Update potential sum for this i atom from the interaction with this j atom. */
 689             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 690             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 691
 692             fscal            = fvdw;
 693
 694             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 695
 696             /* Update vectorial force */
 697             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 698             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 699             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 700
 701             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 702             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 703             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 704
 705             /**************************
 706              * CALCULATE INTERACTIONS *
 707              **************************/
 708
 709             /* REACTION-FIELD ELECTROSTATICS */
 710             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 711             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 712
 713             /* Update potential sum for this i atom from the interaction with this j atom. */
 714             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 715             velecsum         = _mm_add_pd(velecsum,velec);
 716
 717             fscal            = felec;
 718
 719             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 720
 721             /* Update vectorial force */
 722             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 723             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 724             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 725
 726             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 727             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 728             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 729
 730             /**************************
 731              * CALCULATE INTERACTIONS *
 732              **************************/
 733
 734             /* REACTION-FIELD ELECTROSTATICS */
 735             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 736             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 737
 738             /* Update potential sum for this i atom from the interaction with this j atom. */
 739             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 740             velecsum         = _mm_add_pd(velecsum,velec);
 741
 742             fscal            = felec;
 743
 744             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 745
 746             /* Update vectorial force */
 747             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 748             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 749             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 750
 751             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 752             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 753             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             /* REACTION-FIELD ELECTROSTATICS */
 760             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 761             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 762
 763             /* Update potential sum for this i atom from the interaction with this j atom. */
 764             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 765             velecsum         = _mm_add_pd(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 770
 771             /* Update vectorial force */
 772             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 773             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 774             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 775
 776             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 777             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 778             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 779
 780             /**************************
 781              * CALCULATE INTERACTIONS *
 782              **************************/
 783
 784             /* REACTION-FIELD ELECTROSTATICS */
 785             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 786             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 787
 788             /* Update potential sum for this i atom from the interaction with this j atom. */
 789             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 790             velecsum         = _mm_add_pd(velecsum,velec);
 791
 792             fscal            = felec;
 793
 794             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 795
 796             /* Update vectorial force */
 797             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 798             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 799             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 800
 801             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 802             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 803             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 804
 805             /**************************
 806              * CALCULATE INTERACTIONS *
 807              **************************/
 808
 809             /* REACTION-FIELD ELECTROSTATICS */
 810             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 811             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 815             velecsum         = _mm_add_pd(velecsum,velec);
 816
 817             fscal            = felec;
 818
 819             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 820
 821             /* Update vectorial force */
 822             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 823             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 824             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 825
 826             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 827             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 828             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             /* REACTION-FIELD ELECTROSTATICS */
 835             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 836             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 837
 838             /* Update potential sum for this i atom from the interaction with this j atom. */
 839             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 840             velecsum         = _mm_add_pd(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 845
 846             /* Update vectorial force */
 847             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 848             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 849             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 850
 851             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 852             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 853             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 854
 855             /**************************
 856              * CALCULATE INTERACTIONS *
 857              **************************/
 858
 859             /* REACTION-FIELD ELECTROSTATICS */
 860             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 861             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 862
 863             /* Update potential sum for this i atom from the interaction with this j atom. */
 864             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 865             velecsum         = _mm_add_pd(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 870
 871             /* Update vectorial force */
 872             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 873             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 874             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 875
 876             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 877             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 878             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             /* REACTION-FIELD ELECTROSTATICS */
 885             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 886             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 890             velecsum         = _mm_add_pd(velecsum,velec);
 891
 892             fscal            = felec;
 893
 894             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 895
 896             /* Update vectorial force */
 897             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 898             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 899             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 900
 901             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 902             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 903             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             /* REACTION-FIELD ELECTROSTATICS */
 910             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 911             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 912
 913             /* Update potential sum for this i atom from the interaction with this j atom. */
 914             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 915             velecsum         = _mm_add_pd(velecsum,velec);
 916
 917             fscal            = felec;
 918
 919             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 920
 921             /* Update vectorial force */
 922             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 923             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 924             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 925
 926             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 927             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 928             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 929
 930             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 931
 932             /* Inner loop uses 377 flops */
 933         }
 934
 935         /* End of innermost loop */
 936
 937         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 938                                               f+i_coord_offset,fshift+i_shift_offset);
 939
 940         ggid                        = gid[iidx];
 941         /* Update potential energies */
 942         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 943         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 944
 945         /* Increment number of inner iterations */
 946         inneriter                  += j_index_end - j_index_start;
 947
 948         /* Outer loop uses 26 flops */
 949     }
 950
 951     /* Increment number of outer iterations */
 952     outeriter        += nri;
 953
 954     /* Update outer/inner flops */
 955
 956     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*377);
 957 }
 958 /*
 959  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 960  * Electrostatics interaction: ReactionField
 961  * VdW interaction:            CubicSplineTable
 962  * Geometry:                   Water4-Water4
 963  * Calculate force/pot:        Force
 964  */
 965 void
 966 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 967                     (t_nblist * gmx_restrict                nlist,
 968                      rvec * gmx_restrict                    xx,
 969                      rvec * gmx_restrict                    ff,
 970                      t_forcerec * gmx_restrict              fr,
 971                      t_mdatoms * gmx_restrict               mdatoms,
 972                      nb_kernel_data_t * gmx_restrict        kernel_data,
 973                      t_nrnb * gmx_restrict                  nrnb)
 974 {
 975     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 976      * just 0 for non-waters.
 977      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 978      * jnr indices corresponding to data put in the four positions in the SIMD register.
 979      */
 980     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 981     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 982     int              jnrA,jnrB;
 983     int              j_coord_offsetA,j_coord_offsetB;
 984     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 985     real             rcutoff_scalar;
 986     real             *shiftvec,*fshift,*x,*f;
 987     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 988     int              vdwioffset0;
 989     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 990     int              vdwioffset1;
 991     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 992     int              vdwioffset2;
 993     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 994     int              vdwioffset3;
 995     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 996     int              vdwjidx0A,vdwjidx0B;
 997     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 998     int              vdwjidx1A,vdwjidx1B;
 999     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1000     int              vdwjidx2A,vdwjidx2B;
1001     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1002     int              vdwjidx3A,vdwjidx3B;
1003     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1004     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1005     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1006     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1007     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1008     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1009     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1010     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1011     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1012     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1013     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1014     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1015     real             *charge;
1016     int              nvdwtype;
1017     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1018     int              *vdwtype;
1019     real             *vdwparam;
1020     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1021     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1022     __m128i          vfitab;
1023     __m128i          ifour       = _mm_set1_epi32(4);
1024     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1025     real             *vftab;
1026     __m128d          dummy_mask,cutoff_mask;
1027     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1028     __m128d          one     = _mm_set1_pd(1.0);
1029     __m128d          two     = _mm_set1_pd(2.0);
1030     x                = xx[0];
1031     f                = ff[0];
1032
1033     nri              = nlist->nri;
1034     iinr             = nlist->iinr;
1035     jindex           = nlist->jindex;
1036     jjnr             = nlist->jjnr;
1037     shiftidx         = nlist->shift;
1038     gid              = nlist->gid;
1039     shiftvec         = fr->shift_vec[0];
1040     fshift           = fr->fshift[0];
1041     facel            = _mm_set1_pd(fr->epsfac);
1042     charge           = mdatoms->chargeA;
1043     krf              = _mm_set1_pd(fr->ic->k_rf);
1044     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1045     crf              = _mm_set1_pd(fr->ic->c_rf);
1046     nvdwtype         = fr->ntype;
1047     vdwparam         = fr->nbfp;
1048     vdwtype          = mdatoms->typeA;
1049
1050     vftab            = kernel_data->table_vdw->data;
1051     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1052
1053     /* Setup water-specific parameters */
1054     inr              = nlist->iinr[0];
1055     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1056     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1057     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1058     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1059
1060     jq1              = _mm_set1_pd(charge[inr+1]);
1061     jq2              = _mm_set1_pd(charge[inr+2]);
1062     jq3              = _mm_set1_pd(charge[inr+3]);
1063     vdwjidx0A        = 2*vdwtype[inr+0];
1064     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1065     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1066     qq11             = _mm_mul_pd(iq1,jq1);
1067     qq12             = _mm_mul_pd(iq1,jq2);
1068     qq13             = _mm_mul_pd(iq1,jq3);
1069     qq21             = _mm_mul_pd(iq2,jq1);
1070     qq22             = _mm_mul_pd(iq2,jq2);
1071     qq23             = _mm_mul_pd(iq2,jq3);
1072     qq31             = _mm_mul_pd(iq3,jq1);
1073     qq32             = _mm_mul_pd(iq3,jq2);
1074     qq33             = _mm_mul_pd(iq3,jq3);
1075
1076     /* Avoid stupid compiler warnings */
1077     jnrA = jnrB = 0;
1078     j_coord_offsetA = 0;
1079     j_coord_offsetB = 0;
1080
1081     outeriter        = 0;
1082     inneriter        = 0;
1083
1084     /* Start outer loop over neighborlists */
1085     for(iidx=0; iidx<nri; iidx++)
1086     {
1087         /* Load shift vector for this list */
1088         i_shift_offset   = DIM*shiftidx[iidx];
1089
1090         /* Load limits for loop over neighbors */
1091         j_index_start    = jindex[iidx];
1092         j_index_end      = jindex[iidx+1];
1093
1094         /* Get outer coordinate index */
1095         inr              = iinr[iidx];
1096         i_coord_offset   = DIM*inr;
1097
1098         /* Load i particle coords and add shift vector */
1099         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1100                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1101
1102         fix0             = _mm_setzero_pd();
1103         fiy0             = _mm_setzero_pd();
1104         fiz0             = _mm_setzero_pd();
1105         fix1             = _mm_setzero_pd();
1106         fiy1             = _mm_setzero_pd();
1107         fiz1             = _mm_setzero_pd();
1108         fix2             = _mm_setzero_pd();
1109         fiy2             = _mm_setzero_pd();
1110         fiz2             = _mm_setzero_pd();
1111         fix3             = _mm_setzero_pd();
1112         fiy3             = _mm_setzero_pd();
1113         fiz3             = _mm_setzero_pd();
1114
1115         /* Start inner kernel loop */
1116         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1117         {
1118
1119             /* Get j neighbor index, and coordinate index */
1120             jnrA             = jjnr[jidx];
1121             jnrB             = jjnr[jidx+1];
1122             j_coord_offsetA  = DIM*jnrA;
1123             j_coord_offsetB  = DIM*jnrB;
1124
1125             /* load j atom coordinates */
1126             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1127                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1128                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1129
1130             /* Calculate displacement vector */
1131             dx00             = _mm_sub_pd(ix0,jx0);
1132             dy00             = _mm_sub_pd(iy0,jy0);
1133             dz00             = _mm_sub_pd(iz0,jz0);
1134             dx11             = _mm_sub_pd(ix1,jx1);
1135             dy11             = _mm_sub_pd(iy1,jy1);
1136             dz11             = _mm_sub_pd(iz1,jz1);
1137             dx12             = _mm_sub_pd(ix1,jx2);
1138             dy12             = _mm_sub_pd(iy1,jy2);
1139             dz12             = _mm_sub_pd(iz1,jz2);
1140             dx13             = _mm_sub_pd(ix1,jx3);
1141             dy13             = _mm_sub_pd(iy1,jy3);
1142             dz13             = _mm_sub_pd(iz1,jz3);
1143             dx21             = _mm_sub_pd(ix2,jx1);
1144             dy21             = _mm_sub_pd(iy2,jy1);
1145             dz21             = _mm_sub_pd(iz2,jz1);
1146             dx22             = _mm_sub_pd(ix2,jx2);
1147             dy22             = _mm_sub_pd(iy2,jy2);
1148             dz22             = _mm_sub_pd(iz2,jz2);
1149             dx23             = _mm_sub_pd(ix2,jx3);
1150             dy23             = _mm_sub_pd(iy2,jy3);
1151             dz23             = _mm_sub_pd(iz2,jz3);
1152             dx31             = _mm_sub_pd(ix3,jx1);
1153             dy31             = _mm_sub_pd(iy3,jy1);
1154             dz31             = _mm_sub_pd(iz3,jz1);
1155             dx32             = _mm_sub_pd(ix3,jx2);
1156             dy32             = _mm_sub_pd(iy3,jy2);
1157             dz32             = _mm_sub_pd(iz3,jz2);
1158             dx33             = _mm_sub_pd(ix3,jx3);
1159             dy33             = _mm_sub_pd(iy3,jy3);
1160             dz33             = _mm_sub_pd(iz3,jz3);
1161
1162             /* Calculate squared distance and things based on it */
1163             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1164             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1165             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1166             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1167             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1168             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1169             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1170             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1171             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1172             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1173
1174             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1175             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1176             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1177             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1178             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1179             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1180             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1181             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1182             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1183             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1184
1185             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1186             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1187             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1188             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1189             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1190             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1191             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1192             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1193             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1194
1195             fjx0             = _mm_setzero_pd();
1196             fjy0             = _mm_setzero_pd();
1197             fjz0             = _mm_setzero_pd();
1198             fjx1             = _mm_setzero_pd();
1199             fjy1             = _mm_setzero_pd();
1200             fjz1             = _mm_setzero_pd();
1201             fjx2             = _mm_setzero_pd();
1202             fjy2             = _mm_setzero_pd();
1203             fjz2             = _mm_setzero_pd();
1204             fjx3             = _mm_setzero_pd();
1205             fjy3             = _mm_setzero_pd();
1206             fjz3             = _mm_setzero_pd();
1207
1208             /**************************
1209              * CALCULATE INTERACTIONS *
1210              **************************/
1211
1212             r00              = _mm_mul_pd(rsq00,rinv00);
1213
1214             /* Calculate table index by multiplying r with table scale and truncate to integer */
1215             rt               = _mm_mul_pd(r00,vftabscale);
1216             vfitab           = _mm_cvttpd_epi32(rt);
1217 #ifdef __XOP__
1218             vfeps            = _mm_frcz_pd(rt);
1219 #else
1220             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1221 #endif
1222             twovfeps         = _mm_add_pd(vfeps,vfeps);
1223             vfitab           = _mm_slli_epi32(vfitab,3);
1224
1225             /* CUBIC SPLINE TABLE DISPERSION */
1226             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1227             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1228             GMX_MM_TRANSPOSE2_PD(Y,F);
1229             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1230             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1231             GMX_MM_TRANSPOSE2_PD(G,H);
1232             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1233             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1234             fvdw6            = _mm_mul_pd(c6_00,FF);
1235
1236             /* CUBIC SPLINE TABLE REPULSION */
1237             vfitab           = _mm_add_epi32(vfitab,ifour);
1238             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1239             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1240             GMX_MM_TRANSPOSE2_PD(Y,F);
1241             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1242             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1243             GMX_MM_TRANSPOSE2_PD(G,H);
1244             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1245             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1246             fvdw12           = _mm_mul_pd(c12_00,FF);
1247             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1248
1249             fscal            = fvdw;
1250
1251             /* Update vectorial force */
1252             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1253             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1254             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1255
1256             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1257             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1258             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1259
1260             /**************************
1261              * CALCULATE INTERACTIONS *
1262              **************************/
1263
1264             /* REACTION-FIELD ELECTROSTATICS */
1265             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1266
1267             fscal            = felec;
1268
1269             /* Update vectorial force */
1270             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1271             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1272             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1273
1274             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1275             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1276             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1277
1278             /**************************
1279              * CALCULATE INTERACTIONS *
1280              **************************/
1281
1282             /* REACTION-FIELD ELECTROSTATICS */
1283             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1284
1285             fscal            = felec;
1286
1287             /* Update vectorial force */
1288             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1289             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1290             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1291
1292             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1293             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1294             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1295
1296             /**************************
1297              * CALCULATE INTERACTIONS *
1298              **************************/
1299
1300             /* REACTION-FIELD ELECTROSTATICS */
1301             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1302
1303             fscal            = felec;
1304
1305             /* Update vectorial force */
1306             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1307             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1308             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1309
1310             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1311             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1312             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1313
1314             /**************************
1315              * CALCULATE INTERACTIONS *
1316              **************************/
1317
1318             /* REACTION-FIELD ELECTROSTATICS */
1319             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1320
1321             fscal            = felec;
1322
1323             /* Update vectorial force */
1324             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1325             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1326             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1327
1328             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1329             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1330             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1331
1332             /**************************
1333              * CALCULATE INTERACTIONS *
1334              **************************/
1335
1336             /* REACTION-FIELD ELECTROSTATICS */
1337             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1338
1339             fscal            = felec;
1340
1341             /* Update vectorial force */
1342             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1343             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1344             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1345
1346             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1347             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1348             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1349
1350             /**************************
1351              * CALCULATE INTERACTIONS *
1352              **************************/
1353
1354             /* REACTION-FIELD ELECTROSTATICS */
1355             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1356
1357             fscal            = felec;
1358
1359             /* Update vectorial force */
1360             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1361             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1362             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1363
1364             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1365             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1366             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1367
1368             /**************************
1369              * CALCULATE INTERACTIONS *
1370              **************************/
1371
1372             /* REACTION-FIELD ELECTROSTATICS */
1373             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1374
1375             fscal            = felec;
1376
1377             /* Update vectorial force */
1378             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1379             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1380             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1381
1382             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1383             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1384             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1385
1386             /**************************
1387              * CALCULATE INTERACTIONS *
1388              **************************/
1389
1390             /* REACTION-FIELD ELECTROSTATICS */
1391             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1392
1393             fscal            = felec;
1394
1395             /* Update vectorial force */
1396             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1397             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1398             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1399
1400             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1401             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1402             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* REACTION-FIELD ELECTROSTATICS */
1409             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1410
1411             fscal            = felec;
1412
1413             /* Update vectorial force */
1414             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1415             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1416             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1417
1418             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1419             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1420             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1421
1422             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);
1423
1424             /* Inner loop uses 324 flops */
1425         }
1426
1427         if(jidx<j_index_end)
1428         {
1429
1430             jnrA             = jjnr[jidx];
1431             j_coord_offsetA  = DIM*jnrA;
1432
1433             /* load j atom coordinates */
1434             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1435                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1436                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1437
1438             /* Calculate displacement vector */
1439             dx00             = _mm_sub_pd(ix0,jx0);
1440             dy00             = _mm_sub_pd(iy0,jy0);
1441             dz00             = _mm_sub_pd(iz0,jz0);
1442             dx11             = _mm_sub_pd(ix1,jx1);
1443             dy11             = _mm_sub_pd(iy1,jy1);
1444             dz11             = _mm_sub_pd(iz1,jz1);
1445             dx12             = _mm_sub_pd(ix1,jx2);
1446             dy12             = _mm_sub_pd(iy1,jy2);
1447             dz12             = _mm_sub_pd(iz1,jz2);
1448             dx13             = _mm_sub_pd(ix1,jx3);
1449             dy13             = _mm_sub_pd(iy1,jy3);
1450             dz13             = _mm_sub_pd(iz1,jz3);
1451             dx21             = _mm_sub_pd(ix2,jx1);
1452             dy21             = _mm_sub_pd(iy2,jy1);
1453             dz21             = _mm_sub_pd(iz2,jz1);
1454             dx22             = _mm_sub_pd(ix2,jx2);
1455             dy22             = _mm_sub_pd(iy2,jy2);
1456             dz22             = _mm_sub_pd(iz2,jz2);
1457             dx23             = _mm_sub_pd(ix2,jx3);
1458             dy23             = _mm_sub_pd(iy2,jy3);
1459             dz23             = _mm_sub_pd(iz2,jz3);
1460             dx31             = _mm_sub_pd(ix3,jx1);
1461             dy31             = _mm_sub_pd(iy3,jy1);
1462             dz31             = _mm_sub_pd(iz3,jz1);
1463             dx32             = _mm_sub_pd(ix3,jx2);
1464             dy32             = _mm_sub_pd(iy3,jy2);
1465             dz32             = _mm_sub_pd(iz3,jz2);
1466             dx33             = _mm_sub_pd(ix3,jx3);
1467             dy33             = _mm_sub_pd(iy3,jy3);
1468             dz33             = _mm_sub_pd(iz3,jz3);
1469
1470             /* Calculate squared distance and things based on it */
1471             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1472             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1473             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1474             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1475             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1476             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1477             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1478             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1479             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1480             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1481
1482             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1483             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1484             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1485             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1486             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1487             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1488             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1489             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1490             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1491             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1492
1493             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1494             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1495             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1496             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1497             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1498             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1499             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1500             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1501             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1502
1503             fjx0             = _mm_setzero_pd();
1504             fjy0             = _mm_setzero_pd();
1505             fjz0             = _mm_setzero_pd();
1506             fjx1             = _mm_setzero_pd();
1507             fjy1             = _mm_setzero_pd();
1508             fjz1             = _mm_setzero_pd();
1509             fjx2             = _mm_setzero_pd();
1510             fjy2             = _mm_setzero_pd();
1511             fjz2             = _mm_setzero_pd();
1512             fjx3             = _mm_setzero_pd();
1513             fjy3             = _mm_setzero_pd();
1514             fjz3             = _mm_setzero_pd();
1515
1516             /**************************
1517              * CALCULATE INTERACTIONS *
1518              **************************/
1519
1520             r00              = _mm_mul_pd(rsq00,rinv00);
1521
1522             /* Calculate table index by multiplying r with table scale and truncate to integer */
1523             rt               = _mm_mul_pd(r00,vftabscale);
1524             vfitab           = _mm_cvttpd_epi32(rt);
1525 #ifdef __XOP__
1526             vfeps            = _mm_frcz_pd(rt);
1527 #else
1528             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1529 #endif
1530             twovfeps         = _mm_add_pd(vfeps,vfeps);
1531             vfitab           = _mm_slli_epi32(vfitab,3);
1532
1533             /* CUBIC SPLINE TABLE DISPERSION */
1534             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1535             F                = _mm_setzero_pd();
1536             GMX_MM_TRANSPOSE2_PD(Y,F);
1537             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1538             H                = _mm_setzero_pd();
1539             GMX_MM_TRANSPOSE2_PD(G,H);
1540             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1541             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1542             fvdw6            = _mm_mul_pd(c6_00,FF);
1543
1544             /* CUBIC SPLINE TABLE REPULSION */
1545             vfitab           = _mm_add_epi32(vfitab,ifour);
1546             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1547             F                = _mm_setzero_pd();
1548             GMX_MM_TRANSPOSE2_PD(Y,F);
1549             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1550             H                = _mm_setzero_pd();
1551             GMX_MM_TRANSPOSE2_PD(G,H);
1552             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1553             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1554             fvdw12           = _mm_mul_pd(c12_00,FF);
1555             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1556
1557             fscal            = fvdw;
1558
1559             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1560
1561             /* Update vectorial force */
1562             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1563             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1564             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1565
1566             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1567             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1568             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1569
1570             /**************************
1571              * CALCULATE INTERACTIONS *
1572              **************************/
1573
1574             /* REACTION-FIELD ELECTROSTATICS */
1575             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1576
1577             fscal            = felec;
1578
1579             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1580
1581             /* Update vectorial force */
1582             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1583             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1584             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1585
1586             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1587             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1588             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             /* REACTION-FIELD ELECTROSTATICS */
1595             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1600
1601             /* Update vectorial force */
1602             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1603             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1604             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1605
1606             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1607             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1608             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             /* REACTION-FIELD ELECTROSTATICS */
1615             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1616
1617             fscal            = felec;
1618
1619             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1620
1621             /* Update vectorial force */
1622             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1623             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1624             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1625
1626             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1627             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1628             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1629
1630             /**************************
1631              * CALCULATE INTERACTIONS *
1632              **************************/
1633
1634             /* REACTION-FIELD ELECTROSTATICS */
1635             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1636
1637             fscal            = felec;
1638
1639             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1640
1641             /* Update vectorial force */
1642             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1643             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1644             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1645
1646             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1647             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1648             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             /* REACTION-FIELD ELECTROSTATICS */
1655             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1656
1657             fscal            = felec;
1658
1659             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1660
1661             /* Update vectorial force */
1662             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1663             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1664             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1665
1666             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1667             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1668             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1669
1670             /**************************
1671              * CALCULATE INTERACTIONS *
1672              **************************/
1673
1674             /* REACTION-FIELD ELECTROSTATICS */
1675             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1676
1677             fscal            = felec;
1678
1679             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1680
1681             /* Update vectorial force */
1682             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1683             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1684             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1685
1686             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1687             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1688             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1689
1690             /**************************
1691              * CALCULATE INTERACTIONS *
1692              **************************/
1693
1694             /* REACTION-FIELD ELECTROSTATICS */
1695             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1696
1697             fscal            = felec;
1698
1699             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1700
1701             /* Update vectorial force */
1702             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1703             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1704             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1705
1706             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1707             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1708             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1709
1710             /**************************
1711              * CALCULATE INTERACTIONS *
1712              **************************/
1713
1714             /* REACTION-FIELD ELECTROSTATICS */
1715             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1716
1717             fscal            = felec;
1718
1719             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1720
1721             /* Update vectorial force */
1722             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1723             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1724             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1725
1726             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1727             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1728             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1729
1730             /**************************
1731              * CALCULATE INTERACTIONS *
1732              **************************/
1733
1734             /* REACTION-FIELD ELECTROSTATICS */
1735             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1736
1737             fscal            = felec;
1738
1739             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1740
1741             /* Update vectorial force */
1742             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1743             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1744             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1745
1746             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1747             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1748             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1749
1750             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1751
1752             /* Inner loop uses 324 flops */
1753         }
1754
1755         /* End of innermost loop */
1756
1757         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1758                                               f+i_coord_offset,fshift+i_shift_offset);
1759
1760         /* Increment number of inner iterations */
1761         inneriter                  += j_index_end - j_index_start;
1762
1763         /* Outer loop uses 24 flops */
1764     }
1765
1766     /* Increment number of outer iterations */
1767     outeriter        += nri;
1768
1769     /* Update outer/inner flops */
1770
1771     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*324);
1772 }