src/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_single 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_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse4_1_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse4_1_single
  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,C,D refer to j loop unrolling done with SSE, e.g. for the four 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,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     __m128i          gbitab;
  77     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  78     __m128           minushalf = _mm_set1_ps(-0.5);
  79     real             *invsqrta,*dvda,*gbtab;
  80     int              nvdwtype;
  81     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  82     int              *vdwtype;
  83     real             *vdwparam;
  84     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  85     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  86     __m128i          vfitab;
  87     __m128i          ifour       = _mm_set1_epi32(4);
  88     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  89     real             *vftab;
  90     __m128           dummy_mask,cutoff_mask;
  91     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  92     __m128           one     = _mm_set1_ps(1.0);
  93     __m128           two     = _mm_set1_ps(2.0);
  94     x                = xx[0];
  95     f                = ff[0];
  96
  97     nri              = nlist->nri;
  98     iinr             = nlist->iinr;
  99     jindex           = nlist->jindex;
 100     jjnr             = nlist->jjnr;
 101     shiftidx         = nlist->shift;
 102     gid              = nlist->gid;
 103     shiftvec         = fr->shift_vec[0];
 104     fshift           = fr->fshift[0];
 105     facel            = _mm_set1_ps(fr->epsfac);
 106     charge           = mdatoms->chargeA;
 107     nvdwtype         = fr->ntype;
 108     vdwparam         = fr->nbfp;
 109     vdwtype          = mdatoms->typeA;
 110
 111     vftab            = kernel_data->table_vdw->data;
 112     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 113
 114     invsqrta         = fr->invsqrta;
 115     dvda             = fr->dvda;
 116     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 117     gbtab            = fr->gbtab.data;
 118     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 119
 120     /* Avoid stupid compiler warnings */
 121     jnrA = jnrB = jnrC = jnrD = 0;
 122     j_coord_offsetA = 0;
 123     j_coord_offsetB = 0;
 124     j_coord_offsetC = 0;
 125     j_coord_offsetD = 0;
 126
 127     outeriter        = 0;
 128     inneriter        = 0;
 129
 130     for(iidx=0;iidx<4*DIM;iidx++)
 131     {
 132         scratch[iidx] = 0.0;
 133     }
 134
 135     /* Start outer loop over neighborlists */
 136     for(iidx=0; iidx<nri; iidx++)
 137     {
 138         /* Load shift vector for this list */
 139         i_shift_offset   = DIM*shiftidx[iidx];
 140
 141         /* Load limits for loop over neighbors */
 142         j_index_start    = jindex[iidx];
 143         j_index_end      = jindex[iidx+1];
 144
 145         /* Get outer coordinate index */
 146         inr              = iinr[iidx];
 147         i_coord_offset   = DIM*inr;
 148
 149         /* Load i particle coords and add shift vector */
 150         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 151
 152         fix0             = _mm_setzero_ps();
 153         fiy0             = _mm_setzero_ps();
 154         fiz0             = _mm_setzero_ps();
 155
 156         /* Load parameters for i particles */
 157         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 158         isai0            = _mm_load1_ps(invsqrta+inr+0);
 159         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm_setzero_ps();
 163         vgbsum           = _mm_setzero_ps();
 164         vvdwsum          = _mm_setzero_ps();
 165         dvdasum          = _mm_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                               x+j_coord_offsetC,x+j_coord_offsetD,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_ps(ix0,jx0);
 188             dy00             = _mm_sub_ps(iy0,jy0);
 189             dz00             = _mm_sub_ps(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 193
 194             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 195
 196             /* Load parameters for j particles */
 197             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 198                                                               charge+jnrC+0,charge+jnrD+0);
 199             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 200                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 201             vdwjidx0A        = 2*vdwtype[jnrA+0];
 202             vdwjidx0B        = 2*vdwtype[jnrB+0];
 203             vdwjidx0C        = 2*vdwtype[jnrC+0];
 204             vdwjidx0D        = 2*vdwtype[jnrD+0];
 205
 206             /**************************
 207              * CALCULATE INTERACTIONS *
 208              **************************/
 209
 210             r00              = _mm_mul_ps(rsq00,rinv00);
 211
 212             /* Compute parameters for interactions between i and j atoms */
 213             qq00             = _mm_mul_ps(iq0,jq0);
 214             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 215                                          vdwparam+vdwioffset0+vdwjidx0B,
 216                                          vdwparam+vdwioffset0+vdwjidx0C,
 217                                          vdwparam+vdwioffset0+vdwjidx0D,
 218                                          &c6_00,&c12_00);
 219
 220             /* Calculate table index by multiplying r with table scale and truncate to integer */
 221             rt               = _mm_mul_ps(r00,vftabscale);
 222             vfitab           = _mm_cvttps_epi32(rt);
 223             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 224             vfitab           = _mm_slli_epi32(vfitab,3);
 225
 226             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 227             isaprod          = _mm_mul_ps(isai0,isaj0);
 228             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 229             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 230
 231             /* Calculate generalized born table index - this is a separate table from the normal one,
 232              * but we use the same procedure by multiplying r with scale and truncating to integer.
 233              */
 234             rt               = _mm_mul_ps(r00,gbscale);
 235             gbitab           = _mm_cvttps_epi32(rt);
 236             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 237             gbitab           = _mm_slli_epi32(gbitab,2);
 238             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 239             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 240             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 241             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 242             _MM_TRANSPOSE4_PS(Y,F,G,H);
 243             Heps             = _mm_mul_ps(gbeps,H);
 244             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 245             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 246             vgb              = _mm_mul_ps(gbqqfactor,VV);
 247
 248             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 249             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 250             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 251             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 252             fjptrA           = dvda+jnrA;
 253             fjptrB           = dvda+jnrB;
 254             fjptrC           = dvda+jnrC;
 255             fjptrD           = dvda+jnrD;
 256             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 257             velec            = _mm_mul_ps(qq00,rinv00);
 258             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 259
 260             /* CUBIC SPLINE TABLE DISPERSION */
 261             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 262             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 263             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 264             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 265             _MM_TRANSPOSE4_PS(Y,F,G,H);
 266             Heps             = _mm_mul_ps(vfeps,H);
 267             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 268             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 269             vvdw6            = _mm_mul_ps(c6_00,VV);
 270             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 271             fvdw6            = _mm_mul_ps(c6_00,FF);
 272
 273             /* CUBIC SPLINE TABLE REPULSION */
 274             vfitab           = _mm_add_epi32(vfitab,ifour);
 275             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 276             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 277             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 278             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 279             _MM_TRANSPOSE4_PS(Y,F,G,H);
 280             Heps             = _mm_mul_ps(vfeps,H);
 281             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 282             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 283             vvdw12           = _mm_mul_ps(c12_00,VV);
 284             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 285             fvdw12           = _mm_mul_ps(c12_00,FF);
 286             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 287             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 288
 289             /* Update potential sum for this i atom from the interaction with this j atom. */
 290             velecsum         = _mm_add_ps(velecsum,velec);
 291             vgbsum           = _mm_add_ps(vgbsum,vgb);
 292             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 293
 294             fscal            = _mm_add_ps(felec,fvdw);
 295
 296             /* Calculate temporary vectorial force */
 297             tx               = _mm_mul_ps(fscal,dx00);
 298             ty               = _mm_mul_ps(fscal,dy00);
 299             tz               = _mm_mul_ps(fscal,dz00);
 300
 301             /* Update vectorial force */
 302             fix0             = _mm_add_ps(fix0,tx);
 303             fiy0             = _mm_add_ps(fiy0,ty);
 304             fiz0             = _mm_add_ps(fiz0,tz);
 305
 306             fjptrA             = f+j_coord_offsetA;
 307             fjptrB             = f+j_coord_offsetB;
 308             fjptrC             = f+j_coord_offsetC;
 309             fjptrD             = f+j_coord_offsetD;
 310             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 311
 312             /* Inner loop uses 92 flops */
 313         }
 314
 315         if(jidx<j_index_end)
 316         {
 317
 318             /* Get j neighbor index, and coordinate index */
 319             jnrlistA         = jjnr[jidx];
 320             jnrlistB         = jjnr[jidx+1];
 321             jnrlistC         = jjnr[jidx+2];
 322             jnrlistD         = jjnr[jidx+3];
 323             /* Sign of each element will be negative for non-real atoms.
 324              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 325              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 326              */
 327             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 328             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 329             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 330             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 331             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 332             j_coord_offsetA  = DIM*jnrA;
 333             j_coord_offsetB  = DIM*jnrB;
 334             j_coord_offsetC  = DIM*jnrC;
 335             j_coord_offsetD  = DIM*jnrD;
 336
 337             /* load j atom coordinates */
 338             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 339                                               x+j_coord_offsetC,x+j_coord_offsetD,
 340                                               &jx0,&jy0,&jz0);
 341
 342             /* Calculate displacement vector */
 343             dx00             = _mm_sub_ps(ix0,jx0);
 344             dy00             = _mm_sub_ps(iy0,jy0);
 345             dz00             = _mm_sub_ps(iz0,jz0);
 346
 347             /* Calculate squared distance and things based on it */
 348             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 349
 350             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 351
 352             /* Load parameters for j particles */
 353             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 354                                                               charge+jnrC+0,charge+jnrD+0);
 355             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 356                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 357             vdwjidx0A        = 2*vdwtype[jnrA+0];
 358             vdwjidx0B        = 2*vdwtype[jnrB+0];
 359             vdwjidx0C        = 2*vdwtype[jnrC+0];
 360             vdwjidx0D        = 2*vdwtype[jnrD+0];
 361
 362             /**************************
 363              * CALCULATE INTERACTIONS *
 364              **************************/
 365
 366             r00              = _mm_mul_ps(rsq00,rinv00);
 367             r00              = _mm_andnot_ps(dummy_mask,r00);
 368
 369             /* Compute parameters for interactions between i and j atoms */
 370             qq00             = _mm_mul_ps(iq0,jq0);
 371             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 372                                          vdwparam+vdwioffset0+vdwjidx0B,
 373                                          vdwparam+vdwioffset0+vdwjidx0C,
 374                                          vdwparam+vdwioffset0+vdwjidx0D,
 375                                          &c6_00,&c12_00);
 376
 377             /* Calculate table index by multiplying r with table scale and truncate to integer */
 378             rt               = _mm_mul_ps(r00,vftabscale);
 379             vfitab           = _mm_cvttps_epi32(rt);
 380             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 381             vfitab           = _mm_slli_epi32(vfitab,3);
 382
 383             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 384             isaprod          = _mm_mul_ps(isai0,isaj0);
 385             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 386             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 387
 388             /* Calculate generalized born table index - this is a separate table from the normal one,
 389              * but we use the same procedure by multiplying r with scale and truncating to integer.
 390              */
 391             rt               = _mm_mul_ps(r00,gbscale);
 392             gbitab           = _mm_cvttps_epi32(rt);
 393             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 394             gbitab           = _mm_slli_epi32(gbitab,2);
 395             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 396             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 397             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 398             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 399             _MM_TRANSPOSE4_PS(Y,F,G,H);
 400             Heps             = _mm_mul_ps(gbeps,H);
 401             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 402             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 403             vgb              = _mm_mul_ps(gbqqfactor,VV);
 404
 405             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 406             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 407             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 408             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 409             /* 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. */
 410             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 411             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 412             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 413             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 414             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 415             velec            = _mm_mul_ps(qq00,rinv00);
 416             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 417
 418             /* CUBIC SPLINE TABLE DISPERSION */
 419             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 420             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 421             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 422             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 423             _MM_TRANSPOSE4_PS(Y,F,G,H);
 424             Heps             = _mm_mul_ps(vfeps,H);
 425             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 426             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 427             vvdw6            = _mm_mul_ps(c6_00,VV);
 428             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 429             fvdw6            = _mm_mul_ps(c6_00,FF);
 430
 431             /* CUBIC SPLINE TABLE REPULSION */
 432             vfitab           = _mm_add_epi32(vfitab,ifour);
 433             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 434             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 435             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 436             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 437             _MM_TRANSPOSE4_PS(Y,F,G,H);
 438             Heps             = _mm_mul_ps(vfeps,H);
 439             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 440             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 441             vvdw12           = _mm_mul_ps(c12_00,VV);
 442             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 443             fvdw12           = _mm_mul_ps(c12_00,FF);
 444             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 445             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velec            = _mm_andnot_ps(dummy_mask,velec);
 449             velecsum         = _mm_add_ps(velecsum,velec);
 450             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 451             vgbsum           = _mm_add_ps(vgbsum,vgb);
 452             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 453             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 454
 455             fscal            = _mm_add_ps(felec,fvdw);
 456
 457             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 458
 459             /* Calculate temporary vectorial force */
 460             tx               = _mm_mul_ps(fscal,dx00);
 461             ty               = _mm_mul_ps(fscal,dy00);
 462             tz               = _mm_mul_ps(fscal,dz00);
 463
 464             /* Update vectorial force */
 465             fix0             = _mm_add_ps(fix0,tx);
 466             fiy0             = _mm_add_ps(fiy0,ty);
 467             fiz0             = _mm_add_ps(fiz0,tz);
 468
 469             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 470             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 471             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 472             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 473             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 474
 475             /* Inner loop uses 93 flops */
 476         }
 477
 478         /* End of innermost loop */
 479
 480         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 481                                               f+i_coord_offset,fshift+i_shift_offset);
 482
 483         ggid                        = gid[iidx];
 484         /* Update potential energies */
 485         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 486         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 487         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 488         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 489         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 490
 491         /* Increment number of inner iterations */
 492         inneriter                  += j_index_end - j_index_start;
 493
 494         /* Outer loop uses 10 flops */
 495     }
 496
 497     /* Increment number of outer iterations */
 498     outeriter        += nri;
 499
 500     /* Update outer/inner flops */
 501
 502     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*93);
 503 }
 504 /*
 505  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse4_1_single
 506  * Electrostatics interaction: GeneralizedBorn
 507  * VdW interaction:            CubicSplineTable
 508  * Geometry:                   Particle-Particle
 509  * Calculate force/pot:        Force
 510  */
 511 void
 512 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse4_1_single
 513                     (t_nblist * gmx_restrict                nlist,
 514                      rvec * gmx_restrict                    xx,
 515                      rvec * gmx_restrict                    ff,
 516                      t_forcerec * gmx_restrict              fr,
 517                      t_mdatoms * gmx_restrict               mdatoms,
 518                      nb_kernel_data_t * gmx_restrict        kernel_data,
 519                      t_nrnb * gmx_restrict                  nrnb)
 520 {
 521     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 522      * just 0 for non-waters.
 523      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 524      * jnr indices corresponding to data put in the four positions in the SIMD register.
 525      */
 526     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 527     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 528     int              jnrA,jnrB,jnrC,jnrD;
 529     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 530     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 531     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 532     real             rcutoff_scalar;
 533     real             *shiftvec,*fshift,*x,*f;
 534     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 535     real             scratch[4*DIM];
 536     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 537     int              vdwioffset0;
 538     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 539     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 540     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 541     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 542     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 543     real             *charge;
 544     __m128i          gbitab;
 545     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 546     __m128           minushalf = _mm_set1_ps(-0.5);
 547     real             *invsqrta,*dvda,*gbtab;
 548     int              nvdwtype;
 549     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 550     int              *vdwtype;
 551     real             *vdwparam;
 552     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 553     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 554     __m128i          vfitab;
 555     __m128i          ifour       = _mm_set1_epi32(4);
 556     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 557     real             *vftab;
 558     __m128           dummy_mask,cutoff_mask;
 559     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 560     __m128           one     = _mm_set1_ps(1.0);
 561     __m128           two     = _mm_set1_ps(2.0);
 562     x                = xx[0];
 563     f                = ff[0];
 564
 565     nri              = nlist->nri;
 566     iinr             = nlist->iinr;
 567     jindex           = nlist->jindex;
 568     jjnr             = nlist->jjnr;
 569     shiftidx         = nlist->shift;
 570     gid              = nlist->gid;
 571     shiftvec         = fr->shift_vec[0];
 572     fshift           = fr->fshift[0];
 573     facel            = _mm_set1_ps(fr->epsfac);
 574     charge           = mdatoms->chargeA;
 575     nvdwtype         = fr->ntype;
 576     vdwparam         = fr->nbfp;
 577     vdwtype          = mdatoms->typeA;
 578
 579     vftab            = kernel_data->table_vdw->data;
 580     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 581
 582     invsqrta         = fr->invsqrta;
 583     dvda             = fr->dvda;
 584     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 585     gbtab            = fr->gbtab.data;
 586     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 587
 588     /* Avoid stupid compiler warnings */
 589     jnrA = jnrB = jnrC = jnrD = 0;
 590     j_coord_offsetA = 0;
 591     j_coord_offsetB = 0;
 592     j_coord_offsetC = 0;
 593     j_coord_offsetD = 0;
 594
 595     outeriter        = 0;
 596     inneriter        = 0;
 597
 598     for(iidx=0;iidx<4*DIM;iidx++)
 599     {
 600         scratch[iidx] = 0.0;
 601     }
 602
 603     /* Start outer loop over neighborlists */
 604     for(iidx=0; iidx<nri; iidx++)
 605     {
 606         /* Load shift vector for this list */
 607         i_shift_offset   = DIM*shiftidx[iidx];
 608
 609         /* Load limits for loop over neighbors */
 610         j_index_start    = jindex[iidx];
 611         j_index_end      = jindex[iidx+1];
 612
 613         /* Get outer coordinate index */
 614         inr              = iinr[iidx];
 615         i_coord_offset   = DIM*inr;
 616
 617         /* Load i particle coords and add shift vector */
 618         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 619
 620         fix0             = _mm_setzero_ps();
 621         fiy0             = _mm_setzero_ps();
 622         fiz0             = _mm_setzero_ps();
 623
 624         /* Load parameters for i particles */
 625         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 626         isai0            = _mm_load1_ps(invsqrta+inr+0);
 627         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 628
 629         dvdasum          = _mm_setzero_ps();
 630
 631         /* Start inner kernel loop */
 632         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 633         {
 634
 635             /* Get j neighbor index, and coordinate index */
 636             jnrA             = jjnr[jidx];
 637             jnrB             = jjnr[jidx+1];
 638             jnrC             = jjnr[jidx+2];
 639             jnrD             = jjnr[jidx+3];
 640             j_coord_offsetA  = DIM*jnrA;
 641             j_coord_offsetB  = DIM*jnrB;
 642             j_coord_offsetC  = DIM*jnrC;
 643             j_coord_offsetD  = DIM*jnrD;
 644
 645             /* load j atom coordinates */
 646             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 647                                               x+j_coord_offsetC,x+j_coord_offsetD,
 648                                               &jx0,&jy0,&jz0);
 649
 650             /* Calculate displacement vector */
 651             dx00             = _mm_sub_ps(ix0,jx0);
 652             dy00             = _mm_sub_ps(iy0,jy0);
 653             dz00             = _mm_sub_ps(iz0,jz0);
 654
 655             /* Calculate squared distance and things based on it */
 656             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 657
 658             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 659
 660             /* Load parameters for j particles */
 661             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 662                                                               charge+jnrC+0,charge+jnrD+0);
 663             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 664                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 665             vdwjidx0A        = 2*vdwtype[jnrA+0];
 666             vdwjidx0B        = 2*vdwtype[jnrB+0];
 667             vdwjidx0C        = 2*vdwtype[jnrC+0];
 668             vdwjidx0D        = 2*vdwtype[jnrD+0];
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             r00              = _mm_mul_ps(rsq00,rinv00);
 675
 676             /* Compute parameters for interactions between i and j atoms */
 677             qq00             = _mm_mul_ps(iq0,jq0);
 678             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 679                                          vdwparam+vdwioffset0+vdwjidx0B,
 680                                          vdwparam+vdwioffset0+vdwjidx0C,
 681                                          vdwparam+vdwioffset0+vdwjidx0D,
 682                                          &c6_00,&c12_00);
 683
 684             /* Calculate table index by multiplying r with table scale and truncate to integer */
 685             rt               = _mm_mul_ps(r00,vftabscale);
 686             vfitab           = _mm_cvttps_epi32(rt);
 687             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 688             vfitab           = _mm_slli_epi32(vfitab,3);
 689
 690             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 691             isaprod          = _mm_mul_ps(isai0,isaj0);
 692             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 693             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 694
 695             /* Calculate generalized born table index - this is a separate table from the normal one,
 696              * but we use the same procedure by multiplying r with scale and truncating to integer.
 697              */
 698             rt               = _mm_mul_ps(r00,gbscale);
 699             gbitab           = _mm_cvttps_epi32(rt);
 700             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 701             gbitab           = _mm_slli_epi32(gbitab,2);
 702             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 703             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 704             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 705             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 706             _MM_TRANSPOSE4_PS(Y,F,G,H);
 707             Heps             = _mm_mul_ps(gbeps,H);
 708             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 709             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 710             vgb              = _mm_mul_ps(gbqqfactor,VV);
 711
 712             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 713             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 714             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 715             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 716             fjptrA           = dvda+jnrA;
 717             fjptrB           = dvda+jnrB;
 718             fjptrC           = dvda+jnrC;
 719             fjptrD           = dvda+jnrD;
 720             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 721             velec            = _mm_mul_ps(qq00,rinv00);
 722             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 723
 724             /* CUBIC SPLINE TABLE DISPERSION */
 725             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 726             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 727             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 728             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 729             _MM_TRANSPOSE4_PS(Y,F,G,H);
 730             Heps             = _mm_mul_ps(vfeps,H);
 731             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 732             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 733             fvdw6            = _mm_mul_ps(c6_00,FF);
 734
 735             /* CUBIC SPLINE TABLE REPULSION */
 736             vfitab           = _mm_add_epi32(vfitab,ifour);
 737             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 738             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 739             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 740             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 741             _MM_TRANSPOSE4_PS(Y,F,G,H);
 742             Heps             = _mm_mul_ps(vfeps,H);
 743             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 744             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 745             fvdw12           = _mm_mul_ps(c12_00,FF);
 746             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 747
 748             fscal            = _mm_add_ps(felec,fvdw);
 749
 750             /* Calculate temporary vectorial force */
 751             tx               = _mm_mul_ps(fscal,dx00);
 752             ty               = _mm_mul_ps(fscal,dy00);
 753             tz               = _mm_mul_ps(fscal,dz00);
 754
 755             /* Update vectorial force */
 756             fix0             = _mm_add_ps(fix0,tx);
 757             fiy0             = _mm_add_ps(fiy0,ty);
 758             fiz0             = _mm_add_ps(fiz0,tz);
 759
 760             fjptrA             = f+j_coord_offsetA;
 761             fjptrB             = f+j_coord_offsetB;
 762             fjptrC             = f+j_coord_offsetC;
 763             fjptrD             = f+j_coord_offsetD;
 764             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 765
 766             /* Inner loop uses 82 flops */
 767         }
 768
 769         if(jidx<j_index_end)
 770         {
 771
 772             /* Get j neighbor index, and coordinate index */
 773             jnrlistA         = jjnr[jidx];
 774             jnrlistB         = jjnr[jidx+1];
 775             jnrlistC         = jjnr[jidx+2];
 776             jnrlistD         = jjnr[jidx+3];
 777             /* Sign of each element will be negative for non-real atoms.
 778              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 779              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 780              */
 781             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 782             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 783             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 784             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 785             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 786             j_coord_offsetA  = DIM*jnrA;
 787             j_coord_offsetB  = DIM*jnrB;
 788             j_coord_offsetC  = DIM*jnrC;
 789             j_coord_offsetD  = DIM*jnrD;
 790
 791             /* load j atom coordinates */
 792             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 793                                               x+j_coord_offsetC,x+j_coord_offsetD,
 794                                               &jx0,&jy0,&jz0);
 795
 796             /* Calculate displacement vector */
 797             dx00             = _mm_sub_ps(ix0,jx0);
 798             dy00             = _mm_sub_ps(iy0,jy0);
 799             dz00             = _mm_sub_ps(iz0,jz0);
 800
 801             /* Calculate squared distance and things based on it */
 802             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 803
 804             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 805
 806             /* Load parameters for j particles */
 807             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 808                                                               charge+jnrC+0,charge+jnrD+0);
 809             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 810                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 811             vdwjidx0A        = 2*vdwtype[jnrA+0];
 812             vdwjidx0B        = 2*vdwtype[jnrB+0];
 813             vdwjidx0C        = 2*vdwtype[jnrC+0];
 814             vdwjidx0D        = 2*vdwtype[jnrD+0];
 815
 816             /**************************
 817              * CALCULATE INTERACTIONS *
 818              **************************/
 819
 820             r00              = _mm_mul_ps(rsq00,rinv00);
 821             r00              = _mm_andnot_ps(dummy_mask,r00);
 822
 823             /* Compute parameters for interactions between i and j atoms */
 824             qq00             = _mm_mul_ps(iq0,jq0);
 825             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 826                                          vdwparam+vdwioffset0+vdwjidx0B,
 827                                          vdwparam+vdwioffset0+vdwjidx0C,
 828                                          vdwparam+vdwioffset0+vdwjidx0D,
 829                                          &c6_00,&c12_00);
 830
 831             /* Calculate table index by multiplying r with table scale and truncate to integer */
 832             rt               = _mm_mul_ps(r00,vftabscale);
 833             vfitab           = _mm_cvttps_epi32(rt);
 834             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 835             vfitab           = _mm_slli_epi32(vfitab,3);
 836
 837             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 838             isaprod          = _mm_mul_ps(isai0,isaj0);
 839             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 840             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 841
 842             /* Calculate generalized born table index - this is a separate table from the normal one,
 843              * but we use the same procedure by multiplying r with scale and truncating to integer.
 844              */
 845             rt               = _mm_mul_ps(r00,gbscale);
 846             gbitab           = _mm_cvttps_epi32(rt);
 847             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 848             gbitab           = _mm_slli_epi32(gbitab,2);
 849             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 850             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 851             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 852             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 853             _MM_TRANSPOSE4_PS(Y,F,G,H);
 854             Heps             = _mm_mul_ps(gbeps,H);
 855             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 856             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 857             vgb              = _mm_mul_ps(gbqqfactor,VV);
 858
 859             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 860             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 861             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 862             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 863             /* 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. */
 864             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 865             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 866             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 867             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 868             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 869             velec            = _mm_mul_ps(qq00,rinv00);
 870             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 871
 872             /* CUBIC SPLINE TABLE DISPERSION */
 873             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 874             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 875             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 876             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 877             _MM_TRANSPOSE4_PS(Y,F,G,H);
 878             Heps             = _mm_mul_ps(vfeps,H);
 879             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 880             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 881             fvdw6            = _mm_mul_ps(c6_00,FF);
 882
 883             /* CUBIC SPLINE TABLE REPULSION */
 884             vfitab           = _mm_add_epi32(vfitab,ifour);
 885             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 886             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 887             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 888             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 889             _MM_TRANSPOSE4_PS(Y,F,G,H);
 890             Heps             = _mm_mul_ps(vfeps,H);
 891             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 892             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 893             fvdw12           = _mm_mul_ps(c12_00,FF);
 894             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 895
 896             fscal            = _mm_add_ps(felec,fvdw);
 897
 898             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 899
 900             /* Calculate temporary vectorial force */
 901             tx               = _mm_mul_ps(fscal,dx00);
 902             ty               = _mm_mul_ps(fscal,dy00);
 903             tz               = _mm_mul_ps(fscal,dz00);
 904
 905             /* Update vectorial force */
 906             fix0             = _mm_add_ps(fix0,tx);
 907             fiy0             = _mm_add_ps(fiy0,ty);
 908             fiz0             = _mm_add_ps(fiz0,tz);
 909
 910             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 911             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 912             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 913             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 914             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 915
 916             /* Inner loop uses 83 flops */
 917         }
 918
 919         /* End of innermost loop */
 920
 921         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 922                                               f+i_coord_offset,fshift+i_shift_offset);
 923
 924         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 925         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 926
 927         /* Increment number of inner iterations */
 928         inneriter                  += j_index_end - j_index_start;
 929
 930         /* Outer loop uses 7 flops */
 931     }
 932
 933     /* Increment number of outer iterations */
 934     outeriter        += nri;
 935
 936     /* Update outer/inner flops */
 937
 938     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*83);
 939 }