src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse4_1_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 sse4_1_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse4_1_single.h"
  50 #include "kernelutil_x86_sse4_1_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse4_1_single
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse4_1_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128i          gbitab;
  93     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  94     __m128           minushalf = _mm_set1_ps(-0.5);
  95     real             *invsqrta,*dvda,*gbtab;
  96     int              nvdwtype;
  97     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  98     int              *vdwtype;
  99     real             *vdwparam;
 100     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 101     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 102     __m128i          vfitab;
 103     __m128i          ifour       = _mm_set1_epi32(4);
 104     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 105     real             *vftab;
 106     __m128           dummy_mask,cutoff_mask;
 107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 108     __m128           one     = _mm_set1_ps(1.0);
 109     __m128           two     = _mm_set1_ps(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm_set1_ps(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123     nvdwtype         = fr->ntype;
 124     vdwparam         = fr->nbfp;
 125     vdwtype          = mdatoms->typeA;
 126
 127     vftab            = kernel_data->table_vdw->data;
 128     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 129
 130     invsqrta         = fr->invsqrta;
 131     dvda             = fr->dvda;
 132     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 133     gbtab            = fr->gbtab.data;
 134     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 135
 136     /* Avoid stupid compiler warnings */
 137     jnrA = jnrB = jnrC = jnrD = 0;
 138     j_coord_offsetA = 0;
 139     j_coord_offsetB = 0;
 140     j_coord_offsetC = 0;
 141     j_coord_offsetD = 0;
 142
 143     outeriter        = 0;
 144     inneriter        = 0;
 145
 146     for(iidx=0;iidx<4*DIM;iidx++)
 147     {
 148         scratch[iidx] = 0.0;
 149     }
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 167
 168         fix0             = _mm_setzero_ps();
 169         fiy0             = _mm_setzero_ps();
 170         fiz0             = _mm_setzero_ps();
 171
 172         /* Load parameters for i particles */
 173         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 174         isai0            = _mm_load1_ps(invsqrta+inr+0);
 175         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 176
 177         /* Reset potential sums */
 178         velecsum         = _mm_setzero_ps();
 179         vgbsum           = _mm_setzero_ps();
 180         vvdwsum          = _mm_setzero_ps();
 181         dvdasum          = _mm_setzero_ps();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               x+j_coord_offsetC,x+j_coord_offsetD,
 200                                               &jx0,&jy0,&jz0);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm_sub_ps(ix0,jx0);
 204             dy00             = _mm_sub_ps(iy0,jy0);
 205             dz00             = _mm_sub_ps(iz0,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 209
 210             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 211
 212             /* Load parameters for j particles */
 213             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 214                                                               charge+jnrC+0,charge+jnrD+0);
 215             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 216                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 217             vdwjidx0A        = 2*vdwtype[jnrA+0];
 218             vdwjidx0B        = 2*vdwtype[jnrB+0];
 219             vdwjidx0C        = 2*vdwtype[jnrC+0];
 220             vdwjidx0D        = 2*vdwtype[jnrD+0];
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             r00              = _mm_mul_ps(rsq00,rinv00);
 227
 228             /* Compute parameters for interactions between i and j atoms */
 229             qq00             = _mm_mul_ps(iq0,jq0);
 230             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 231                                          vdwparam+vdwioffset0+vdwjidx0B,
 232                                          vdwparam+vdwioffset0+vdwjidx0C,
 233                                          vdwparam+vdwioffset0+vdwjidx0D,
 234                                          &c6_00,&c12_00);
 235
 236             /* Calculate table index by multiplying r with table scale and truncate to integer */
 237             rt               = _mm_mul_ps(r00,vftabscale);
 238             vfitab           = _mm_cvttps_epi32(rt);
 239             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 240             vfitab           = _mm_slli_epi32(vfitab,3);
 241
 242             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 243             isaprod          = _mm_mul_ps(isai0,isaj0);
 244             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 245             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 246
 247             /* Calculate generalized born table index - this is a separate table from the normal one,
 248              * but we use the same procedure by multiplying r with scale and truncating to integer.
 249              */
 250             rt               = _mm_mul_ps(r00,gbscale);
 251             gbitab           = _mm_cvttps_epi32(rt);
 252             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 253             gbitab           = _mm_slli_epi32(gbitab,2);
 254             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 255             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 256             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 257             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 258             _MM_TRANSPOSE4_PS(Y,F,G,H);
 259             Heps             = _mm_mul_ps(gbeps,H);
 260             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 261             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 262             vgb              = _mm_mul_ps(gbqqfactor,VV);
 263
 264             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 265             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 266             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 267             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 268             fjptrA           = dvda+jnrA;
 269             fjptrB           = dvda+jnrB;
 270             fjptrC           = dvda+jnrC;
 271             fjptrD           = dvda+jnrD;
 272             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 273             velec            = _mm_mul_ps(qq00,rinv00);
 274             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 275
 276             /* CUBIC SPLINE TABLE DISPERSION */
 277             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 278             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 279             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 280             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 281             _MM_TRANSPOSE4_PS(Y,F,G,H);
 282             Heps             = _mm_mul_ps(vfeps,H);
 283             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 284             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 285             vvdw6            = _mm_mul_ps(c6_00,VV);
 286             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 287             fvdw6            = _mm_mul_ps(c6_00,FF);
 288
 289             /* CUBIC SPLINE TABLE REPULSION */
 290             vfitab           = _mm_add_epi32(vfitab,ifour);
 291             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 292             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 293             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 294             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 295             _MM_TRANSPOSE4_PS(Y,F,G,H);
 296             Heps             = _mm_mul_ps(vfeps,H);
 297             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 298             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 299             vvdw12           = _mm_mul_ps(c12_00,VV);
 300             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 301             fvdw12           = _mm_mul_ps(c12_00,FF);
 302             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 303             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velecsum         = _mm_add_ps(velecsum,velec);
 307             vgbsum           = _mm_add_ps(vgbsum,vgb);
 308             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 309
 310             fscal            = _mm_add_ps(felec,fvdw);
 311
 312             /* Calculate temporary vectorial force */
 313             tx               = _mm_mul_ps(fscal,dx00);
 314             ty               = _mm_mul_ps(fscal,dy00);
 315             tz               = _mm_mul_ps(fscal,dz00);
 316
 317             /* Update vectorial force */
 318             fix0             = _mm_add_ps(fix0,tx);
 319             fiy0             = _mm_add_ps(fiy0,ty);
 320             fiz0             = _mm_add_ps(fiz0,tz);
 321
 322             fjptrA             = f+j_coord_offsetA;
 323             fjptrB             = f+j_coord_offsetB;
 324             fjptrC             = f+j_coord_offsetC;
 325             fjptrD             = f+j_coord_offsetD;
 326             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 327
 328             /* Inner loop uses 92 flops */
 329         }
 330
 331         if(jidx<j_index_end)
 332         {
 333
 334             /* Get j neighbor index, and coordinate index */
 335             jnrlistA         = jjnr[jidx];
 336             jnrlistB         = jjnr[jidx+1];
 337             jnrlistC         = jjnr[jidx+2];
 338             jnrlistD         = jjnr[jidx+3];
 339             /* Sign of each element will be negative for non-real atoms.
 340              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 341              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 342              */
 343             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 344             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 345             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 346             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 347             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 348             j_coord_offsetA  = DIM*jnrA;
 349             j_coord_offsetB  = DIM*jnrB;
 350             j_coord_offsetC  = DIM*jnrC;
 351             j_coord_offsetD  = DIM*jnrD;
 352
 353             /* load j atom coordinates */
 354             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 355                                               x+j_coord_offsetC,x+j_coord_offsetD,
 356                                               &jx0,&jy0,&jz0);
 357
 358             /* Calculate displacement vector */
 359             dx00             = _mm_sub_ps(ix0,jx0);
 360             dy00             = _mm_sub_ps(iy0,jy0);
 361             dz00             = _mm_sub_ps(iz0,jz0);
 362
 363             /* Calculate squared distance and things based on it */
 364             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 365
 366             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 367
 368             /* Load parameters for j particles */
 369             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 370                                                               charge+jnrC+0,charge+jnrD+0);
 371             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 372                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 373             vdwjidx0A        = 2*vdwtype[jnrA+0];
 374             vdwjidx0B        = 2*vdwtype[jnrB+0];
 375             vdwjidx0C        = 2*vdwtype[jnrC+0];
 376             vdwjidx0D        = 2*vdwtype[jnrD+0];
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             r00              = _mm_mul_ps(rsq00,rinv00);
 383             r00              = _mm_andnot_ps(dummy_mask,r00);
 384
 385             /* Compute parameters for interactions between i and j atoms */
 386             qq00             = _mm_mul_ps(iq0,jq0);
 387             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 388                                          vdwparam+vdwioffset0+vdwjidx0B,
 389                                          vdwparam+vdwioffset0+vdwjidx0C,
 390                                          vdwparam+vdwioffset0+vdwjidx0D,
 391                                          &c6_00,&c12_00);
 392
 393             /* Calculate table index by multiplying r with table scale and truncate to integer */
 394             rt               = _mm_mul_ps(r00,vftabscale);
 395             vfitab           = _mm_cvttps_epi32(rt);
 396             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 397             vfitab           = _mm_slli_epi32(vfitab,3);
 398
 399             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 400             isaprod          = _mm_mul_ps(isai0,isaj0);
 401             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 402             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 403
 404             /* Calculate generalized born table index - this is a separate table from the normal one,
 405              * but we use the same procedure by multiplying r with scale and truncating to integer.
 406              */
 407             rt               = _mm_mul_ps(r00,gbscale);
 408             gbitab           = _mm_cvttps_epi32(rt);
 409             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 410             gbitab           = _mm_slli_epi32(gbitab,2);
 411             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 412             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 413             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 414             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 415             _MM_TRANSPOSE4_PS(Y,F,G,H);
 416             Heps             = _mm_mul_ps(gbeps,H);
 417             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 418             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 419             vgb              = _mm_mul_ps(gbqqfactor,VV);
 420
 421             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 422             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 423             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 424             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 425             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 426             /* 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. */
 427             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 428             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 429             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 430             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 431             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 432             velec            = _mm_mul_ps(qq00,rinv00);
 433             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 434
 435             /* CUBIC SPLINE TABLE DISPERSION */
 436             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 437             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 438             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 439             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 440             _MM_TRANSPOSE4_PS(Y,F,G,H);
 441             Heps             = _mm_mul_ps(vfeps,H);
 442             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 443             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 444             vvdw6            = _mm_mul_ps(c6_00,VV);
 445             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 446             fvdw6            = _mm_mul_ps(c6_00,FF);
 447
 448             /* CUBIC SPLINE TABLE REPULSION */
 449             vfitab           = _mm_add_epi32(vfitab,ifour);
 450             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 451             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 452             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 453             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 454             _MM_TRANSPOSE4_PS(Y,F,G,H);
 455             Heps             = _mm_mul_ps(vfeps,H);
 456             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 457             VV               = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
 458             vvdw12           = _mm_mul_ps(c12_00,VV);
 459             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 460             fvdw12           = _mm_mul_ps(c12_00,FF);
 461             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 462             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 463
 464             /* Update potential sum for this i atom from the interaction with this j atom. */
 465             velec            = _mm_andnot_ps(dummy_mask,velec);
 466             velecsum         = _mm_add_ps(velecsum,velec);
 467             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 468             vgbsum           = _mm_add_ps(vgbsum,vgb);
 469             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 470             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 471
 472             fscal            = _mm_add_ps(felec,fvdw);
 473
 474             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 475
 476             /* Calculate temporary vectorial force */
 477             tx               = _mm_mul_ps(fscal,dx00);
 478             ty               = _mm_mul_ps(fscal,dy00);
 479             tz               = _mm_mul_ps(fscal,dz00);
 480
 481             /* Update vectorial force */
 482             fix0             = _mm_add_ps(fix0,tx);
 483             fiy0             = _mm_add_ps(fiy0,ty);
 484             fiz0             = _mm_add_ps(fiz0,tz);
 485
 486             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 487             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 488             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 489             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 490             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 491
 492             /* Inner loop uses 93 flops */
 493         }
 494
 495         /* End of innermost loop */
 496
 497         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 498                                               f+i_coord_offset,fshift+i_shift_offset);
 499
 500         ggid                        = gid[iidx];
 501         /* Update potential energies */
 502         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 503         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 504         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 505         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 506         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 507
 508         /* Increment number of inner iterations */
 509         inneriter                  += j_index_end - j_index_start;
 510
 511         /* Outer loop uses 10 flops */
 512     }
 513
 514     /* Increment number of outer iterations */
 515     outeriter        += nri;
 516
 517     /* Update outer/inner flops */
 518
 519     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*93);
 520 }
 521 /*
 522  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse4_1_single
 523  * Electrostatics interaction: GeneralizedBorn
 524  * VdW interaction:            CubicSplineTable
 525  * Geometry:                   Particle-Particle
 526  * Calculate force/pot:        Force
 527  */
 528 void
 529 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse4_1_single
 530                     (t_nblist                    * gmx_restrict       nlist,
 531                      rvec                        * gmx_restrict          xx,
 532                      rvec                        * gmx_restrict          ff,
 533                      t_forcerec                  * gmx_restrict          fr,
 534                      t_mdatoms                   * gmx_restrict     mdatoms,
 535                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 536                      t_nrnb                      * gmx_restrict        nrnb)
 537 {
 538     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 539      * just 0 for non-waters.
 540      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 541      * jnr indices corresponding to data put in the four positions in the SIMD register.
 542      */
 543     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 544     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 545     int              jnrA,jnrB,jnrC,jnrD;
 546     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 547     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 548     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 549     real             rcutoff_scalar;
 550     real             *shiftvec,*fshift,*x,*f;
 551     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 552     real             scratch[4*DIM];
 553     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 554     int              vdwioffset0;
 555     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 556     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 557     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 558     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 559     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 560     real             *charge;
 561     __m128i          gbitab;
 562     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 563     __m128           minushalf = _mm_set1_ps(-0.5);
 564     real             *invsqrta,*dvda,*gbtab;
 565     int              nvdwtype;
 566     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 567     int              *vdwtype;
 568     real             *vdwparam;
 569     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 570     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 571     __m128i          vfitab;
 572     __m128i          ifour       = _mm_set1_epi32(4);
 573     __m128           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 574     real             *vftab;
 575     __m128           dummy_mask,cutoff_mask;
 576     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 577     __m128           one     = _mm_set1_ps(1.0);
 578     __m128           two     = _mm_set1_ps(2.0);
 579     x                = xx[0];
 580     f                = ff[0];
 581
 582     nri              = nlist->nri;
 583     iinr             = nlist->iinr;
 584     jindex           = nlist->jindex;
 585     jjnr             = nlist->jjnr;
 586     shiftidx         = nlist->shift;
 587     gid              = nlist->gid;
 588     shiftvec         = fr->shift_vec[0];
 589     fshift           = fr->fshift[0];
 590     facel            = _mm_set1_ps(fr->epsfac);
 591     charge           = mdatoms->chargeA;
 592     nvdwtype         = fr->ntype;
 593     vdwparam         = fr->nbfp;
 594     vdwtype          = mdatoms->typeA;
 595
 596     vftab            = kernel_data->table_vdw->data;
 597     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 598
 599     invsqrta         = fr->invsqrta;
 600     dvda             = fr->dvda;
 601     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 602     gbtab            = fr->gbtab.data;
 603     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 604
 605     /* Avoid stupid compiler warnings */
 606     jnrA = jnrB = jnrC = jnrD = 0;
 607     j_coord_offsetA = 0;
 608     j_coord_offsetB = 0;
 609     j_coord_offsetC = 0;
 610     j_coord_offsetD = 0;
 611
 612     outeriter        = 0;
 613     inneriter        = 0;
 614
 615     for(iidx=0;iidx<4*DIM;iidx++)
 616     {
 617         scratch[iidx] = 0.0;
 618     }
 619
 620     /* Start outer loop over neighborlists */
 621     for(iidx=0; iidx<nri; iidx++)
 622     {
 623         /* Load shift vector for this list */
 624         i_shift_offset   = DIM*shiftidx[iidx];
 625
 626         /* Load limits for loop over neighbors */
 627         j_index_start    = jindex[iidx];
 628         j_index_end      = jindex[iidx+1];
 629
 630         /* Get outer coordinate index */
 631         inr              = iinr[iidx];
 632         i_coord_offset   = DIM*inr;
 633
 634         /* Load i particle coords and add shift vector */
 635         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 636
 637         fix0             = _mm_setzero_ps();
 638         fiy0             = _mm_setzero_ps();
 639         fiz0             = _mm_setzero_ps();
 640
 641         /* Load parameters for i particles */
 642         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 643         isai0            = _mm_load1_ps(invsqrta+inr+0);
 644         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 645
 646         dvdasum          = _mm_setzero_ps();
 647
 648         /* Start inner kernel loop */
 649         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 650         {
 651
 652             /* Get j neighbor index, and coordinate index */
 653             jnrA             = jjnr[jidx];
 654             jnrB             = jjnr[jidx+1];
 655             jnrC             = jjnr[jidx+2];
 656             jnrD             = jjnr[jidx+3];
 657             j_coord_offsetA  = DIM*jnrA;
 658             j_coord_offsetB  = DIM*jnrB;
 659             j_coord_offsetC  = DIM*jnrC;
 660             j_coord_offsetD  = DIM*jnrD;
 661
 662             /* load j atom coordinates */
 663             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 664                                               x+j_coord_offsetC,x+j_coord_offsetD,
 665                                               &jx0,&jy0,&jz0);
 666
 667             /* Calculate displacement vector */
 668             dx00             = _mm_sub_ps(ix0,jx0);
 669             dy00             = _mm_sub_ps(iy0,jy0);
 670             dz00             = _mm_sub_ps(iz0,jz0);
 671
 672             /* Calculate squared distance and things based on it */
 673             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 674
 675             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 676
 677             /* Load parameters for j particles */
 678             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 679                                                               charge+jnrC+0,charge+jnrD+0);
 680             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 681                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 682             vdwjidx0A        = 2*vdwtype[jnrA+0];
 683             vdwjidx0B        = 2*vdwtype[jnrB+0];
 684             vdwjidx0C        = 2*vdwtype[jnrC+0];
 685             vdwjidx0D        = 2*vdwtype[jnrD+0];
 686
 687             /**************************
 688              * CALCULATE INTERACTIONS *
 689              **************************/
 690
 691             r00              = _mm_mul_ps(rsq00,rinv00);
 692
 693             /* Compute parameters for interactions between i and j atoms */
 694             qq00             = _mm_mul_ps(iq0,jq0);
 695             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 696                                          vdwparam+vdwioffset0+vdwjidx0B,
 697                                          vdwparam+vdwioffset0+vdwjidx0C,
 698                                          vdwparam+vdwioffset0+vdwjidx0D,
 699                                          &c6_00,&c12_00);
 700
 701             /* Calculate table index by multiplying r with table scale and truncate to integer */
 702             rt               = _mm_mul_ps(r00,vftabscale);
 703             vfitab           = _mm_cvttps_epi32(rt);
 704             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 705             vfitab           = _mm_slli_epi32(vfitab,3);
 706
 707             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 708             isaprod          = _mm_mul_ps(isai0,isaj0);
 709             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 710             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 711
 712             /* Calculate generalized born table index - this is a separate table from the normal one,
 713              * but we use the same procedure by multiplying r with scale and truncating to integer.
 714              */
 715             rt               = _mm_mul_ps(r00,gbscale);
 716             gbitab           = _mm_cvttps_epi32(rt);
 717             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 718             gbitab           = _mm_slli_epi32(gbitab,2);
 719             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 720             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 721             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 722             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 723             _MM_TRANSPOSE4_PS(Y,F,G,H);
 724             Heps             = _mm_mul_ps(gbeps,H);
 725             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 726             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 727             vgb              = _mm_mul_ps(gbqqfactor,VV);
 728
 729             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 730             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 731             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 732             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 733             fjptrA           = dvda+jnrA;
 734             fjptrB           = dvda+jnrB;
 735             fjptrC           = dvda+jnrC;
 736             fjptrD           = dvda+jnrD;
 737             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 738             velec            = _mm_mul_ps(qq00,rinv00);
 739             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 740
 741             /* CUBIC SPLINE TABLE DISPERSION */
 742             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 743             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 744             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 745             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 746             _MM_TRANSPOSE4_PS(Y,F,G,H);
 747             Heps             = _mm_mul_ps(vfeps,H);
 748             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 749             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 750             fvdw6            = _mm_mul_ps(c6_00,FF);
 751
 752             /* CUBIC SPLINE TABLE REPULSION */
 753             vfitab           = _mm_add_epi32(vfitab,ifour);
 754             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 755             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 756             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 757             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 758             _MM_TRANSPOSE4_PS(Y,F,G,H);
 759             Heps             = _mm_mul_ps(vfeps,H);
 760             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 761             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 762             fvdw12           = _mm_mul_ps(c12_00,FF);
 763             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 764
 765             fscal            = _mm_add_ps(felec,fvdw);
 766
 767             /* Calculate temporary vectorial force */
 768             tx               = _mm_mul_ps(fscal,dx00);
 769             ty               = _mm_mul_ps(fscal,dy00);
 770             tz               = _mm_mul_ps(fscal,dz00);
 771
 772             /* Update vectorial force */
 773             fix0             = _mm_add_ps(fix0,tx);
 774             fiy0             = _mm_add_ps(fiy0,ty);
 775             fiz0             = _mm_add_ps(fiz0,tz);
 776
 777             fjptrA             = f+j_coord_offsetA;
 778             fjptrB             = f+j_coord_offsetB;
 779             fjptrC             = f+j_coord_offsetC;
 780             fjptrD             = f+j_coord_offsetD;
 781             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 782
 783             /* Inner loop uses 82 flops */
 784         }
 785
 786         if(jidx<j_index_end)
 787         {
 788
 789             /* Get j neighbor index, and coordinate index */
 790             jnrlistA         = jjnr[jidx];
 791             jnrlistB         = jjnr[jidx+1];
 792             jnrlistC         = jjnr[jidx+2];
 793             jnrlistD         = jjnr[jidx+3];
 794             /* Sign of each element will be negative for non-real atoms.
 795              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 796              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 797              */
 798             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 799             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 800             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 801             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 802             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 803             j_coord_offsetA  = DIM*jnrA;
 804             j_coord_offsetB  = DIM*jnrB;
 805             j_coord_offsetC  = DIM*jnrC;
 806             j_coord_offsetD  = DIM*jnrD;
 807
 808             /* load j atom coordinates */
 809             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 810                                               x+j_coord_offsetC,x+j_coord_offsetD,
 811                                               &jx0,&jy0,&jz0);
 812
 813             /* Calculate displacement vector */
 814             dx00             = _mm_sub_ps(ix0,jx0);
 815             dy00             = _mm_sub_ps(iy0,jy0);
 816             dz00             = _mm_sub_ps(iz0,jz0);
 817
 818             /* Calculate squared distance and things based on it */
 819             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 820
 821             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 822
 823             /* Load parameters for j particles */
 824             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 825                                                               charge+jnrC+0,charge+jnrD+0);
 826             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 827                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 828             vdwjidx0A        = 2*vdwtype[jnrA+0];
 829             vdwjidx0B        = 2*vdwtype[jnrB+0];
 830             vdwjidx0C        = 2*vdwtype[jnrC+0];
 831             vdwjidx0D        = 2*vdwtype[jnrD+0];
 832
 833             /**************************
 834              * CALCULATE INTERACTIONS *
 835              **************************/
 836
 837             r00              = _mm_mul_ps(rsq00,rinv00);
 838             r00              = _mm_andnot_ps(dummy_mask,r00);
 839
 840             /* Compute parameters for interactions between i and j atoms */
 841             qq00             = _mm_mul_ps(iq0,jq0);
 842             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 843                                          vdwparam+vdwioffset0+vdwjidx0B,
 844                                          vdwparam+vdwioffset0+vdwjidx0C,
 845                                          vdwparam+vdwioffset0+vdwjidx0D,
 846                                          &c6_00,&c12_00);
 847
 848             /* Calculate table index by multiplying r with table scale and truncate to integer */
 849             rt               = _mm_mul_ps(r00,vftabscale);
 850             vfitab           = _mm_cvttps_epi32(rt);
 851             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 852             vfitab           = _mm_slli_epi32(vfitab,3);
 853
 854             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 855             isaprod          = _mm_mul_ps(isai0,isaj0);
 856             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 857             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 858
 859             /* Calculate generalized born table index - this is a separate table from the normal one,
 860              * but we use the same procedure by multiplying r with scale and truncating to integer.
 861              */
 862             rt               = _mm_mul_ps(r00,gbscale);
 863             gbitab           = _mm_cvttps_epi32(rt);
 864             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 865             gbitab           = _mm_slli_epi32(gbitab,2);
 866             Y                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
 867             F                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
 868             G                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
 869             H                = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
 870             _MM_TRANSPOSE4_PS(Y,F,G,H);
 871             Heps             = _mm_mul_ps(gbeps,H);
 872             Fp               = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
 873             VV               = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
 874             vgb              = _mm_mul_ps(gbqqfactor,VV);
 875
 876             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 877             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 878             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
 879             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 880             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 881             /* 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. */
 882             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 883             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 884             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 885             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 886             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 887             velec            = _mm_mul_ps(qq00,rinv00);
 888             felec            = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
 889
 890             /* CUBIC SPLINE TABLE DISPERSION */
 891             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 892             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 893             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 894             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 895             _MM_TRANSPOSE4_PS(Y,F,G,H);
 896             Heps             = _mm_mul_ps(vfeps,H);
 897             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 898             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 899             fvdw6            = _mm_mul_ps(c6_00,FF);
 900
 901             /* CUBIC SPLINE TABLE REPULSION */
 902             vfitab           = _mm_add_epi32(vfitab,ifour);
 903             Y                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
 904             F                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
 905             G                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
 906             H                = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
 907             _MM_TRANSPOSE4_PS(Y,F,G,H);
 908             Heps             = _mm_mul_ps(vfeps,H);
 909             Fp               = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
 910             FF               = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
 911             fvdw12           = _mm_mul_ps(c12_00,FF);
 912             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 913
 914             fscal            = _mm_add_ps(felec,fvdw);
 915
 916             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 917
 918             /* Calculate temporary vectorial force */
 919             tx               = _mm_mul_ps(fscal,dx00);
 920             ty               = _mm_mul_ps(fscal,dy00);
 921             tz               = _mm_mul_ps(fscal,dz00);
 922
 923             /* Update vectorial force */
 924             fix0             = _mm_add_ps(fix0,tx);
 925             fiy0             = _mm_add_ps(fiy0,ty);
 926             fiz0             = _mm_add_ps(fiz0,tz);
 927
 928             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 929             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 930             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 931             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 932             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 933
 934             /* Inner loop uses 83 flops */
 935         }
 936
 937         /* End of innermost loop */
 938
 939         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 940                                               f+i_coord_offset,fshift+i_shift_offset);
 941
 942         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 943         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 944
 945         /* Increment number of inner iterations */
 946         inneriter                  += j_index_end - j_index_start;
 947
 948         /* Outer loop uses 7 flops */
 949     }
 950
 951     /* Increment number of outer iterations */
 952     outeriter        += nri;
 953
 954     /* Update outer/inner flops */
 955
 956     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*83);
 957 }