src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwLJ_GeomP1P1_avx_128_fma_single.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_single 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_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  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,C,D refer to j loop unrolling done with AVX_128, e.g. for the four 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,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     __m128i          gbitab;
  91     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  92     __m128           minushalf = _mm_set1_ps(-0.5);
  93     real             *invsqrta,*dvda,*gbtab;
  94     int              nvdwtype;
  95     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  99     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 103     real             *vftab;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(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_ps(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     invsqrta         = fr->invsqrta;
 126     dvda             = fr->dvda;
 127     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 128     gbtab            = fr->gbtab.data;
 129     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 130
 131     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = jnrC = jnrD = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135     j_coord_offsetC = 0;
 136     j_coord_offsetD = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm_setzero_ps();
 164         fiy0             = _mm_setzero_ps();
 165         fiz0             = _mm_setzero_ps();
 166
 167         /* Load parameters for i particles */
 168         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 169         isai0            = _mm_load1_ps(invsqrta+inr+0);
 170         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm_setzero_ps();
 174         vgbsum           = _mm_setzero_ps();
 175         vvdwsum          = _mm_setzero_ps();
 176         dvdasum          = _mm_setzero_ps();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189             j_coord_offsetC  = DIM*jnrC;
 190             j_coord_offsetD  = DIM*jnrD;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 194                                               x+j_coord_offsetC,x+j_coord_offsetD,
 195                                               &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm_sub_ps(ix0,jx0);
 199             dy00             = _mm_sub_ps(iy0,jy0);
 200             dz00             = _mm_sub_ps(iz0,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 204
 205             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 206
 207             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 211                                                               charge+jnrC+0,charge+jnrD+0);
 212             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 213                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 214             vdwjidx0A        = 2*vdwtype[jnrA+0];
 215             vdwjidx0B        = 2*vdwtype[jnrB+0];
 216             vdwjidx0C        = 2*vdwtype[jnrC+0];
 217             vdwjidx0D        = 2*vdwtype[jnrD+0];
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             r00              = _mm_mul_ps(rsq00,rinv00);
 224
 225             /* Compute parameters for interactions between i and j atoms */
 226             qq00             = _mm_mul_ps(iq0,jq0);
 227             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 228                                          vdwparam+vdwioffset0+vdwjidx0B,
 229                                          vdwparam+vdwioffset0+vdwjidx0C,
 230                                          vdwparam+vdwioffset0+vdwjidx0D,
 231                                          &c6_00,&c12_00);
 232
 233             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 234             isaprod          = _mm_mul_ps(isai0,isaj0);
 235             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 236             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 237
 238             /* Calculate generalized born table index - this is a separate table from the normal one,
 239              * but we use the same procedure by multiplying r with scale and truncating to integer.
 240              */
 241             rt               = _mm_mul_ps(r00,gbscale);
 242             gbitab           = _mm_cvttps_epi32(rt);
 243 #ifdef __XOP__
 244             gbeps            = _mm_frcz_ps(rt);
 245 #else
 246             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 247 #endif
 248             gbitab           = _mm_slli_epi32(gbitab,2);
 249
 250             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 251             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 252             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 253             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 254             _MM_TRANSPOSE4_PS(Y,F,G,H);
 255             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 256             VV               = _mm_macc_ps(gbeps,Fp,Y);
 257             vgb              = _mm_mul_ps(gbqqfactor,VV);
 258
 259             twogbeps         = _mm_add_ps(gbeps,gbeps);
 260             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 261             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 262             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 263             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 264             fjptrA           = dvda+jnrA;
 265             fjptrB           = dvda+jnrB;
 266             fjptrC           = dvda+jnrC;
 267             fjptrD           = dvda+jnrD;
 268             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 269             velec            = _mm_mul_ps(qq00,rinv00);
 270             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 271
 272             /* LENNARD-JONES DISPERSION/REPULSION */
 273
 274             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 275             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 276             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 277             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 278             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 279
 280             /* Update potential sum for this i atom from the interaction with this j atom. */
 281             velecsum         = _mm_add_ps(velecsum,velec);
 282             vgbsum           = _mm_add_ps(vgbsum,vgb);
 283             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 284
 285             fscal            = _mm_add_ps(felec,fvdw);
 286
 287              /* Update vectorial force */
 288             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 289             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 290             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 291
 292             fjptrA             = f+j_coord_offsetA;
 293             fjptrB             = f+j_coord_offsetB;
 294             fjptrC             = f+j_coord_offsetC;
 295             fjptrD             = f+j_coord_offsetD;
 296             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 297                                                    _mm_mul_ps(dx00,fscal),
 298                                                    _mm_mul_ps(dy00,fscal),
 299                                                    _mm_mul_ps(dz00,fscal));
 300
 301             /* Inner loop uses 74 flops */
 302         }
 303
 304         if(jidx<j_index_end)
 305         {
 306
 307             /* Get j neighbor index, and coordinate index */
 308             jnrlistA         = jjnr[jidx];
 309             jnrlistB         = jjnr[jidx+1];
 310             jnrlistC         = jjnr[jidx+2];
 311             jnrlistD         = jjnr[jidx+3];
 312             /* Sign of each element will be negative for non-real atoms.
 313              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 314              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 315              */
 316             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 317             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 318             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 319             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 320             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 321             j_coord_offsetA  = DIM*jnrA;
 322             j_coord_offsetB  = DIM*jnrB;
 323             j_coord_offsetC  = DIM*jnrC;
 324             j_coord_offsetD  = DIM*jnrD;
 325
 326             /* load j atom coordinates */
 327             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 328                                               x+j_coord_offsetC,x+j_coord_offsetD,
 329                                               &jx0,&jy0,&jz0);
 330
 331             /* Calculate displacement vector */
 332             dx00             = _mm_sub_ps(ix0,jx0);
 333             dy00             = _mm_sub_ps(iy0,jy0);
 334             dz00             = _mm_sub_ps(iz0,jz0);
 335
 336             /* Calculate squared distance and things based on it */
 337             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 338
 339             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 340
 341             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 342
 343             /* Load parameters for j particles */
 344             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 345                                                               charge+jnrC+0,charge+jnrD+0);
 346             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 347                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 348             vdwjidx0A        = 2*vdwtype[jnrA+0];
 349             vdwjidx0B        = 2*vdwtype[jnrB+0];
 350             vdwjidx0C        = 2*vdwtype[jnrC+0];
 351             vdwjidx0D        = 2*vdwtype[jnrD+0];
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             r00              = _mm_mul_ps(rsq00,rinv00);
 358             r00              = _mm_andnot_ps(dummy_mask,r00);
 359
 360             /* Compute parameters for interactions between i and j atoms */
 361             qq00             = _mm_mul_ps(iq0,jq0);
 362             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 363                                          vdwparam+vdwioffset0+vdwjidx0B,
 364                                          vdwparam+vdwioffset0+vdwjidx0C,
 365                                          vdwparam+vdwioffset0+vdwjidx0D,
 366                                          &c6_00,&c12_00);
 367
 368             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 369             isaprod          = _mm_mul_ps(isai0,isaj0);
 370             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 371             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 372
 373             /* Calculate generalized born table index - this is a separate table from the normal one,
 374              * but we use the same procedure by multiplying r with scale and truncating to integer.
 375              */
 376             rt               = _mm_mul_ps(r00,gbscale);
 377             gbitab           = _mm_cvttps_epi32(rt);
 378 #ifdef __XOP__
 379             gbeps            = _mm_frcz_ps(rt);
 380 #else
 381             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 382 #endif
 383             gbitab           = _mm_slli_epi32(gbitab,2);
 384
 385             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 386             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 387             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 388             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 389             _MM_TRANSPOSE4_PS(Y,F,G,H);
 390             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 391             VV               = _mm_macc_ps(gbeps,Fp,Y);
 392             vgb              = _mm_mul_ps(gbqqfactor,VV);
 393
 394             twogbeps         = _mm_add_ps(gbeps,gbeps);
 395             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 396             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 397             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 398             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 399             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 400             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 401             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 402             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 403             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 404             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 405             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 406             velec            = _mm_mul_ps(qq00,rinv00);
 407             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 408
 409             /* LENNARD-JONES DISPERSION/REPULSION */
 410
 411             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 412             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 413             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 414             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 415             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velec            = _mm_andnot_ps(dummy_mask,velec);
 419             velecsum         = _mm_add_ps(velecsum,velec);
 420             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 421             vgbsum           = _mm_add_ps(vgbsum,vgb);
 422             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 423             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 424
 425             fscal            = _mm_add_ps(felec,fvdw);
 426
 427             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 428
 429              /* Update vectorial force */
 430             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 431             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 432             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 433
 434             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 435             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 436             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 437             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 438             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 439                                                    _mm_mul_ps(dx00,fscal),
 440                                                    _mm_mul_ps(dy00,fscal),
 441                                                    _mm_mul_ps(dz00,fscal));
 442
 443             /* Inner loop uses 75 flops */
 444         }
 445
 446         /* End of innermost loop */
 447
 448         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 449                                               f+i_coord_offset,fshift+i_shift_offset);
 450
 451         ggid                        = gid[iidx];
 452         /* Update potential energies */
 453         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 454         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 455         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 456         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 457         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 458
 459         /* Increment number of inner iterations */
 460         inneriter                  += j_index_end - j_index_start;
 461
 462         /* Outer loop uses 10 flops */
 463     }
 464
 465     /* Increment number of outer iterations */
 466     outeriter        += nri;
 467
 468     /* Update outer/inner flops */
 469
 470     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*75);
 471 }
 472 /*
 473  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_single
 474  * Electrostatics interaction: GeneralizedBorn
 475  * VdW interaction:            LennardJones
 476  * Geometry:                   Particle-Particle
 477  * Calculate force/pot:        Force
 478  */
 479 void
 480 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_single
 481                     (t_nblist                    * gmx_restrict       nlist,
 482                      rvec                        * gmx_restrict          xx,
 483                      rvec                        * gmx_restrict          ff,
 484                      t_forcerec                  * gmx_restrict          fr,
 485                      t_mdatoms                   * gmx_restrict     mdatoms,
 486                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 487                      t_nrnb                      * gmx_restrict        nrnb)
 488 {
 489     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 490      * just 0 for non-waters.
 491      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 492      * jnr indices corresponding to data put in the four positions in the SIMD register.
 493      */
 494     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 495     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 496     int              jnrA,jnrB,jnrC,jnrD;
 497     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 498     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 499     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 500     real             rcutoff_scalar;
 501     real             *shiftvec,*fshift,*x,*f;
 502     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 503     real             scratch[4*DIM];
 504     __m128           fscal,rcutoff,rcutoff2,jidxall;
 505     int              vdwioffset0;
 506     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 507     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 508     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 509     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 510     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 511     real             *charge;
 512     __m128i          gbitab;
 513     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 514     __m128           minushalf = _mm_set1_ps(-0.5);
 515     real             *invsqrta,*dvda,*gbtab;
 516     int              nvdwtype;
 517     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 518     int              *vdwtype;
 519     real             *vdwparam;
 520     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 521     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 522     __m128i          vfitab;
 523     __m128i          ifour       = _mm_set1_epi32(4);
 524     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 525     real             *vftab;
 526     __m128           dummy_mask,cutoff_mask;
 527     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 528     __m128           one     = _mm_set1_ps(1.0);
 529     __m128           two     = _mm_set1_ps(2.0);
 530     x                = xx[0];
 531     f                = ff[0];
 532
 533     nri              = nlist->nri;
 534     iinr             = nlist->iinr;
 535     jindex           = nlist->jindex;
 536     jjnr             = nlist->jjnr;
 537     shiftidx         = nlist->shift;
 538     gid              = nlist->gid;
 539     shiftvec         = fr->shift_vec[0];
 540     fshift           = fr->fshift[0];
 541     facel            = _mm_set1_ps(fr->epsfac);
 542     charge           = mdatoms->chargeA;
 543     nvdwtype         = fr->ntype;
 544     vdwparam         = fr->nbfp;
 545     vdwtype          = mdatoms->typeA;
 546
 547     invsqrta         = fr->invsqrta;
 548     dvda             = fr->dvda;
 549     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 550     gbtab            = fr->gbtab.data;
 551     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 552
 553     /* Avoid stupid compiler warnings */
 554     jnrA = jnrB = jnrC = jnrD = 0;
 555     j_coord_offsetA = 0;
 556     j_coord_offsetB = 0;
 557     j_coord_offsetC = 0;
 558     j_coord_offsetD = 0;
 559
 560     outeriter        = 0;
 561     inneriter        = 0;
 562
 563     for(iidx=0;iidx<4*DIM;iidx++)
 564     {
 565         scratch[iidx] = 0.0;
 566     }
 567
 568     /* Start outer loop over neighborlists */
 569     for(iidx=0; iidx<nri; iidx++)
 570     {
 571         /* Load shift vector for this list */
 572         i_shift_offset   = DIM*shiftidx[iidx];
 573
 574         /* Load limits for loop over neighbors */
 575         j_index_start    = jindex[iidx];
 576         j_index_end      = jindex[iidx+1];
 577
 578         /* Get outer coordinate index */
 579         inr              = iinr[iidx];
 580         i_coord_offset   = DIM*inr;
 581
 582         /* Load i particle coords and add shift vector */
 583         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 584
 585         fix0             = _mm_setzero_ps();
 586         fiy0             = _mm_setzero_ps();
 587         fiz0             = _mm_setzero_ps();
 588
 589         /* Load parameters for i particles */
 590         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 591         isai0            = _mm_load1_ps(invsqrta+inr+0);
 592         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 593
 594         dvdasum          = _mm_setzero_ps();
 595
 596         /* Start inner kernel loop */
 597         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 598         {
 599
 600             /* Get j neighbor index, and coordinate index */
 601             jnrA             = jjnr[jidx];
 602             jnrB             = jjnr[jidx+1];
 603             jnrC             = jjnr[jidx+2];
 604             jnrD             = jjnr[jidx+3];
 605             j_coord_offsetA  = DIM*jnrA;
 606             j_coord_offsetB  = DIM*jnrB;
 607             j_coord_offsetC  = DIM*jnrC;
 608             j_coord_offsetD  = DIM*jnrD;
 609
 610             /* load j atom coordinates */
 611             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 612                                               x+j_coord_offsetC,x+j_coord_offsetD,
 613                                               &jx0,&jy0,&jz0);
 614
 615             /* Calculate displacement vector */
 616             dx00             = _mm_sub_ps(ix0,jx0);
 617             dy00             = _mm_sub_ps(iy0,jy0);
 618             dz00             = _mm_sub_ps(iz0,jz0);
 619
 620             /* Calculate squared distance and things based on it */
 621             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 622
 623             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 624
 625             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 626
 627             /* Load parameters for j particles */
 628             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 629                                                               charge+jnrC+0,charge+jnrD+0);
 630             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 631                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 632             vdwjidx0A        = 2*vdwtype[jnrA+0];
 633             vdwjidx0B        = 2*vdwtype[jnrB+0];
 634             vdwjidx0C        = 2*vdwtype[jnrC+0];
 635             vdwjidx0D        = 2*vdwtype[jnrD+0];
 636
 637             /**************************
 638              * CALCULATE INTERACTIONS *
 639              **************************/
 640
 641             r00              = _mm_mul_ps(rsq00,rinv00);
 642
 643             /* Compute parameters for interactions between i and j atoms */
 644             qq00             = _mm_mul_ps(iq0,jq0);
 645             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 646                                          vdwparam+vdwioffset0+vdwjidx0B,
 647                                          vdwparam+vdwioffset0+vdwjidx0C,
 648                                          vdwparam+vdwioffset0+vdwjidx0D,
 649                                          &c6_00,&c12_00);
 650
 651             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 652             isaprod          = _mm_mul_ps(isai0,isaj0);
 653             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 654             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 655
 656             /* Calculate generalized born table index - this is a separate table from the normal one,
 657              * but we use the same procedure by multiplying r with scale and truncating to integer.
 658              */
 659             rt               = _mm_mul_ps(r00,gbscale);
 660             gbitab           = _mm_cvttps_epi32(rt);
 661 #ifdef __XOP__
 662             gbeps            = _mm_frcz_ps(rt);
 663 #else
 664             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 665 #endif
 666             gbitab           = _mm_slli_epi32(gbitab,2);
 667
 668             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 669             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 670             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 671             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 672             _MM_TRANSPOSE4_PS(Y,F,G,H);
 673             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 674             VV               = _mm_macc_ps(gbeps,Fp,Y);
 675             vgb              = _mm_mul_ps(gbqqfactor,VV);
 676
 677             twogbeps         = _mm_add_ps(gbeps,gbeps);
 678             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 679             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 680             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 681             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 682             fjptrA           = dvda+jnrA;
 683             fjptrB           = dvda+jnrB;
 684             fjptrC           = dvda+jnrC;
 685             fjptrD           = dvda+jnrD;
 686             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 687             velec            = _mm_mul_ps(qq00,rinv00);
 688             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 689
 690             /* LENNARD-JONES DISPERSION/REPULSION */
 691
 692             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 693             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 694
 695             fscal            = _mm_add_ps(felec,fvdw);
 696
 697              /* Update vectorial force */
 698             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 699             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 700             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 701
 702             fjptrA             = f+j_coord_offsetA;
 703             fjptrB             = f+j_coord_offsetB;
 704             fjptrC             = f+j_coord_offsetC;
 705             fjptrD             = f+j_coord_offsetD;
 706             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 707                                                    _mm_mul_ps(dx00,fscal),
 708                                                    _mm_mul_ps(dy00,fscal),
 709                                                    _mm_mul_ps(dz00,fscal));
 710
 711             /* Inner loop uses 67 flops */
 712         }
 713
 714         if(jidx<j_index_end)
 715         {
 716
 717             /* Get j neighbor index, and coordinate index */
 718             jnrlistA         = jjnr[jidx];
 719             jnrlistB         = jjnr[jidx+1];
 720             jnrlistC         = jjnr[jidx+2];
 721             jnrlistD         = jjnr[jidx+3];
 722             /* Sign of each element will be negative for non-real atoms.
 723              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 724              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 725              */
 726             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 727             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 728             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 729             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 730             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 731             j_coord_offsetA  = DIM*jnrA;
 732             j_coord_offsetB  = DIM*jnrB;
 733             j_coord_offsetC  = DIM*jnrC;
 734             j_coord_offsetD  = DIM*jnrD;
 735
 736             /* load j atom coordinates */
 737             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 738                                               x+j_coord_offsetC,x+j_coord_offsetD,
 739                                               &jx0,&jy0,&jz0);
 740
 741             /* Calculate displacement vector */
 742             dx00             = _mm_sub_ps(ix0,jx0);
 743             dy00             = _mm_sub_ps(iy0,jy0);
 744             dz00             = _mm_sub_ps(iz0,jz0);
 745
 746             /* Calculate squared distance and things based on it */
 747             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 748
 749             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 750
 751             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 752
 753             /* Load parameters for j particles */
 754             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 755                                                               charge+jnrC+0,charge+jnrD+0);
 756             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 757                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 758             vdwjidx0A        = 2*vdwtype[jnrA+0];
 759             vdwjidx0B        = 2*vdwtype[jnrB+0];
 760             vdwjidx0C        = 2*vdwtype[jnrC+0];
 761             vdwjidx0D        = 2*vdwtype[jnrD+0];
 762
 763             /**************************
 764              * CALCULATE INTERACTIONS *
 765              **************************/
 766
 767             r00              = _mm_mul_ps(rsq00,rinv00);
 768             r00              = _mm_andnot_ps(dummy_mask,r00);
 769
 770             /* Compute parameters for interactions between i and j atoms */
 771             qq00             = _mm_mul_ps(iq0,jq0);
 772             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 773                                          vdwparam+vdwioffset0+vdwjidx0B,
 774                                          vdwparam+vdwioffset0+vdwjidx0C,
 775                                          vdwparam+vdwioffset0+vdwjidx0D,
 776                                          &c6_00,&c12_00);
 777
 778             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 779             isaprod          = _mm_mul_ps(isai0,isaj0);
 780             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 781             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 782
 783             /* Calculate generalized born table index - this is a separate table from the normal one,
 784              * but we use the same procedure by multiplying r with scale and truncating to integer.
 785              */
 786             rt               = _mm_mul_ps(r00,gbscale);
 787             gbitab           = _mm_cvttps_epi32(rt);
 788 #ifdef __XOP__
 789             gbeps            = _mm_frcz_ps(rt);
 790 #else
 791             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 792 #endif
 793             gbitab           = _mm_slli_epi32(gbitab,2);
 794
 795             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 796             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 797             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 798             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 799             _MM_TRANSPOSE4_PS(Y,F,G,H);
 800             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 801             VV               = _mm_macc_ps(gbeps,Fp,Y);
 802             vgb              = _mm_mul_ps(gbqqfactor,VV);
 803
 804             twogbeps         = _mm_add_ps(gbeps,gbeps);
 805             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 806             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 807             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 808             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 809             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 810             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 811             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 812             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 813             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 814             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 815             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 816             velec            = _mm_mul_ps(qq00,rinv00);
 817             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 818
 819             /* LENNARD-JONES DISPERSION/REPULSION */
 820
 821             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 822             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 823
 824             fscal            = _mm_add_ps(felec,fvdw);
 825
 826             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 827
 828              /* Update vectorial force */
 829             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 830             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 831             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 832
 833             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 834             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 835             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 836             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 837             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 838                                                    _mm_mul_ps(dx00,fscal),
 839                                                    _mm_mul_ps(dy00,fscal),
 840                                                    _mm_mul_ps(dz00,fscal));
 841
 842             /* Inner loop uses 68 flops */
 843         }
 844
 845         /* End of innermost loop */
 846
 847         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 848                                               f+i_coord_offset,fshift+i_shift_offset);
 849
 850         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 851         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 852
 853         /* Increment number of inner iterations */
 854         inneriter                  += j_index_end - j_index_start;
 855
 856         /* Outer loop uses 7 flops */
 857     }
 858
 859     /* Increment number of outer iterations */
 860     outeriter        += nri;
 861
 862     /* Update outer/inner flops */
 863
 864     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*68);
 865 }