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, 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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/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_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with 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             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 441             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 442             /* 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. */
 443             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 444             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 445             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 446             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 447             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 448             velec            = _mm_mul_ps(qq00,rinv00);
 449             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 450
 451             /* CUBIC SPLINE TABLE DISPERSION */
 452             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 453             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 454             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 455             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 456             _MM_TRANSPOSE4_PS(Y,F,G,H);
 457             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 458             VV               = _mm_macc_ps(vfeps,Fp,Y);
 459             vvdw6            = _mm_mul_ps(c6_00,VV);
 460             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 461             fvdw6            = _mm_mul_ps(c6_00,FF);
 462
 463             /* CUBIC SPLINE TABLE REPULSION */
 464             vfitab           = _mm_add_epi32(vfitab,ifour);
 465             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 466             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 467             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 468             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 469             _MM_TRANSPOSE4_PS(Y,F,G,H);
 470             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 471             VV               = _mm_macc_ps(vfeps,Fp,Y);
 472             vvdw12           = _mm_mul_ps(c12_00,VV);
 473             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 474             fvdw12           = _mm_mul_ps(c12_00,FF);
 475             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 476             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velec            = _mm_andnot_ps(dummy_mask,velec);
 480             velecsum         = _mm_add_ps(velecsum,velec);
 481             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 482             vgbsum           = _mm_add_ps(vgbsum,vgb);
 483             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 484             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 485
 486             fscal            = _mm_add_ps(felec,fvdw);
 487
 488             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 489
 490              /* Update vectorial force */
 491             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 492             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 493             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 494
 495             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 496             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 497             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 498             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 499             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 500                                                    _mm_mul_ps(dx00,fscal),
 501                                                    _mm_mul_ps(dy00,fscal),
 502                                                    _mm_mul_ps(dz00,fscal));
 503
 504             /* Inner loop uses 96 flops */
 505         }
 506
 507         /* End of innermost loop */
 508
 509         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 510                                               f+i_coord_offset,fshift+i_shift_offset);
 511
 512         ggid                        = gid[iidx];
 513         /* Update potential energies */
 514         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 515         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 516         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 517         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 518         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 519
 520         /* Increment number of inner iterations */
 521         inneriter                  += j_index_end - j_index_start;
 522
 523         /* Outer loop uses 10 flops */
 524     }
 525
 526     /* Increment number of outer iterations */
 527     outeriter        += nri;
 528
 529     /* Update outer/inner flops */
 530
 531     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*96);
 532 }
 533 /*
 534  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 535  * Electrostatics interaction: GeneralizedBorn
 536  * VdW interaction:            CubicSplineTable
 537  * Geometry:                   Particle-Particle
 538  * Calculate force/pot:        Force
 539  */
 540 void
 541 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 542                     (t_nblist                    * gmx_restrict       nlist,
 543                      rvec                        * gmx_restrict          xx,
 544                      rvec                        * gmx_restrict          ff,
 545                      t_forcerec                  * gmx_restrict          fr,
 546                      t_mdatoms                   * gmx_restrict     mdatoms,
 547                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 548                      t_nrnb                      * gmx_restrict        nrnb)
 549 {
 550     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 551      * just 0 for non-waters.
 552      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 553      * jnr indices corresponding to data put in the four positions in the SIMD register.
 554      */
 555     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 556     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 557     int              jnrA,jnrB,jnrC,jnrD;
 558     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 559     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 560     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 561     real             rcutoff_scalar;
 562     real             *shiftvec,*fshift,*x,*f;
 563     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 564     real             scratch[4*DIM];
 565     __m128           fscal,rcutoff,rcutoff2,jidxall;
 566     int              vdwioffset0;
 567     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 568     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 569     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 570     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 571     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 572     real             *charge;
 573     __m128i          gbitab;
 574     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 575     __m128           minushalf = _mm_set1_ps(-0.5);
 576     real             *invsqrta,*dvda,*gbtab;
 577     int              nvdwtype;
 578     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 579     int              *vdwtype;
 580     real             *vdwparam;
 581     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 582     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 583     __m128i          vfitab;
 584     __m128i          ifour       = _mm_set1_epi32(4);
 585     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 586     real             *vftab;
 587     __m128           dummy_mask,cutoff_mask;
 588     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 589     __m128           one     = _mm_set1_ps(1.0);
 590     __m128           two     = _mm_set1_ps(2.0);
 591     x                = xx[0];
 592     f                = ff[0];
 593
 594     nri              = nlist->nri;
 595     iinr             = nlist->iinr;
 596     jindex           = nlist->jindex;
 597     jjnr             = nlist->jjnr;
 598     shiftidx         = nlist->shift;
 599     gid              = nlist->gid;
 600     shiftvec         = fr->shift_vec[0];
 601     fshift           = fr->fshift[0];
 602     facel            = _mm_set1_ps(fr->epsfac);
 603     charge           = mdatoms->chargeA;
 604     nvdwtype         = fr->ntype;
 605     vdwparam         = fr->nbfp;
 606     vdwtype          = mdatoms->typeA;
 607
 608     vftab            = kernel_data->table_vdw->data;
 609     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 610
 611     invsqrta         = fr->invsqrta;
 612     dvda             = fr->dvda;
 613     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 614     gbtab            = fr->gbtab.data;
 615     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 616
 617     /* Avoid stupid compiler warnings */
 618     jnrA = jnrB = jnrC = jnrD = 0;
 619     j_coord_offsetA = 0;
 620     j_coord_offsetB = 0;
 621     j_coord_offsetC = 0;
 622     j_coord_offsetD = 0;
 623
 624     outeriter        = 0;
 625     inneriter        = 0;
 626
 627     for(iidx=0;iidx<4*DIM;iidx++)
 628     {
 629         scratch[iidx] = 0.0;
 630     }
 631
 632     /* Start outer loop over neighborlists */
 633     for(iidx=0; iidx<nri; iidx++)
 634     {
 635         /* Load shift vector for this list */
 636         i_shift_offset   = DIM*shiftidx[iidx];
 637
 638         /* Load limits for loop over neighbors */
 639         j_index_start    = jindex[iidx];
 640         j_index_end      = jindex[iidx+1];
 641
 642         /* Get outer coordinate index */
 643         inr              = iinr[iidx];
 644         i_coord_offset   = DIM*inr;
 645
 646         /* Load i particle coords and add shift vector */
 647         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 648
 649         fix0             = _mm_setzero_ps();
 650         fiy0             = _mm_setzero_ps();
 651         fiz0             = _mm_setzero_ps();
 652
 653         /* Load parameters for i particles */
 654         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 655         isai0            = _mm_load1_ps(invsqrta+inr+0);
 656         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 657
 658         dvdasum          = _mm_setzero_ps();
 659
 660         /* Start inner kernel loop */
 661         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 662         {
 663
 664             /* Get j neighbor index, and coordinate index */
 665             jnrA             = jjnr[jidx];
 666             jnrB             = jjnr[jidx+1];
 667             jnrC             = jjnr[jidx+2];
 668             jnrD             = jjnr[jidx+3];
 669             j_coord_offsetA  = DIM*jnrA;
 670             j_coord_offsetB  = DIM*jnrB;
 671             j_coord_offsetC  = DIM*jnrC;
 672             j_coord_offsetD  = DIM*jnrD;
 673
 674             /* load j atom coordinates */
 675             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 676                                               x+j_coord_offsetC,x+j_coord_offsetD,
 677                                               &jx0,&jy0,&jz0);
 678
 679             /* Calculate displacement vector */
 680             dx00             = _mm_sub_ps(ix0,jx0);
 681             dy00             = _mm_sub_ps(iy0,jy0);
 682             dz00             = _mm_sub_ps(iz0,jz0);
 683
 684             /* Calculate squared distance and things based on it */
 685             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 686
 687             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 688
 689             /* Load parameters for j particles */
 690             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 691                                                               charge+jnrC+0,charge+jnrD+0);
 692             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 693                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 694             vdwjidx0A        = 2*vdwtype[jnrA+0];
 695             vdwjidx0B        = 2*vdwtype[jnrB+0];
 696             vdwjidx0C        = 2*vdwtype[jnrC+0];
 697             vdwjidx0D        = 2*vdwtype[jnrD+0];
 698
 699             /**************************
 700              * CALCULATE INTERACTIONS *
 701              **************************/
 702
 703             r00              = _mm_mul_ps(rsq00,rinv00);
 704
 705             /* Compute parameters for interactions between i and j atoms */
 706             qq00             = _mm_mul_ps(iq0,jq0);
 707             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 708                                          vdwparam+vdwioffset0+vdwjidx0B,
 709                                          vdwparam+vdwioffset0+vdwjidx0C,
 710                                          vdwparam+vdwioffset0+vdwjidx0D,
 711                                          &c6_00,&c12_00);
 712
 713             /* Calculate table index by multiplying r with table scale and truncate to integer */
 714             rt               = _mm_mul_ps(r00,vftabscale);
 715             vfitab           = _mm_cvttps_epi32(rt);
 716 #ifdef __XOP__
 717             vfeps            = _mm_frcz_ps(rt);
 718 #else
 719             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 720 #endif
 721             twovfeps         = _mm_add_ps(vfeps,vfeps);
 722             vfitab           = _mm_slli_epi32(vfitab,3);
 723
 724             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 725             isaprod          = _mm_mul_ps(isai0,isaj0);
 726             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 727             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 728
 729             /* Calculate generalized born table index - this is a separate table from the normal one,
 730              * but we use the same procedure by multiplying r with scale and truncating to integer.
 731              */
 732             rt               = _mm_mul_ps(r00,gbscale);
 733             gbitab           = _mm_cvttps_epi32(rt);
 734 #ifdef __XOP__
 735             gbeps            = _mm_frcz_ps(rt);
 736 #else
 737             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 738 #endif
 739             gbitab           = _mm_slli_epi32(gbitab,2);
 740
 741             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 742             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 743             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 744             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 745             _MM_TRANSPOSE4_PS(Y,F,G,H);
 746             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 747             VV               = _mm_macc_ps(gbeps,Fp,Y);
 748             vgb              = _mm_mul_ps(gbqqfactor,VV);
 749
 750             twogbeps         = _mm_add_ps(gbeps,gbeps);
 751             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 752             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 753             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 754             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 755             fjptrA           = dvda+jnrA;
 756             fjptrB           = dvda+jnrB;
 757             fjptrC           = dvda+jnrC;
 758             fjptrD           = dvda+jnrD;
 759             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 760             velec            = _mm_mul_ps(qq00,rinv00);
 761             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 762
 763             /* CUBIC SPLINE TABLE DISPERSION */
 764             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 765             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 766             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 767             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 768             _MM_TRANSPOSE4_PS(Y,F,G,H);
 769             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 770             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 771             fvdw6            = _mm_mul_ps(c6_00,FF);
 772
 773             /* CUBIC SPLINE TABLE REPULSION */
 774             vfitab           = _mm_add_epi32(vfitab,ifour);
 775             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 776             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 777             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 778             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 779             _MM_TRANSPOSE4_PS(Y,F,G,H);
 780             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 781             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 782             fvdw12           = _mm_mul_ps(c12_00,FF);
 783             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 784
 785             fscal            = _mm_add_ps(felec,fvdw);
 786
 787              /* Update vectorial force */
 788             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 789             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 790             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 791
 792             fjptrA             = f+j_coord_offsetA;
 793             fjptrB             = f+j_coord_offsetB;
 794             fjptrC             = f+j_coord_offsetC;
 795             fjptrD             = f+j_coord_offsetD;
 796             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 797                                                    _mm_mul_ps(dx00,fscal),
 798                                                    _mm_mul_ps(dy00,fscal),
 799                                                    _mm_mul_ps(dz00,fscal));
 800
 801             /* Inner loop uses 85 flops */
 802         }
 803
 804         if(jidx<j_index_end)
 805         {
 806
 807             /* Get j neighbor index, and coordinate index */
 808             jnrlistA         = jjnr[jidx];
 809             jnrlistB         = jjnr[jidx+1];
 810             jnrlistC         = jjnr[jidx+2];
 811             jnrlistD         = jjnr[jidx+3];
 812             /* Sign of each element will be negative for non-real atoms.
 813              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 814              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 815              */
 816             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 817             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 818             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 819             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 820             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 821             j_coord_offsetA  = DIM*jnrA;
 822             j_coord_offsetB  = DIM*jnrB;
 823             j_coord_offsetC  = DIM*jnrC;
 824             j_coord_offsetD  = DIM*jnrD;
 825
 826             /* load j atom coordinates */
 827             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 828                                               x+j_coord_offsetC,x+j_coord_offsetD,
 829                                               &jx0,&jy0,&jz0);
 830
 831             /* Calculate displacement vector */
 832             dx00             = _mm_sub_ps(ix0,jx0);
 833             dy00             = _mm_sub_ps(iy0,jy0);
 834             dz00             = _mm_sub_ps(iz0,jz0);
 835
 836             /* Calculate squared distance and things based on it */
 837             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 838
 839             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 840
 841             /* Load parameters for j particles */
 842             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 843                                                               charge+jnrC+0,charge+jnrD+0);
 844             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 845                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 846             vdwjidx0A        = 2*vdwtype[jnrA+0];
 847             vdwjidx0B        = 2*vdwtype[jnrB+0];
 848             vdwjidx0C        = 2*vdwtype[jnrC+0];
 849             vdwjidx0D        = 2*vdwtype[jnrD+0];
 850
 851             /**************************
 852              * CALCULATE INTERACTIONS *
 853              **************************/
 854
 855             r00              = _mm_mul_ps(rsq00,rinv00);
 856             r00              = _mm_andnot_ps(dummy_mask,r00);
 857
 858             /* Compute parameters for interactions between i and j atoms */
 859             qq00             = _mm_mul_ps(iq0,jq0);
 860             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 861                                          vdwparam+vdwioffset0+vdwjidx0B,
 862                                          vdwparam+vdwioffset0+vdwjidx0C,
 863                                          vdwparam+vdwioffset0+vdwjidx0D,
 864                                          &c6_00,&c12_00);
 865
 866             /* Calculate table index by multiplying r with table scale and truncate to integer */
 867             rt               = _mm_mul_ps(r00,vftabscale);
 868             vfitab           = _mm_cvttps_epi32(rt);
 869 #ifdef __XOP__
 870             vfeps            = _mm_frcz_ps(rt);
 871 #else
 872             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 873 #endif
 874             twovfeps         = _mm_add_ps(vfeps,vfeps);
 875             vfitab           = _mm_slli_epi32(vfitab,3);
 876
 877             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 878             isaprod          = _mm_mul_ps(isai0,isaj0);
 879             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 880             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 881
 882             /* Calculate generalized born table index - this is a separate table from the normal one,
 883              * but we use the same procedure by multiplying r with scale and truncating to integer.
 884              */
 885             rt               = _mm_mul_ps(r00,gbscale);
 886             gbitab           = _mm_cvttps_epi32(rt);
 887 #ifdef __XOP__
 888             gbeps            = _mm_frcz_ps(rt);
 889 #else
 890             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 891 #endif
 892             gbitab           = _mm_slli_epi32(gbitab,2);
 893
 894             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 895             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 896             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 897             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 898             _MM_TRANSPOSE4_PS(Y,F,G,H);
 899             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 900             VV               = _mm_macc_ps(gbeps,Fp,Y);
 901             vgb              = _mm_mul_ps(gbqqfactor,VV);
 902
 903             twogbeps         = _mm_add_ps(gbeps,gbeps);
 904             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 905             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 906             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 907             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 908             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 909             /* 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. */
 910             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 911             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 912             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 913             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 914             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 915             velec            = _mm_mul_ps(qq00,rinv00);
 916             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 917
 918             /* CUBIC SPLINE TABLE DISPERSION */
 919             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 920             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 921             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 922             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 923             _MM_TRANSPOSE4_PS(Y,F,G,H);
 924             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 925             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 926             fvdw6            = _mm_mul_ps(c6_00,FF);
 927
 928             /* CUBIC SPLINE TABLE REPULSION */
 929             vfitab           = _mm_add_epi32(vfitab,ifour);
 930             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 931             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 932             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 933             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 934             _MM_TRANSPOSE4_PS(Y,F,G,H);
 935             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 936             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 937             fvdw12           = _mm_mul_ps(c12_00,FF);
 938             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 939
 940             fscal            = _mm_add_ps(felec,fvdw);
 941
 942             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 943
 944              /* Update vectorial force */
 945             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 946             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 947             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 948
 949             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 950             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 951             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 952             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 953             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 954                                                    _mm_mul_ps(dx00,fscal),
 955                                                    _mm_mul_ps(dy00,fscal),
 956                                                    _mm_mul_ps(dz00,fscal));
 957
 958             /* Inner loop uses 86 flops */
 959         }
 960
 961         /* End of innermost loop */
 962
 963         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 964                                               f+i_coord_offset,fshift+i_shift_offset);
 965
 966         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 967         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 968
 969         /* Increment number of inner iterations */
 970         inneriter                  += j_index_end - j_index_start;
 971
 972         /* Outer loop uses 7 flops */
 973     }
 974
 975     /* Increment number of outer iterations */
 976     outeriter        += nri;
 977
 978     /* Update outer/inner flops */
 979
 980     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*86);
 981 }