src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_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_avx_128_fma_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_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 AVX_128, 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           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,twogbeps,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,twovfeps,vftabscale,Y,F,G,H,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 #ifdef __XOP__
 240             vfeps            = _mm_frcz_ps(rt);
 241 #else
 242             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 243 #endif
 244             twovfeps         = _mm_add_ps(vfeps,vfeps);
 245             vfitab           = _mm_slli_epi32(vfitab,3);
 246
 247             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 248             isaprod          = _mm_mul_ps(isai0,isaj0);
 249             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 250             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 251
 252             /* Calculate generalized born table index - this is a separate table from the normal one,
 253              * but we use the same procedure by multiplying r with scale and truncating to integer.
 254              */
 255             rt               = _mm_mul_ps(r00,gbscale);
 256             gbitab           = _mm_cvttps_epi32(rt);
 257 #ifdef __XOP__
 258             gbeps            = _mm_frcz_ps(rt);
 259 #else
 260             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 261 #endif
 262             gbitab           = _mm_slli_epi32(gbitab,2);
 263
 264             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 265             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 266             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 267             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 268             _MM_TRANSPOSE4_PS(Y,F,G,H);
 269             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 270             VV               = _mm_macc_ps(gbeps,Fp,Y);
 271             vgb              = _mm_mul_ps(gbqqfactor,VV);
 272
 273             twogbeps         = _mm_add_ps(gbeps,gbeps);
 274             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 275             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 276             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 277             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 278             fjptrA           = dvda+jnrA;
 279             fjptrB           = dvda+jnrB;
 280             fjptrC           = dvda+jnrC;
 281             fjptrD           = dvda+jnrD;
 282             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 283             velec            = _mm_mul_ps(qq00,rinv00);
 284             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 285
 286             /* CUBIC SPLINE TABLE DISPERSION */
 287             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 288             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 289             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 290             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 291             _MM_TRANSPOSE4_PS(Y,F,G,H);
 292             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 293             VV               = _mm_macc_ps(vfeps,Fp,Y);
 294             vvdw6            = _mm_mul_ps(c6_00,VV);
 295             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 296             fvdw6            = _mm_mul_ps(c6_00,FF);
 297
 298             /* CUBIC SPLINE TABLE REPULSION */
 299             vfitab           = _mm_add_epi32(vfitab,ifour);
 300             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 301             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 302             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 303             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 304             _MM_TRANSPOSE4_PS(Y,F,G,H);
 305             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 306             VV               = _mm_macc_ps(vfeps,Fp,Y);
 307             vvdw12           = _mm_mul_ps(c12_00,VV);
 308             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 309             fvdw12           = _mm_mul_ps(c12_00,FF);
 310             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 311             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 312
 313             /* Update potential sum for this i atom from the interaction with this j atom. */
 314             velecsum         = _mm_add_ps(velecsum,velec);
 315             vgbsum           = _mm_add_ps(vgbsum,vgb);
 316             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 317
 318             fscal            = _mm_add_ps(felec,fvdw);
 319
 320              /* Update vectorial force */
 321             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 322             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 323             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 324
 325             fjptrA             = f+j_coord_offsetA;
 326             fjptrB             = f+j_coord_offsetB;
 327             fjptrC             = f+j_coord_offsetC;
 328             fjptrD             = f+j_coord_offsetD;
 329             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 330                                                    _mm_mul_ps(dx00,fscal),
 331                                                    _mm_mul_ps(dy00,fscal),
 332                                                    _mm_mul_ps(dz00,fscal));
 333
 334             /* Inner loop uses 95 flops */
 335         }
 336
 337         if(jidx<j_index_end)
 338         {
 339
 340             /* Get j neighbor index, and coordinate index */
 341             jnrlistA         = jjnr[jidx];
 342             jnrlistB         = jjnr[jidx+1];
 343             jnrlistC         = jjnr[jidx+2];
 344             jnrlistD         = jjnr[jidx+3];
 345             /* Sign of each element will be negative for non-real atoms.
 346              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 347              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 348              */
 349             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 350             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 351             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 352             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 353             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 354             j_coord_offsetA  = DIM*jnrA;
 355             j_coord_offsetB  = DIM*jnrB;
 356             j_coord_offsetC  = DIM*jnrC;
 357             j_coord_offsetD  = DIM*jnrD;
 358
 359             /* load j atom coordinates */
 360             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 361                                               x+j_coord_offsetC,x+j_coord_offsetD,
 362                                               &jx0,&jy0,&jz0);
 363
 364             /* Calculate displacement vector */
 365             dx00             = _mm_sub_ps(ix0,jx0);
 366             dy00             = _mm_sub_ps(iy0,jy0);
 367             dz00             = _mm_sub_ps(iz0,jz0);
 368
 369             /* Calculate squared distance and things based on it */
 370             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 371
 372             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 373
 374             /* Load parameters for j particles */
 375             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 376                                                               charge+jnrC+0,charge+jnrD+0);
 377             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 378                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 379             vdwjidx0A        = 2*vdwtype[jnrA+0];
 380             vdwjidx0B        = 2*vdwtype[jnrB+0];
 381             vdwjidx0C        = 2*vdwtype[jnrC+0];
 382             vdwjidx0D        = 2*vdwtype[jnrD+0];
 383
 384             /**************************
 385              * CALCULATE INTERACTIONS *
 386              **************************/
 387
 388             r00              = _mm_mul_ps(rsq00,rinv00);
 389             r00              = _mm_andnot_ps(dummy_mask,r00);
 390
 391             /* Compute parameters for interactions between i and j atoms */
 392             qq00             = _mm_mul_ps(iq0,jq0);
 393             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 394                                          vdwparam+vdwioffset0+vdwjidx0B,
 395                                          vdwparam+vdwioffset0+vdwjidx0C,
 396                                          vdwparam+vdwioffset0+vdwjidx0D,
 397                                          &c6_00,&c12_00);
 398
 399             /* Calculate table index by multiplying r with table scale and truncate to integer */
 400             rt               = _mm_mul_ps(r00,vftabscale);
 401             vfitab           = _mm_cvttps_epi32(rt);
 402 #ifdef __XOP__
 403             vfeps            = _mm_frcz_ps(rt);
 404 #else
 405             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 406 #endif
 407             twovfeps         = _mm_add_ps(vfeps,vfeps);
 408             vfitab           = _mm_slli_epi32(vfitab,3);
 409
 410             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 411             isaprod          = _mm_mul_ps(isai0,isaj0);
 412             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 413             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 414
 415             /* Calculate generalized born table index - this is a separate table from the normal one,
 416              * but we use the same procedure by multiplying r with scale and truncating to integer.
 417              */
 418             rt               = _mm_mul_ps(r00,gbscale);
 419             gbitab           = _mm_cvttps_epi32(rt);
 420 #ifdef __XOP__
 421             gbeps            = _mm_frcz_ps(rt);
 422 #else
 423             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 424 #endif
 425             gbitab           = _mm_slli_epi32(gbitab,2);
 426
 427             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 428             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 429             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 430             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 431             _MM_TRANSPOSE4_PS(Y,F,G,H);
 432             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 433             VV               = _mm_macc_ps(gbeps,Fp,Y);
 434             vgb              = _mm_mul_ps(gbqqfactor,VV);
 435
 436             twogbeps         = _mm_add_ps(gbeps,gbeps);
 437             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 438             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 439             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 440             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 441             /* 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. */
 442             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 443             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 444             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 445             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 446             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 447             velec            = _mm_mul_ps(qq00,rinv00);
 448             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 449
 450             /* CUBIC SPLINE TABLE DISPERSION */
 451             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 452             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 453             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 454             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 455             _MM_TRANSPOSE4_PS(Y,F,G,H);
 456             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 457             VV               = _mm_macc_ps(vfeps,Fp,Y);
 458             vvdw6            = _mm_mul_ps(c6_00,VV);
 459             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 460             fvdw6            = _mm_mul_ps(c6_00,FF);
 461
 462             /* CUBIC SPLINE TABLE REPULSION */
 463             vfitab           = _mm_add_epi32(vfitab,ifour);
 464             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 465             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 466             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 467             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 468             _MM_TRANSPOSE4_PS(Y,F,G,H);
 469             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 470             VV               = _mm_macc_ps(vfeps,Fp,Y);
 471             vvdw12           = _mm_mul_ps(c12_00,VV);
 472             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 473             fvdw12           = _mm_mul_ps(c12_00,FF);
 474             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 475             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velec            = _mm_andnot_ps(dummy_mask,velec);
 479             velecsum         = _mm_add_ps(velecsum,velec);
 480             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 481             vgbsum           = _mm_add_ps(vgbsum,vgb);
 482             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 483             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 484
 485             fscal            = _mm_add_ps(felec,fvdw);
 486
 487             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 488
 489              /* Update vectorial force */
 490             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 491             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 492             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 493
 494             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 495             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 496             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 497             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 498             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 499                                                    _mm_mul_ps(dx00,fscal),
 500                                                    _mm_mul_ps(dy00,fscal),
 501                                                    _mm_mul_ps(dz00,fscal));
 502
 503             /* Inner loop uses 96 flops */
 504         }
 505
 506         /* End of innermost loop */
 507
 508         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 509                                               f+i_coord_offset,fshift+i_shift_offset);
 510
 511         ggid                        = gid[iidx];
 512         /* Update potential energies */
 513         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 514         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 515         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 516         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 517         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 518
 519         /* Increment number of inner iterations */
 520         inneriter                  += j_index_end - j_index_start;
 521
 522         /* Outer loop uses 10 flops */
 523     }
 524
 525     /* Increment number of outer iterations */
 526     outeriter        += nri;
 527
 528     /* Update outer/inner flops */
 529
 530     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*96);
 531 }
 532 /*
 533  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 534  * Electrostatics interaction: GeneralizedBorn
 535  * VdW interaction:            CubicSplineTable
 536  * Geometry:                   Particle-Particle
 537  * Calculate force/pot:        Force
 538  */
 539 void
 540 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 541                     (t_nblist * gmx_restrict                nlist,
 542                      rvec * gmx_restrict                    xx,
 543                      rvec * gmx_restrict                    ff,
 544                      t_forcerec * gmx_restrict              fr,
 545                      t_mdatoms * gmx_restrict               mdatoms,
 546                      nb_kernel_data_t * gmx_restrict        kernel_data,
 547                      t_nrnb * gmx_restrict                  nrnb)
 548 {
 549     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 550      * just 0 for non-waters.
 551      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 552      * jnr indices corresponding to data put in the four positions in the SIMD register.
 553      */
 554     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 555     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 556     int              jnrA,jnrB,jnrC,jnrD;
 557     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 558     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 559     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 560     real             rcutoff_scalar;
 561     real             *shiftvec,*fshift,*x,*f;
 562     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 563     real             scratch[4*DIM];
 564     __m128           fscal,rcutoff,rcutoff2,jidxall;
 565     int              vdwioffset0;
 566     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 567     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 568     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 569     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 570     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 571     real             *charge;
 572     __m128i          gbitab;
 573     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 574     __m128           minushalf = _mm_set1_ps(-0.5);
 575     real             *invsqrta,*dvda,*gbtab;
 576     int              nvdwtype;
 577     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 578     int              *vdwtype;
 579     real             *vdwparam;
 580     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 581     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 582     __m128i          vfitab;
 583     __m128i          ifour       = _mm_set1_epi32(4);
 584     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 585     real             *vftab;
 586     __m128           dummy_mask,cutoff_mask;
 587     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 588     __m128           one     = _mm_set1_ps(1.0);
 589     __m128           two     = _mm_set1_ps(2.0);
 590     x                = xx[0];
 591     f                = ff[0];
 592
 593     nri              = nlist->nri;
 594     iinr             = nlist->iinr;
 595     jindex           = nlist->jindex;
 596     jjnr             = nlist->jjnr;
 597     shiftidx         = nlist->shift;
 598     gid              = nlist->gid;
 599     shiftvec         = fr->shift_vec[0];
 600     fshift           = fr->fshift[0];
 601     facel            = _mm_set1_ps(fr->epsfac);
 602     charge           = mdatoms->chargeA;
 603     nvdwtype         = fr->ntype;
 604     vdwparam         = fr->nbfp;
 605     vdwtype          = mdatoms->typeA;
 606
 607     vftab            = kernel_data->table_vdw->data;
 608     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 609
 610     invsqrta         = fr->invsqrta;
 611     dvda             = fr->dvda;
 612     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 613     gbtab            = fr->gbtab.data;
 614     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 615
 616     /* Avoid stupid compiler warnings */
 617     jnrA = jnrB = jnrC = jnrD = 0;
 618     j_coord_offsetA = 0;
 619     j_coord_offsetB = 0;
 620     j_coord_offsetC = 0;
 621     j_coord_offsetD = 0;
 622
 623     outeriter        = 0;
 624     inneriter        = 0;
 625
 626     for(iidx=0;iidx<4*DIM;iidx++)
 627     {
 628         scratch[iidx] = 0.0;
 629     }
 630
 631     /* Start outer loop over neighborlists */
 632     for(iidx=0; iidx<nri; iidx++)
 633     {
 634         /* Load shift vector for this list */
 635         i_shift_offset   = DIM*shiftidx[iidx];
 636
 637         /* Load limits for loop over neighbors */
 638         j_index_start    = jindex[iidx];
 639         j_index_end      = jindex[iidx+1];
 640
 641         /* Get outer coordinate index */
 642         inr              = iinr[iidx];
 643         i_coord_offset   = DIM*inr;
 644
 645         /* Load i particle coords and add shift vector */
 646         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 647
 648         fix0             = _mm_setzero_ps();
 649         fiy0             = _mm_setzero_ps();
 650         fiz0             = _mm_setzero_ps();
 651
 652         /* Load parameters for i particles */
 653         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 654         isai0            = _mm_load1_ps(invsqrta+inr+0);
 655         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 656
 657         dvdasum          = _mm_setzero_ps();
 658
 659         /* Start inner kernel loop */
 660         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 661         {
 662
 663             /* Get j neighbor index, and coordinate index */
 664             jnrA             = jjnr[jidx];
 665             jnrB             = jjnr[jidx+1];
 666             jnrC             = jjnr[jidx+2];
 667             jnrD             = jjnr[jidx+3];
 668             j_coord_offsetA  = DIM*jnrA;
 669             j_coord_offsetB  = DIM*jnrB;
 670             j_coord_offsetC  = DIM*jnrC;
 671             j_coord_offsetD  = DIM*jnrD;
 672
 673             /* load j atom coordinates */
 674             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 675                                               x+j_coord_offsetC,x+j_coord_offsetD,
 676                                               &jx0,&jy0,&jz0);
 677
 678             /* Calculate displacement vector */
 679             dx00             = _mm_sub_ps(ix0,jx0);
 680             dy00             = _mm_sub_ps(iy0,jy0);
 681             dz00             = _mm_sub_ps(iz0,jz0);
 682
 683             /* Calculate squared distance and things based on it */
 684             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 685
 686             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 687
 688             /* Load parameters for j particles */
 689             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 690                                                               charge+jnrC+0,charge+jnrD+0);
 691             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 692                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 693             vdwjidx0A        = 2*vdwtype[jnrA+0];
 694             vdwjidx0B        = 2*vdwtype[jnrB+0];
 695             vdwjidx0C        = 2*vdwtype[jnrC+0];
 696             vdwjidx0D        = 2*vdwtype[jnrD+0];
 697
 698             /**************************
 699              * CALCULATE INTERACTIONS *
 700              **************************/
 701
 702             r00              = _mm_mul_ps(rsq00,rinv00);
 703
 704             /* Compute parameters for interactions between i and j atoms */
 705             qq00             = _mm_mul_ps(iq0,jq0);
 706             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 707                                          vdwparam+vdwioffset0+vdwjidx0B,
 708                                          vdwparam+vdwioffset0+vdwjidx0C,
 709                                          vdwparam+vdwioffset0+vdwjidx0D,
 710                                          &c6_00,&c12_00);
 711
 712             /* Calculate table index by multiplying r with table scale and truncate to integer */
 713             rt               = _mm_mul_ps(r00,vftabscale);
 714             vfitab           = _mm_cvttps_epi32(rt);
 715 #ifdef __XOP__
 716             vfeps            = _mm_frcz_ps(rt);
 717 #else
 718             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 719 #endif
 720             twovfeps         = _mm_add_ps(vfeps,vfeps);
 721             vfitab           = _mm_slli_epi32(vfitab,3);
 722
 723             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 724             isaprod          = _mm_mul_ps(isai0,isaj0);
 725             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 726             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 727
 728             /* Calculate generalized born table index - this is a separate table from the normal one,
 729              * but we use the same procedure by multiplying r with scale and truncating to integer.
 730              */
 731             rt               = _mm_mul_ps(r00,gbscale);
 732             gbitab           = _mm_cvttps_epi32(rt);
 733 #ifdef __XOP__
 734             gbeps            = _mm_frcz_ps(rt);
 735 #else
 736             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 737 #endif
 738             gbitab           = _mm_slli_epi32(gbitab,2);
 739
 740             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 741             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 742             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 743             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 744             _MM_TRANSPOSE4_PS(Y,F,G,H);
 745             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 746             VV               = _mm_macc_ps(gbeps,Fp,Y);
 747             vgb              = _mm_mul_ps(gbqqfactor,VV);
 748
 749             twogbeps         = _mm_add_ps(gbeps,gbeps);
 750             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 751             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 752             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 753             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 754             fjptrA           = dvda+jnrA;
 755             fjptrB           = dvda+jnrB;
 756             fjptrC           = dvda+jnrC;
 757             fjptrD           = dvda+jnrD;
 758             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 759             velec            = _mm_mul_ps(qq00,rinv00);
 760             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 761
 762             /* CUBIC SPLINE TABLE DISPERSION */
 763             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 764             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 765             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 766             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 767             _MM_TRANSPOSE4_PS(Y,F,G,H);
 768             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 769             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 770             fvdw6            = _mm_mul_ps(c6_00,FF);
 771
 772             /* CUBIC SPLINE TABLE REPULSION */
 773             vfitab           = _mm_add_epi32(vfitab,ifour);
 774             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 775             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 776             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 777             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 778             _MM_TRANSPOSE4_PS(Y,F,G,H);
 779             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 780             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 781             fvdw12           = _mm_mul_ps(c12_00,FF);
 782             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 783
 784             fscal            = _mm_add_ps(felec,fvdw);
 785
 786              /* Update vectorial force */
 787             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 788             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 789             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 790
 791             fjptrA             = f+j_coord_offsetA;
 792             fjptrB             = f+j_coord_offsetB;
 793             fjptrC             = f+j_coord_offsetC;
 794             fjptrD             = f+j_coord_offsetD;
 795             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 796                                                    _mm_mul_ps(dx00,fscal),
 797                                                    _mm_mul_ps(dy00,fscal),
 798                                                    _mm_mul_ps(dz00,fscal));
 799
 800             /* Inner loop uses 85 flops */
 801         }
 802
 803         if(jidx<j_index_end)
 804         {
 805
 806             /* Get j neighbor index, and coordinate index */
 807             jnrlistA         = jjnr[jidx];
 808             jnrlistB         = jjnr[jidx+1];
 809             jnrlistC         = jjnr[jidx+2];
 810             jnrlistD         = jjnr[jidx+3];
 811             /* Sign of each element will be negative for non-real atoms.
 812              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 813              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 814              */
 815             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 816             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 817             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 818             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 819             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 820             j_coord_offsetA  = DIM*jnrA;
 821             j_coord_offsetB  = DIM*jnrB;
 822             j_coord_offsetC  = DIM*jnrC;
 823             j_coord_offsetD  = DIM*jnrD;
 824
 825             /* load j atom coordinates */
 826             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 827                                               x+j_coord_offsetC,x+j_coord_offsetD,
 828                                               &jx0,&jy0,&jz0);
 829
 830             /* Calculate displacement vector */
 831             dx00             = _mm_sub_ps(ix0,jx0);
 832             dy00             = _mm_sub_ps(iy0,jy0);
 833             dz00             = _mm_sub_ps(iz0,jz0);
 834
 835             /* Calculate squared distance and things based on it */
 836             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 837
 838             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 839
 840             /* Load parameters for j particles */
 841             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 842                                                               charge+jnrC+0,charge+jnrD+0);
 843             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 844                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 845             vdwjidx0A        = 2*vdwtype[jnrA+0];
 846             vdwjidx0B        = 2*vdwtype[jnrB+0];
 847             vdwjidx0C        = 2*vdwtype[jnrC+0];
 848             vdwjidx0D        = 2*vdwtype[jnrD+0];
 849
 850             /**************************
 851              * CALCULATE INTERACTIONS *
 852              **************************/
 853
 854             r00              = _mm_mul_ps(rsq00,rinv00);
 855             r00              = _mm_andnot_ps(dummy_mask,r00);
 856
 857             /* Compute parameters for interactions between i and j atoms */
 858             qq00             = _mm_mul_ps(iq0,jq0);
 859             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 860                                          vdwparam+vdwioffset0+vdwjidx0B,
 861                                          vdwparam+vdwioffset0+vdwjidx0C,
 862                                          vdwparam+vdwioffset0+vdwjidx0D,
 863                                          &c6_00,&c12_00);
 864
 865             /* Calculate table index by multiplying r with table scale and truncate to integer */
 866             rt               = _mm_mul_ps(r00,vftabscale);
 867             vfitab           = _mm_cvttps_epi32(rt);
 868 #ifdef __XOP__
 869             vfeps            = _mm_frcz_ps(rt);
 870 #else
 871             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 872 #endif
 873             twovfeps         = _mm_add_ps(vfeps,vfeps);
 874             vfitab           = _mm_slli_epi32(vfitab,3);
 875
 876             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 877             isaprod          = _mm_mul_ps(isai0,isaj0);
 878             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 879             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 880
 881             /* Calculate generalized born table index - this is a separate table from the normal one,
 882              * but we use the same procedure by multiplying r with scale and truncating to integer.
 883              */
 884             rt               = _mm_mul_ps(r00,gbscale);
 885             gbitab           = _mm_cvttps_epi32(rt);
 886 #ifdef __XOP__
 887             gbeps            = _mm_frcz_ps(rt);
 888 #else
 889             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 890 #endif
 891             gbitab           = _mm_slli_epi32(gbitab,2);
 892
 893             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 894             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 895             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 896             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 897             _MM_TRANSPOSE4_PS(Y,F,G,H);
 898             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 899             VV               = _mm_macc_ps(gbeps,Fp,Y);
 900             vgb              = _mm_mul_ps(gbqqfactor,VV);
 901
 902             twogbeps         = _mm_add_ps(gbeps,gbeps);
 903             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 904             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 905             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 906             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 907             /* 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. */
 908             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 909             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 910             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 911             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 912             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 913             velec            = _mm_mul_ps(qq00,rinv00);
 914             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 915
 916             /* CUBIC SPLINE TABLE DISPERSION */
 917             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 918             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 919             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 920             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 921             _MM_TRANSPOSE4_PS(Y,F,G,H);
 922             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 923             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 924             fvdw6            = _mm_mul_ps(c6_00,FF);
 925
 926             /* CUBIC SPLINE TABLE REPULSION */
 927             vfitab           = _mm_add_epi32(vfitab,ifour);
 928             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 929             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 930             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 931             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 932             _MM_TRANSPOSE4_PS(Y,F,G,H);
 933             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 934             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 935             fvdw12           = _mm_mul_ps(c12_00,FF);
 936             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 937
 938             fscal            = _mm_add_ps(felec,fvdw);
 939
 940             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 941
 942              /* Update vectorial force */
 943             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 944             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 945             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 946
 947             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 948             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 949             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 950             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 951             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 952                                                    _mm_mul_ps(dx00,fscal),
 953                                                    _mm_mul_ps(dy00,fscal),
 954                                                    _mm_mul_ps(dz00,fscal));
 955
 956             /* Inner loop uses 86 flops */
 957         }
 958
 959         /* End of innermost loop */
 960
 961         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 962                                               f+i_coord_offset,fshift+i_shift_offset);
 963
 964         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 965         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 966
 967         /* Increment number of inner iterations */
 968         inneriter                  += j_index_end - j_index_start;
 969
 970         /* Outer loop uses 7 flops */
 971     }
 972
 973     /* Increment number of outer iterations */
 974     outeriter        += nri;
 975
 976     /* Update outer/inner flops */
 977
 978     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*86);
 979 }