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