src/gromacs/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) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  85     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128i          gbitab;
  90     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  91     __m128           minushalf = _mm_set1_ps(-0.5);
  92     real             *invsqrta,*dvda,*gbtab;
  93     int              nvdwtype;
  94     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  98     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  99     __m128i          vfitab;
 100     __m128i          ifour       = _mm_set1_epi32(4);
 101     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 102     real             *vftab;
 103     __m128           dummy_mask,cutoff_mask;
 104     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 105     __m128           one     = _mm_set1_ps(1.0);
 106     __m128           two     = _mm_set1_ps(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_ps(fr->ic->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     vftab            = kernel_data->table_vdw->data;
 125     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 126
 127     invsqrta         = fr->invsqrta;
 128     dvda             = fr->dvda;
 129     gbtabscale       = _mm_set1_ps(fr->gbtab->scale);
 130     gbtab            = fr->gbtab->data;
 131     gbinvepsdiff     = _mm_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 132
 133     /* Avoid stupid compiler warnings */
 134     jnrA = jnrB = jnrC = jnrD = 0;
 135     j_coord_offsetA = 0;
 136     j_coord_offsetB = 0;
 137     j_coord_offsetC = 0;
 138     j_coord_offsetD = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     for(iidx=0;iidx<4*DIM;iidx++)
 144     {
 145         scratch[iidx] = 0.0;
 146     }
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 164
 165         fix0             = _mm_setzero_ps();
 166         fiy0             = _mm_setzero_ps();
 167         fiz0             = _mm_setzero_ps();
 168
 169         /* Load parameters for i particles */
 170         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 171         isai0            = _mm_load1_ps(invsqrta+inr+0);
 172         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_ps();
 176         vgbsum           = _mm_setzero_ps();
 177         vvdwsum          = _mm_setzero_ps();
 178         dvdasum          = _mm_setzero_ps();
 179
 180         /* Start inner kernel loop */
 181         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 182         {
 183
 184             /* Get j neighbor index, and coordinate index */
 185             jnrA             = jjnr[jidx];
 186             jnrB             = jjnr[jidx+1];
 187             jnrC             = jjnr[jidx+2];
 188             jnrD             = jjnr[jidx+3];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193
 194             /* load j atom coordinates */
 195             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 196                                               x+j_coord_offsetC,x+j_coord_offsetD,
 197                                               &jx0,&jy0,&jz0);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _mm_sub_ps(ix0,jx0);
 201             dy00             = _mm_sub_ps(iy0,jy0);
 202             dz00             = _mm_sub_ps(iz0,jz0);
 203
 204             /* Calculate squared distance and things based on it */
 205             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 206
 207             rinv00           = avx128fma_invsqrt_f(rsq00);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 211                                                               charge+jnrC+0,charge+jnrD+0);
 212             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 213                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 214             vdwjidx0A        = 2*vdwtype[jnrA+0];
 215             vdwjidx0B        = 2*vdwtype[jnrB+0];
 216             vdwjidx0C        = 2*vdwtype[jnrC+0];
 217             vdwjidx0D        = 2*vdwtype[jnrD+0];
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             r00              = _mm_mul_ps(rsq00,rinv00);
 224
 225             /* Compute parameters for interactions between i and j atoms */
 226             qq00             = _mm_mul_ps(iq0,jq0);
 227             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 228                                          vdwparam+vdwioffset0+vdwjidx0B,
 229                                          vdwparam+vdwioffset0+vdwjidx0C,
 230                                          vdwparam+vdwioffset0+vdwjidx0D,
 231                                          &c6_00,&c12_00);
 232
 233             /* Calculate table index by multiplying r with table scale and truncate to integer */
 234             rt               = _mm_mul_ps(r00,vftabscale);
 235             vfitab           = _mm_cvttps_epi32(rt);
 236 #ifdef __XOP__
 237             vfeps            = _mm_frcz_ps(rt);
 238 #else
 239             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 240 #endif
 241             twovfeps         = _mm_add_ps(vfeps,vfeps);
 242             vfitab           = _mm_slli_epi32(vfitab,3);
 243
 244             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 245             isaprod          = _mm_mul_ps(isai0,isaj0);
 246             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 247             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 248
 249             /* Calculate generalized born table index - this is a separate table from the normal one,
 250              * but we use the same procedure by multiplying r with scale and truncating to integer.
 251              */
 252             rt               = _mm_mul_ps(r00,gbscale);
 253             gbitab           = _mm_cvttps_epi32(rt);
 254 #ifdef __XOP__
 255             gbeps            = _mm_frcz_ps(rt);
 256 #else
 257             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 258 #endif
 259             gbitab           = _mm_slli_epi32(gbitab,2);
 260
 261             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 262             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 263             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 264             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 265             _MM_TRANSPOSE4_PS(Y,F,G,H);
 266             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 267             VV               = _mm_macc_ps(gbeps,Fp,Y);
 268             vgb              = _mm_mul_ps(gbqqfactor,VV);
 269
 270             twogbeps         = _mm_add_ps(gbeps,gbeps);
 271             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 272             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 273             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 274             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 275             fjptrA           = dvda+jnrA;
 276             fjptrB           = dvda+jnrB;
 277             fjptrC           = dvda+jnrC;
 278             fjptrD           = dvda+jnrD;
 279             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 280             velec            = _mm_mul_ps(qq00,rinv00);
 281             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 282
 283             /* CUBIC SPLINE TABLE DISPERSION */
 284             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 285             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 286             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 287             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 288             _MM_TRANSPOSE4_PS(Y,F,G,H);
 289             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 290             VV               = _mm_macc_ps(vfeps,Fp,Y);
 291             vvdw6            = _mm_mul_ps(c6_00,VV);
 292             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 293             fvdw6            = _mm_mul_ps(c6_00,FF);
 294
 295             /* CUBIC SPLINE TABLE REPULSION */
 296             vfitab           = _mm_add_epi32(vfitab,ifour);
 297             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 298             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 299             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 300             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 301             _MM_TRANSPOSE4_PS(Y,F,G,H);
 302             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 303             VV               = _mm_macc_ps(vfeps,Fp,Y);
 304             vvdw12           = _mm_mul_ps(c12_00,VV);
 305             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 306             fvdw12           = _mm_mul_ps(c12_00,FF);
 307             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 308             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velecsum         = _mm_add_ps(velecsum,velec);
 312             vgbsum           = _mm_add_ps(vgbsum,vgb);
 313             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 314
 315             fscal            = _mm_add_ps(felec,fvdw);
 316
 317              /* Update vectorial force */
 318             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 319             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 320             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 321
 322             fjptrA             = f+j_coord_offsetA;
 323             fjptrB             = f+j_coord_offsetB;
 324             fjptrC             = f+j_coord_offsetC;
 325             fjptrD             = f+j_coord_offsetD;
 326             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 327                                                    _mm_mul_ps(dx00,fscal),
 328                                                    _mm_mul_ps(dy00,fscal),
 329                                                    _mm_mul_ps(dz00,fscal));
 330
 331             /* Inner loop uses 95 flops */
 332         }
 333
 334         if(jidx<j_index_end)
 335         {
 336
 337             /* Get j neighbor index, and coordinate index */
 338             jnrlistA         = jjnr[jidx];
 339             jnrlistB         = jjnr[jidx+1];
 340             jnrlistC         = jjnr[jidx+2];
 341             jnrlistD         = jjnr[jidx+3];
 342             /* Sign of each element will be negative for non-real atoms.
 343              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 344              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 345              */
 346             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 347             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 348             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 349             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 350             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 351             j_coord_offsetA  = DIM*jnrA;
 352             j_coord_offsetB  = DIM*jnrB;
 353             j_coord_offsetC  = DIM*jnrC;
 354             j_coord_offsetD  = DIM*jnrD;
 355
 356             /* load j atom coordinates */
 357             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 358                                               x+j_coord_offsetC,x+j_coord_offsetD,
 359                                               &jx0,&jy0,&jz0);
 360
 361             /* Calculate displacement vector */
 362             dx00             = _mm_sub_ps(ix0,jx0);
 363             dy00             = _mm_sub_ps(iy0,jy0);
 364             dz00             = _mm_sub_ps(iz0,jz0);
 365
 366             /* Calculate squared distance and things based on it */
 367             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 368
 369             rinv00           = avx128fma_invsqrt_f(rsq00);
 370
 371             /* Load parameters for j particles */
 372             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 373                                                               charge+jnrC+0,charge+jnrD+0);
 374             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 375                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 376             vdwjidx0A        = 2*vdwtype[jnrA+0];
 377             vdwjidx0B        = 2*vdwtype[jnrB+0];
 378             vdwjidx0C        = 2*vdwtype[jnrC+0];
 379             vdwjidx0D        = 2*vdwtype[jnrD+0];
 380
 381             /**************************
 382              * CALCULATE INTERACTIONS *
 383              **************************/
 384
 385             r00              = _mm_mul_ps(rsq00,rinv00);
 386             r00              = _mm_andnot_ps(dummy_mask,r00);
 387
 388             /* Compute parameters for interactions between i and j atoms */
 389             qq00             = _mm_mul_ps(iq0,jq0);
 390             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 391                                          vdwparam+vdwioffset0+vdwjidx0B,
 392                                          vdwparam+vdwioffset0+vdwjidx0C,
 393                                          vdwparam+vdwioffset0+vdwjidx0D,
 394                                          &c6_00,&c12_00);
 395
 396             /* Calculate table index by multiplying r with table scale and truncate to integer */
 397             rt               = _mm_mul_ps(r00,vftabscale);
 398             vfitab           = _mm_cvttps_epi32(rt);
 399 #ifdef __XOP__
 400             vfeps            = _mm_frcz_ps(rt);
 401 #else
 402             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 403 #endif
 404             twovfeps         = _mm_add_ps(vfeps,vfeps);
 405             vfitab           = _mm_slli_epi32(vfitab,3);
 406
 407             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 408             isaprod          = _mm_mul_ps(isai0,isaj0);
 409             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 410             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 411
 412             /* Calculate generalized born table index - this is a separate table from the normal one,
 413              * but we use the same procedure by multiplying r with scale and truncating to integer.
 414              */
 415             rt               = _mm_mul_ps(r00,gbscale);
 416             gbitab           = _mm_cvttps_epi32(rt);
 417 #ifdef __XOP__
 418             gbeps            = _mm_frcz_ps(rt);
 419 #else
 420             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 421 #endif
 422             gbitab           = _mm_slli_epi32(gbitab,2);
 423
 424             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 425             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 426             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 427             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 428             _MM_TRANSPOSE4_PS(Y,F,G,H);
 429             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 430             VV               = _mm_macc_ps(gbeps,Fp,Y);
 431             vgb              = _mm_mul_ps(gbqqfactor,VV);
 432
 433             twogbeps         = _mm_add_ps(gbeps,gbeps);
 434             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 435             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 436             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 437             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 438             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 439             /* 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. */
 440             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 441             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 442             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 443             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 444             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 445             velec            = _mm_mul_ps(qq00,rinv00);
 446             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 447
 448             /* CUBIC SPLINE TABLE DISPERSION */
 449             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 450             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 451             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 452             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 453             _MM_TRANSPOSE4_PS(Y,F,G,H);
 454             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 455             VV               = _mm_macc_ps(vfeps,Fp,Y);
 456             vvdw6            = _mm_mul_ps(c6_00,VV);
 457             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 458             fvdw6            = _mm_mul_ps(c6_00,FF);
 459
 460             /* CUBIC SPLINE TABLE REPULSION */
 461             vfitab           = _mm_add_epi32(vfitab,ifour);
 462             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 463             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 464             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 465             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 466             _MM_TRANSPOSE4_PS(Y,F,G,H);
 467             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 468             VV               = _mm_macc_ps(vfeps,Fp,Y);
 469             vvdw12           = _mm_mul_ps(c12_00,VV);
 470             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 471             fvdw12           = _mm_mul_ps(c12_00,FF);
 472             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 473             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velec            = _mm_andnot_ps(dummy_mask,velec);
 477             velecsum         = _mm_add_ps(velecsum,velec);
 478             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 479             vgbsum           = _mm_add_ps(vgbsum,vgb);
 480             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 481             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 482
 483             fscal            = _mm_add_ps(felec,fvdw);
 484
 485             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 486
 487              /* Update vectorial force */
 488             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 489             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 490             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 491
 492             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 493             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 494             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 495             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 496             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 497                                                    _mm_mul_ps(dx00,fscal),
 498                                                    _mm_mul_ps(dy00,fscal),
 499                                                    _mm_mul_ps(dz00,fscal));
 500
 501             /* Inner loop uses 96 flops */
 502         }
 503
 504         /* End of innermost loop */
 505
 506         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 507                                               f+i_coord_offset,fshift+i_shift_offset);
 508
 509         ggid                        = gid[iidx];
 510         /* Update potential energies */
 511         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 512         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 513         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 514         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 515         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 516
 517         /* Increment number of inner iterations */
 518         inneriter                  += j_index_end - j_index_start;
 519
 520         /* Outer loop uses 10 flops */
 521     }
 522
 523     /* Increment number of outer iterations */
 524     outeriter        += nri;
 525
 526     /* Update outer/inner flops */
 527
 528     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*96);
 529 }
 530 /*
 531  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 532  * Electrostatics interaction: GeneralizedBorn
 533  * VdW interaction:            CubicSplineTable
 534  * Geometry:                   Particle-Particle
 535  * Calculate force/pot:        Force
 536  */
 537 void
 538 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 539                     (t_nblist                    * gmx_restrict       nlist,
 540                      rvec                        * gmx_restrict          xx,
 541                      rvec                        * gmx_restrict          ff,
 542                      struct t_forcerec           * gmx_restrict          fr,
 543                      t_mdatoms                   * gmx_restrict     mdatoms,
 544                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 545                      t_nrnb                      * gmx_restrict        nrnb)
 546 {
 547     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 548      * just 0 for non-waters.
 549      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 550      * jnr indices corresponding to data put in the four positions in the SIMD register.
 551      */
 552     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 553     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 554     int              jnrA,jnrB,jnrC,jnrD;
 555     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 556     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 557     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 558     real             rcutoff_scalar;
 559     real             *shiftvec,*fshift,*x,*f;
 560     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 561     real             scratch[4*DIM];
 562     __m128           fscal,rcutoff,rcutoff2,jidxall;
 563     int              vdwioffset0;
 564     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 565     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 566     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 567     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 568     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 569     real             *charge;
 570     __m128i          gbitab;
 571     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 572     __m128           minushalf = _mm_set1_ps(-0.5);
 573     real             *invsqrta,*dvda,*gbtab;
 574     int              nvdwtype;
 575     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 576     int              *vdwtype;
 577     real             *vdwparam;
 578     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 579     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 580     __m128i          vfitab;
 581     __m128i          ifour       = _mm_set1_epi32(4);
 582     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 583     real             *vftab;
 584     __m128           dummy_mask,cutoff_mask;
 585     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 586     __m128           one     = _mm_set1_ps(1.0);
 587     __m128           two     = _mm_set1_ps(2.0);
 588     x                = xx[0];
 589     f                = ff[0];
 590
 591     nri              = nlist->nri;
 592     iinr             = nlist->iinr;
 593     jindex           = nlist->jindex;
 594     jjnr             = nlist->jjnr;
 595     shiftidx         = nlist->shift;
 596     gid              = nlist->gid;
 597     shiftvec         = fr->shift_vec[0];
 598     fshift           = fr->fshift[0];
 599     facel            = _mm_set1_ps(fr->ic->epsfac);
 600     charge           = mdatoms->chargeA;
 601     nvdwtype         = fr->ntype;
 602     vdwparam         = fr->nbfp;
 603     vdwtype          = mdatoms->typeA;
 604
 605     vftab            = kernel_data->table_vdw->data;
 606     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 607
 608     invsqrta         = fr->invsqrta;
 609     dvda             = fr->dvda;
 610     gbtabscale       = _mm_set1_ps(fr->gbtab->scale);
 611     gbtab            = fr->gbtab->data;
 612     gbinvepsdiff     = _mm_set1_ps((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 613
 614     /* Avoid stupid compiler warnings */
 615     jnrA = jnrB = jnrC = jnrD = 0;
 616     j_coord_offsetA = 0;
 617     j_coord_offsetB = 0;
 618     j_coord_offsetC = 0;
 619     j_coord_offsetD = 0;
 620
 621     outeriter        = 0;
 622     inneriter        = 0;
 623
 624     for(iidx=0;iidx<4*DIM;iidx++)
 625     {
 626         scratch[iidx] = 0.0;
 627     }
 628
 629     /* Start outer loop over neighborlists */
 630     for(iidx=0; iidx<nri; iidx++)
 631     {
 632         /* Load shift vector for this list */
 633         i_shift_offset   = DIM*shiftidx[iidx];
 634
 635         /* Load limits for loop over neighbors */
 636         j_index_start    = jindex[iidx];
 637         j_index_end      = jindex[iidx+1];
 638
 639         /* Get outer coordinate index */
 640         inr              = iinr[iidx];
 641         i_coord_offset   = DIM*inr;
 642
 643         /* Load i particle coords and add shift vector */
 644         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 645
 646         fix0             = _mm_setzero_ps();
 647         fiy0             = _mm_setzero_ps();
 648         fiz0             = _mm_setzero_ps();
 649
 650         /* Load parameters for i particles */
 651         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 652         isai0            = _mm_load1_ps(invsqrta+inr+0);
 653         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 654
 655         dvdasum          = _mm_setzero_ps();
 656
 657         /* Start inner kernel loop */
 658         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 659         {
 660
 661             /* Get j neighbor index, and coordinate index */
 662             jnrA             = jjnr[jidx];
 663             jnrB             = jjnr[jidx+1];
 664             jnrC             = jjnr[jidx+2];
 665             jnrD             = jjnr[jidx+3];
 666             j_coord_offsetA  = DIM*jnrA;
 667             j_coord_offsetB  = DIM*jnrB;
 668             j_coord_offsetC  = DIM*jnrC;
 669             j_coord_offsetD  = DIM*jnrD;
 670
 671             /* load j atom coordinates */
 672             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 673                                               x+j_coord_offsetC,x+j_coord_offsetD,
 674                                               &jx0,&jy0,&jz0);
 675
 676             /* Calculate displacement vector */
 677             dx00             = _mm_sub_ps(ix0,jx0);
 678             dy00             = _mm_sub_ps(iy0,jy0);
 679             dz00             = _mm_sub_ps(iz0,jz0);
 680
 681             /* Calculate squared distance and things based on it */
 682             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 683
 684             rinv00           = avx128fma_invsqrt_f(rsq00);
 685
 686             /* Load parameters for j particles */
 687             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 688                                                               charge+jnrC+0,charge+jnrD+0);
 689             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 690                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 691             vdwjidx0A        = 2*vdwtype[jnrA+0];
 692             vdwjidx0B        = 2*vdwtype[jnrB+0];
 693             vdwjidx0C        = 2*vdwtype[jnrC+0];
 694             vdwjidx0D        = 2*vdwtype[jnrD+0];
 695
 696             /**************************
 697              * CALCULATE INTERACTIONS *
 698              **************************/
 699
 700             r00              = _mm_mul_ps(rsq00,rinv00);
 701
 702             /* Compute parameters for interactions between i and j atoms */
 703             qq00             = _mm_mul_ps(iq0,jq0);
 704             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 705                                          vdwparam+vdwioffset0+vdwjidx0B,
 706                                          vdwparam+vdwioffset0+vdwjidx0C,
 707                                          vdwparam+vdwioffset0+vdwjidx0D,
 708                                          &c6_00,&c12_00);
 709
 710             /* Calculate table index by multiplying r with table scale and truncate to integer */
 711             rt               = _mm_mul_ps(r00,vftabscale);
 712             vfitab           = _mm_cvttps_epi32(rt);
 713 #ifdef __XOP__
 714             vfeps            = _mm_frcz_ps(rt);
 715 #else
 716             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 717 #endif
 718             twovfeps         = _mm_add_ps(vfeps,vfeps);
 719             vfitab           = _mm_slli_epi32(vfitab,3);
 720
 721             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 722             isaprod          = _mm_mul_ps(isai0,isaj0);
 723             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 724             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 725
 726             /* Calculate generalized born table index - this is a separate table from the normal one,
 727              * but we use the same procedure by multiplying r with scale and truncating to integer.
 728              */
 729             rt               = _mm_mul_ps(r00,gbscale);
 730             gbitab           = _mm_cvttps_epi32(rt);
 731 #ifdef __XOP__
 732             gbeps            = _mm_frcz_ps(rt);
 733 #else
 734             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 735 #endif
 736             gbitab           = _mm_slli_epi32(gbitab,2);
 737
 738             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 739             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 740             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 741             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 742             _MM_TRANSPOSE4_PS(Y,F,G,H);
 743             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 744             VV               = _mm_macc_ps(gbeps,Fp,Y);
 745             vgb              = _mm_mul_ps(gbqqfactor,VV);
 746
 747             twogbeps         = _mm_add_ps(gbeps,gbeps);
 748             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 749             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 750             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 751             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 752             fjptrA           = dvda+jnrA;
 753             fjptrB           = dvda+jnrB;
 754             fjptrC           = dvda+jnrC;
 755             fjptrD           = dvda+jnrD;
 756             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 757             velec            = _mm_mul_ps(qq00,rinv00);
 758             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 759
 760             /* CUBIC SPLINE TABLE DISPERSION */
 761             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 762             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 763             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 764             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 765             _MM_TRANSPOSE4_PS(Y,F,G,H);
 766             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 767             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 768             fvdw6            = _mm_mul_ps(c6_00,FF);
 769
 770             /* CUBIC SPLINE TABLE REPULSION */
 771             vfitab           = _mm_add_epi32(vfitab,ifour);
 772             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 773             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 774             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 775             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 776             _MM_TRANSPOSE4_PS(Y,F,G,H);
 777             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 778             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 779             fvdw12           = _mm_mul_ps(c12_00,FF);
 780             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 781
 782             fscal            = _mm_add_ps(felec,fvdw);
 783
 784              /* Update vectorial force */
 785             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 786             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 787             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 788
 789             fjptrA             = f+j_coord_offsetA;
 790             fjptrB             = f+j_coord_offsetB;
 791             fjptrC             = f+j_coord_offsetC;
 792             fjptrD             = f+j_coord_offsetD;
 793             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 794                                                    _mm_mul_ps(dx00,fscal),
 795                                                    _mm_mul_ps(dy00,fscal),
 796                                                    _mm_mul_ps(dz00,fscal));
 797
 798             /* Inner loop uses 85 flops */
 799         }
 800
 801         if(jidx<j_index_end)
 802         {
 803
 804             /* Get j neighbor index, and coordinate index */
 805             jnrlistA         = jjnr[jidx];
 806             jnrlistB         = jjnr[jidx+1];
 807             jnrlistC         = jjnr[jidx+2];
 808             jnrlistD         = jjnr[jidx+3];
 809             /* Sign of each element will be negative for non-real atoms.
 810              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 811              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 812              */
 813             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 814             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 815             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 816             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 817             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 818             j_coord_offsetA  = DIM*jnrA;
 819             j_coord_offsetB  = DIM*jnrB;
 820             j_coord_offsetC  = DIM*jnrC;
 821             j_coord_offsetD  = DIM*jnrD;
 822
 823             /* load j atom coordinates */
 824             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 825                                               x+j_coord_offsetC,x+j_coord_offsetD,
 826                                               &jx0,&jy0,&jz0);
 827
 828             /* Calculate displacement vector */
 829             dx00             = _mm_sub_ps(ix0,jx0);
 830             dy00             = _mm_sub_ps(iy0,jy0);
 831             dz00             = _mm_sub_ps(iz0,jz0);
 832
 833             /* Calculate squared distance and things based on it */
 834             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 835
 836             rinv00           = avx128fma_invsqrt_f(rsq00);
 837
 838             /* Load parameters for j particles */
 839             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 840                                                               charge+jnrC+0,charge+jnrD+0);
 841             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 842                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 843             vdwjidx0A        = 2*vdwtype[jnrA+0];
 844             vdwjidx0B        = 2*vdwtype[jnrB+0];
 845             vdwjidx0C        = 2*vdwtype[jnrC+0];
 846             vdwjidx0D        = 2*vdwtype[jnrD+0];
 847
 848             /**************************
 849              * CALCULATE INTERACTIONS *
 850              **************************/
 851
 852             r00              = _mm_mul_ps(rsq00,rinv00);
 853             r00              = _mm_andnot_ps(dummy_mask,r00);
 854
 855             /* Compute parameters for interactions between i and j atoms */
 856             qq00             = _mm_mul_ps(iq0,jq0);
 857             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 858                                          vdwparam+vdwioffset0+vdwjidx0B,
 859                                          vdwparam+vdwioffset0+vdwjidx0C,
 860                                          vdwparam+vdwioffset0+vdwjidx0D,
 861                                          &c6_00,&c12_00);
 862
 863             /* Calculate table index by multiplying r with table scale and truncate to integer */
 864             rt               = _mm_mul_ps(r00,vftabscale);
 865             vfitab           = _mm_cvttps_epi32(rt);
 866 #ifdef __XOP__
 867             vfeps            = _mm_frcz_ps(rt);
 868 #else
 869             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 870 #endif
 871             twovfeps         = _mm_add_ps(vfeps,vfeps);
 872             vfitab           = _mm_slli_epi32(vfitab,3);
 873
 874             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 875             isaprod          = _mm_mul_ps(isai0,isaj0);
 876             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 877             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 878
 879             /* Calculate generalized born table index - this is a separate table from the normal one,
 880              * but we use the same procedure by multiplying r with scale and truncating to integer.
 881              */
 882             rt               = _mm_mul_ps(r00,gbscale);
 883             gbitab           = _mm_cvttps_epi32(rt);
 884 #ifdef __XOP__
 885             gbeps            = _mm_frcz_ps(rt);
 886 #else
 887             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 888 #endif
 889             gbitab           = _mm_slli_epi32(gbitab,2);
 890
 891             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 892             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 893             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 894             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 895             _MM_TRANSPOSE4_PS(Y,F,G,H);
 896             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 897             VV               = _mm_macc_ps(gbeps,Fp,Y);
 898             vgb              = _mm_mul_ps(gbqqfactor,VV);
 899
 900             twogbeps         = _mm_add_ps(gbeps,gbeps);
 901             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 902             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 903             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 904             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 905             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 906             /* 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. */
 907             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 908             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 909             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 910             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 911             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 912             velec            = _mm_mul_ps(qq00,rinv00);
 913             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 914
 915             /* CUBIC SPLINE TABLE DISPERSION */
 916             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 917             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 918             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 919             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 920             _MM_TRANSPOSE4_PS(Y,F,G,H);
 921             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 922             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 923             fvdw6            = _mm_mul_ps(c6_00,FF);
 924
 925             /* CUBIC SPLINE TABLE REPULSION */
 926             vfitab           = _mm_add_epi32(vfitab,ifour);
 927             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 928             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 929             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 930             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 931             _MM_TRANSPOSE4_PS(Y,F,G,H);
 932             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 933             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 934             fvdw12           = _mm_mul_ps(c12_00,FF);
 935             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 936
 937             fscal            = _mm_add_ps(felec,fvdw);
 938
 939             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 940
 941              /* Update vectorial force */
 942             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 943             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 944             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 945
 946             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 947             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 948             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 949             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 950             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 951                                                    _mm_mul_ps(dx00,fscal),
 952                                                    _mm_mul_ps(dy00,fscal),
 953                                                    _mm_mul_ps(dz00,fscal));
 954
 955             /* Inner loop uses 86 flops */
 956         }
 957
 958         /* End of innermost loop */
 959
 960         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 961                                               f+i_coord_offset,fshift+i_shift_offset);
 962
 963         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 964         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 965
 966         /* Increment number of inner iterations */
 967         inneriter                  += j_index_end - j_index_start;
 968
 969         /* Outer loop uses 7 flops */
 970     }
 971
 972     /* Increment number of outer iterations */
 973     outeriter        += nri;
 974
 975     /* Update outer/inner flops */
 976
 977     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*86);
 978 }