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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     __m128i          gbitab;
  91     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  92     __m128           minushalf = _mm_set1_ps(-0.5);
  93     real             *invsqrta,*dvda,*gbtab;
  94     int              nvdwtype;
  95     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  99     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 103     real             *vftab;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_ps(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_vdw->data;
 126     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 127
 128     invsqrta         = fr->invsqrta;
 129     dvda             = fr->dvda;
 130     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 131     gbtab            = fr->gbtab.data;
 132     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = jnrC = jnrD = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138     j_coord_offsetC = 0;
 139     j_coord_offsetD = 0;
 140
 141     outeriter        = 0;
 142     inneriter        = 0;
 143
 144     for(iidx=0;iidx<4*DIM;iidx++)
 145     {
 146         scratch[iidx] = 0.0;
 147     }
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm_setzero_ps();
 167         fiy0             = _mm_setzero_ps();
 168         fiz0             = _mm_setzero_ps();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 172         isai0            = _mm_load1_ps(invsqrta+inr+0);
 173         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_ps();
 177         vgbsum           = _mm_setzero_ps();
 178         vvdwsum          = _mm_setzero_ps();
 179         dvdasum          = _mm_setzero_ps();
 180
 181         /* Start inner kernel loop */
 182         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 183         {
 184
 185             /* Get j neighbor index, and coordinate index */
 186             jnrA             = jjnr[jidx];
 187             jnrB             = jjnr[jidx+1];
 188             jnrC             = jjnr[jidx+2];
 189             jnrD             = jjnr[jidx+3];
 190             j_coord_offsetA  = DIM*jnrA;
 191             j_coord_offsetB  = DIM*jnrB;
 192             j_coord_offsetC  = DIM*jnrC;
 193             j_coord_offsetD  = DIM*jnrD;
 194
 195             /* load j atom coordinates */
 196             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                               x+j_coord_offsetC,x+j_coord_offsetD,
 198                                               &jx0,&jy0,&jz0);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm_sub_ps(ix0,jx0);
 202             dy00             = _mm_sub_ps(iy0,jy0);
 203             dz00             = _mm_sub_ps(iz0,jz0);
 204
 205             /* Calculate squared distance and things based on it */
 206             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 207
 208             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 209
 210             /* Load parameters for j particles */
 211             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 212                                                               charge+jnrC+0,charge+jnrD+0);
 213             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 214                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 215             vdwjidx0A        = 2*vdwtype[jnrA+0];
 216             vdwjidx0B        = 2*vdwtype[jnrB+0];
 217             vdwjidx0C        = 2*vdwtype[jnrC+0];
 218             vdwjidx0D        = 2*vdwtype[jnrD+0];
 219
 220             /**************************
 221              * CALCULATE INTERACTIONS *
 222              **************************/
 223
 224             r00              = _mm_mul_ps(rsq00,rinv00);
 225
 226             /* Compute parameters for interactions between i and j atoms */
 227             qq00             = _mm_mul_ps(iq0,jq0);
 228             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 229                                          vdwparam+vdwioffset0+vdwjidx0B,
 230                                          vdwparam+vdwioffset0+vdwjidx0C,
 231                                          vdwparam+vdwioffset0+vdwjidx0D,
 232                                          &c6_00,&c12_00);
 233
 234             /* Calculate table index by multiplying r with table scale and truncate to integer */
 235             rt               = _mm_mul_ps(r00,vftabscale);
 236             vfitab           = _mm_cvttps_epi32(rt);
 237 #ifdef __XOP__
 238             vfeps            = _mm_frcz_ps(rt);
 239 #else
 240             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 241 #endif
 242             twovfeps         = _mm_add_ps(vfeps,vfeps);
 243             vfitab           = _mm_slli_epi32(vfitab,3);
 244
 245             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 246             isaprod          = _mm_mul_ps(isai0,isaj0);
 247             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 248             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 249
 250             /* Calculate generalized born table index - this is a separate table from the normal one,
 251              * but we use the same procedure by multiplying r with scale and truncating to integer.
 252              */
 253             rt               = _mm_mul_ps(r00,gbscale);
 254             gbitab           = _mm_cvttps_epi32(rt);
 255 #ifdef __XOP__
 256             gbeps            = _mm_frcz_ps(rt);
 257 #else
 258             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 259 #endif
 260             gbitab           = _mm_slli_epi32(gbitab,2);
 261
 262             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 263             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 264             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 265             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 266             _MM_TRANSPOSE4_PS(Y,F,G,H);
 267             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 268             VV               = _mm_macc_ps(gbeps,Fp,Y);
 269             vgb              = _mm_mul_ps(gbqqfactor,VV);
 270
 271             twogbeps         = _mm_add_ps(gbeps,gbeps);
 272             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 273             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 274             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 275             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 276             fjptrA           = dvda+jnrA;
 277             fjptrB           = dvda+jnrB;
 278             fjptrC           = dvda+jnrC;
 279             fjptrD           = dvda+jnrD;
 280             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 281             velec            = _mm_mul_ps(qq00,rinv00);
 282             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 283
 284             /* CUBIC SPLINE TABLE DISPERSION */
 285             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 286             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 287             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 288             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 289             _MM_TRANSPOSE4_PS(Y,F,G,H);
 290             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 291             VV               = _mm_macc_ps(vfeps,Fp,Y);
 292             vvdw6            = _mm_mul_ps(c6_00,VV);
 293             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 294             fvdw6            = _mm_mul_ps(c6_00,FF);
 295
 296             /* CUBIC SPLINE TABLE REPULSION */
 297             vfitab           = _mm_add_epi32(vfitab,ifour);
 298             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 299             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 300             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 301             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 302             _MM_TRANSPOSE4_PS(Y,F,G,H);
 303             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 304             VV               = _mm_macc_ps(vfeps,Fp,Y);
 305             vvdw12           = _mm_mul_ps(c12_00,VV);
 306             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 307             fvdw12           = _mm_mul_ps(c12_00,FF);
 308             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 309             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 310
 311             /* Update potential sum for this i atom from the interaction with this j atom. */
 312             velecsum         = _mm_add_ps(velecsum,velec);
 313             vgbsum           = _mm_add_ps(vgbsum,vgb);
 314             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 315
 316             fscal            = _mm_add_ps(felec,fvdw);
 317
 318              /* Update vectorial force */
 319             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 320             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 321             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 322
 323             fjptrA             = f+j_coord_offsetA;
 324             fjptrB             = f+j_coord_offsetB;
 325             fjptrC             = f+j_coord_offsetC;
 326             fjptrD             = f+j_coord_offsetD;
 327             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 328                                                    _mm_mul_ps(dx00,fscal),
 329                                                    _mm_mul_ps(dy00,fscal),
 330                                                    _mm_mul_ps(dz00,fscal));
 331
 332             /* Inner loop uses 95 flops */
 333         }
 334
 335         if(jidx<j_index_end)
 336         {
 337
 338             /* Get j neighbor index, and coordinate index */
 339             jnrlistA         = jjnr[jidx];
 340             jnrlistB         = jjnr[jidx+1];
 341             jnrlistC         = jjnr[jidx+2];
 342             jnrlistD         = jjnr[jidx+3];
 343             /* Sign of each element will be negative for non-real atoms.
 344              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 345              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 346              */
 347             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 348             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 349             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 350             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 351             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 352             j_coord_offsetA  = DIM*jnrA;
 353             j_coord_offsetB  = DIM*jnrB;
 354             j_coord_offsetC  = DIM*jnrC;
 355             j_coord_offsetD  = DIM*jnrD;
 356
 357             /* load j atom coordinates */
 358             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 359                                               x+j_coord_offsetC,x+j_coord_offsetD,
 360                                               &jx0,&jy0,&jz0);
 361
 362             /* Calculate displacement vector */
 363             dx00             = _mm_sub_ps(ix0,jx0);
 364             dy00             = _mm_sub_ps(iy0,jy0);
 365             dz00             = _mm_sub_ps(iz0,jz0);
 366
 367             /* Calculate squared distance and things based on it */
 368             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 369
 370             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 371
 372             /* Load parameters for j particles */
 373             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 374                                                               charge+jnrC+0,charge+jnrD+0);
 375             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 376                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 377             vdwjidx0A        = 2*vdwtype[jnrA+0];
 378             vdwjidx0B        = 2*vdwtype[jnrB+0];
 379             vdwjidx0C        = 2*vdwtype[jnrC+0];
 380             vdwjidx0D        = 2*vdwtype[jnrD+0];
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             r00              = _mm_mul_ps(rsq00,rinv00);
 387             r00              = _mm_andnot_ps(dummy_mask,r00);
 388
 389             /* Compute parameters for interactions between i and j atoms */
 390             qq00             = _mm_mul_ps(iq0,jq0);
 391             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 392                                          vdwparam+vdwioffset0+vdwjidx0B,
 393                                          vdwparam+vdwioffset0+vdwjidx0C,
 394                                          vdwparam+vdwioffset0+vdwjidx0D,
 395                                          &c6_00,&c12_00);
 396
 397             /* Calculate table index by multiplying r with table scale and truncate to integer */
 398             rt               = _mm_mul_ps(r00,vftabscale);
 399             vfitab           = _mm_cvttps_epi32(rt);
 400 #ifdef __XOP__
 401             vfeps            = _mm_frcz_ps(rt);
 402 #else
 403             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 404 #endif
 405             twovfeps         = _mm_add_ps(vfeps,vfeps);
 406             vfitab           = _mm_slli_epi32(vfitab,3);
 407
 408             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 409             isaprod          = _mm_mul_ps(isai0,isaj0);
 410             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 411             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 412
 413             /* Calculate generalized born table index - this is a separate table from the normal one,
 414              * but we use the same procedure by multiplying r with scale and truncating to integer.
 415              */
 416             rt               = _mm_mul_ps(r00,gbscale);
 417             gbitab           = _mm_cvttps_epi32(rt);
 418 #ifdef __XOP__
 419             gbeps            = _mm_frcz_ps(rt);
 420 #else
 421             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 422 #endif
 423             gbitab           = _mm_slli_epi32(gbitab,2);
 424
 425             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 426             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 427             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 428             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 429             _MM_TRANSPOSE4_PS(Y,F,G,H);
 430             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 431             VV               = _mm_macc_ps(gbeps,Fp,Y);
 432             vgb              = _mm_mul_ps(gbqqfactor,VV);
 433
 434             twogbeps         = _mm_add_ps(gbeps,gbeps);
 435             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 436             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 437             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 438             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 439             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 440             /* 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. */
 441             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 442             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 443             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 444             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 445             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 446             velec            = _mm_mul_ps(qq00,rinv00);
 447             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 448
 449             /* CUBIC SPLINE TABLE DISPERSION */
 450             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 451             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 452             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 453             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 454             _MM_TRANSPOSE4_PS(Y,F,G,H);
 455             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 456             VV               = _mm_macc_ps(vfeps,Fp,Y);
 457             vvdw6            = _mm_mul_ps(c6_00,VV);
 458             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 459             fvdw6            = _mm_mul_ps(c6_00,FF);
 460
 461             /* CUBIC SPLINE TABLE REPULSION */
 462             vfitab           = _mm_add_epi32(vfitab,ifour);
 463             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 464             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 465             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 466             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 467             _MM_TRANSPOSE4_PS(Y,F,G,H);
 468             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 469             VV               = _mm_macc_ps(vfeps,Fp,Y);
 470             vvdw12           = _mm_mul_ps(c12_00,VV);
 471             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 472             fvdw12           = _mm_mul_ps(c12_00,FF);
 473             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 474             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 475
 476             /* Update potential sum for this i atom from the interaction with this j atom. */
 477             velec            = _mm_andnot_ps(dummy_mask,velec);
 478             velecsum         = _mm_add_ps(velecsum,velec);
 479             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 480             vgbsum           = _mm_add_ps(vgbsum,vgb);
 481             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 482             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 483
 484             fscal            = _mm_add_ps(felec,fvdw);
 485
 486             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 487
 488              /* Update vectorial force */
 489             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 490             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 491             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 492
 493             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 494             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 495             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 496             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 497             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 498                                                    _mm_mul_ps(dx00,fscal),
 499                                                    _mm_mul_ps(dy00,fscal),
 500                                                    _mm_mul_ps(dz00,fscal));
 501
 502             /* Inner loop uses 96 flops */
 503         }
 504
 505         /* End of innermost loop */
 506
 507         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 508                                               f+i_coord_offset,fshift+i_shift_offset);
 509
 510         ggid                        = gid[iidx];
 511         /* Update potential energies */
 512         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 513         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 514         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 515         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 516         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 517
 518         /* Increment number of inner iterations */
 519         inneriter                  += j_index_end - j_index_start;
 520
 521         /* Outer loop uses 10 flops */
 522     }
 523
 524     /* Increment number of outer iterations */
 525     outeriter        += nri;
 526
 527     /* Update outer/inner flops */
 528
 529     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*96);
 530 }
 531 /*
 532  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 533  * Electrostatics interaction: GeneralizedBorn
 534  * VdW interaction:            CubicSplineTable
 535  * Geometry:                   Particle-Particle
 536  * Calculate force/pot:        Force
 537  */
 538 void
 539 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 540                     (t_nblist                    * gmx_restrict       nlist,
 541                      rvec                        * gmx_restrict          xx,
 542                      rvec                        * gmx_restrict          ff,
 543                      t_forcerec                  * gmx_restrict          fr,
 544                      t_mdatoms                   * gmx_restrict     mdatoms,
 545                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 546                      t_nrnb                      * gmx_restrict        nrnb)
 547 {
 548     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 549      * just 0 for non-waters.
 550      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 551      * jnr indices corresponding to data put in the four positions in the SIMD register.
 552      */
 553     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 554     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 555     int              jnrA,jnrB,jnrC,jnrD;
 556     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 557     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 558     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 559     real             rcutoff_scalar;
 560     real             *shiftvec,*fshift,*x,*f;
 561     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 562     real             scratch[4*DIM];
 563     __m128           fscal,rcutoff,rcutoff2,jidxall;
 564     int              vdwioffset0;
 565     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 566     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 567     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 568     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 569     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 570     real             *charge;
 571     __m128i          gbitab;
 572     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 573     __m128           minushalf = _mm_set1_ps(-0.5);
 574     real             *invsqrta,*dvda,*gbtab;
 575     int              nvdwtype;
 576     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 577     int              *vdwtype;
 578     real             *vdwparam;
 579     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 580     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 581     __m128i          vfitab;
 582     __m128i          ifour       = _mm_set1_epi32(4);
 583     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 584     real             *vftab;
 585     __m128           dummy_mask,cutoff_mask;
 586     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 587     __m128           one     = _mm_set1_ps(1.0);
 588     __m128           two     = _mm_set1_ps(2.0);
 589     x                = xx[0];
 590     f                = ff[0];
 591
 592     nri              = nlist->nri;
 593     iinr             = nlist->iinr;
 594     jindex           = nlist->jindex;
 595     jjnr             = nlist->jjnr;
 596     shiftidx         = nlist->shift;
 597     gid              = nlist->gid;
 598     shiftvec         = fr->shift_vec[0];
 599     fshift           = fr->fshift[0];
 600     facel            = _mm_set1_ps(fr->epsfac);
 601     charge           = mdatoms->chargeA;
 602     nvdwtype         = fr->ntype;
 603     vdwparam         = fr->nbfp;
 604     vdwtype          = mdatoms->typeA;
 605
 606     vftab            = kernel_data->table_vdw->data;
 607     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 608
 609     invsqrta         = fr->invsqrta;
 610     dvda             = fr->dvda;
 611     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 612     gbtab            = fr->gbtab.data;
 613     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 614
 615     /* Avoid stupid compiler warnings */
 616     jnrA = jnrB = jnrC = jnrD = 0;
 617     j_coord_offsetA = 0;
 618     j_coord_offsetB = 0;
 619     j_coord_offsetC = 0;
 620     j_coord_offsetD = 0;
 621
 622     outeriter        = 0;
 623     inneriter        = 0;
 624
 625     for(iidx=0;iidx<4*DIM;iidx++)
 626     {
 627         scratch[iidx] = 0.0;
 628     }
 629
 630     /* Start outer loop over neighborlists */
 631     for(iidx=0; iidx<nri; iidx++)
 632     {
 633         /* Load shift vector for this list */
 634         i_shift_offset   = DIM*shiftidx[iidx];
 635
 636         /* Load limits for loop over neighbors */
 637         j_index_start    = jindex[iidx];
 638         j_index_end      = jindex[iidx+1];
 639
 640         /* Get outer coordinate index */
 641         inr              = iinr[iidx];
 642         i_coord_offset   = DIM*inr;
 643
 644         /* Load i particle coords and add shift vector */
 645         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 646
 647         fix0             = _mm_setzero_ps();
 648         fiy0             = _mm_setzero_ps();
 649         fiz0             = _mm_setzero_ps();
 650
 651         /* Load parameters for i particles */
 652         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 653         isai0            = _mm_load1_ps(invsqrta+inr+0);
 654         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 655
 656         dvdasum          = _mm_setzero_ps();
 657
 658         /* Start inner kernel loop */
 659         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 660         {
 661
 662             /* Get j neighbor index, and coordinate index */
 663             jnrA             = jjnr[jidx];
 664             jnrB             = jjnr[jidx+1];
 665             jnrC             = jjnr[jidx+2];
 666             jnrD             = jjnr[jidx+3];
 667             j_coord_offsetA  = DIM*jnrA;
 668             j_coord_offsetB  = DIM*jnrB;
 669             j_coord_offsetC  = DIM*jnrC;
 670             j_coord_offsetD  = DIM*jnrD;
 671
 672             /* load j atom coordinates */
 673             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 674                                               x+j_coord_offsetC,x+j_coord_offsetD,
 675                                               &jx0,&jy0,&jz0);
 676
 677             /* Calculate displacement vector */
 678             dx00             = _mm_sub_ps(ix0,jx0);
 679             dy00             = _mm_sub_ps(iy0,jy0);
 680             dz00             = _mm_sub_ps(iz0,jz0);
 681
 682             /* Calculate squared distance and things based on it */
 683             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 684
 685             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 686
 687             /* Load parameters for j particles */
 688             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 689                                                               charge+jnrC+0,charge+jnrD+0);
 690             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 691                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 692             vdwjidx0A        = 2*vdwtype[jnrA+0];
 693             vdwjidx0B        = 2*vdwtype[jnrB+0];
 694             vdwjidx0C        = 2*vdwtype[jnrC+0];
 695             vdwjidx0D        = 2*vdwtype[jnrD+0];
 696
 697             /**************************
 698              * CALCULATE INTERACTIONS *
 699              **************************/
 700
 701             r00              = _mm_mul_ps(rsq00,rinv00);
 702
 703             /* Compute parameters for interactions between i and j atoms */
 704             qq00             = _mm_mul_ps(iq0,jq0);
 705             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 706                                          vdwparam+vdwioffset0+vdwjidx0B,
 707                                          vdwparam+vdwioffset0+vdwjidx0C,
 708                                          vdwparam+vdwioffset0+vdwjidx0D,
 709                                          &c6_00,&c12_00);
 710
 711             /* Calculate table index by multiplying r with table scale and truncate to integer */
 712             rt               = _mm_mul_ps(r00,vftabscale);
 713             vfitab           = _mm_cvttps_epi32(rt);
 714 #ifdef __XOP__
 715             vfeps            = _mm_frcz_ps(rt);
 716 #else
 717             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 718 #endif
 719             twovfeps         = _mm_add_ps(vfeps,vfeps);
 720             vfitab           = _mm_slli_epi32(vfitab,3);
 721
 722             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 723             isaprod          = _mm_mul_ps(isai0,isaj0);
 724             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 725             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 726
 727             /* Calculate generalized born table index - this is a separate table from the normal one,
 728              * but we use the same procedure by multiplying r with scale and truncating to integer.
 729              */
 730             rt               = _mm_mul_ps(r00,gbscale);
 731             gbitab           = _mm_cvttps_epi32(rt);
 732 #ifdef __XOP__
 733             gbeps            = _mm_frcz_ps(rt);
 734 #else
 735             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 736 #endif
 737             gbitab           = _mm_slli_epi32(gbitab,2);
 738
 739             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 740             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 741             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 742             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 743             _MM_TRANSPOSE4_PS(Y,F,G,H);
 744             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 745             VV               = _mm_macc_ps(gbeps,Fp,Y);
 746             vgb              = _mm_mul_ps(gbqqfactor,VV);
 747
 748             twogbeps         = _mm_add_ps(gbeps,gbeps);
 749             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 750             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 751             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 752             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 753             fjptrA           = dvda+jnrA;
 754             fjptrB           = dvda+jnrB;
 755             fjptrC           = dvda+jnrC;
 756             fjptrD           = dvda+jnrD;
 757             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 758             velec            = _mm_mul_ps(qq00,rinv00);
 759             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 760
 761             /* CUBIC SPLINE TABLE DISPERSION */
 762             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 763             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 764             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 765             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 766             _MM_TRANSPOSE4_PS(Y,F,G,H);
 767             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 768             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 769             fvdw6            = _mm_mul_ps(c6_00,FF);
 770
 771             /* CUBIC SPLINE TABLE REPULSION */
 772             vfitab           = _mm_add_epi32(vfitab,ifour);
 773             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 774             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 775             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 776             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 777             _MM_TRANSPOSE4_PS(Y,F,G,H);
 778             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 779             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 780             fvdw12           = _mm_mul_ps(c12_00,FF);
 781             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 782
 783             fscal            = _mm_add_ps(felec,fvdw);
 784
 785              /* Update vectorial force */
 786             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 787             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 788             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 789
 790             fjptrA             = f+j_coord_offsetA;
 791             fjptrB             = f+j_coord_offsetB;
 792             fjptrC             = f+j_coord_offsetC;
 793             fjptrD             = f+j_coord_offsetD;
 794             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 795                                                    _mm_mul_ps(dx00,fscal),
 796                                                    _mm_mul_ps(dy00,fscal),
 797                                                    _mm_mul_ps(dz00,fscal));
 798
 799             /* Inner loop uses 85 flops */
 800         }
 801
 802         if(jidx<j_index_end)
 803         {
 804
 805             /* Get j neighbor index, and coordinate index */
 806             jnrlistA         = jjnr[jidx];
 807             jnrlistB         = jjnr[jidx+1];
 808             jnrlistC         = jjnr[jidx+2];
 809             jnrlistD         = jjnr[jidx+3];
 810             /* Sign of each element will be negative for non-real atoms.
 811              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 812              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 813              */
 814             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 815             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 816             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 817             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 818             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 819             j_coord_offsetA  = DIM*jnrA;
 820             j_coord_offsetB  = DIM*jnrB;
 821             j_coord_offsetC  = DIM*jnrC;
 822             j_coord_offsetD  = DIM*jnrD;
 823
 824             /* load j atom coordinates */
 825             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 826                                               x+j_coord_offsetC,x+j_coord_offsetD,
 827                                               &jx0,&jy0,&jz0);
 828
 829             /* Calculate displacement vector */
 830             dx00             = _mm_sub_ps(ix0,jx0);
 831             dy00             = _mm_sub_ps(iy0,jy0);
 832             dz00             = _mm_sub_ps(iz0,jz0);
 833
 834             /* Calculate squared distance and things based on it */
 835             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 836
 837             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 838
 839             /* Load parameters for j particles */
 840             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 841                                                               charge+jnrC+0,charge+jnrD+0);
 842             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 843                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 844             vdwjidx0A        = 2*vdwtype[jnrA+0];
 845             vdwjidx0B        = 2*vdwtype[jnrB+0];
 846             vdwjidx0C        = 2*vdwtype[jnrC+0];
 847             vdwjidx0D        = 2*vdwtype[jnrD+0];
 848
 849             /**************************
 850              * CALCULATE INTERACTIONS *
 851              **************************/
 852
 853             r00              = _mm_mul_ps(rsq00,rinv00);
 854             r00              = _mm_andnot_ps(dummy_mask,r00);
 855
 856             /* Compute parameters for interactions between i and j atoms */
 857             qq00             = _mm_mul_ps(iq0,jq0);
 858             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 859                                          vdwparam+vdwioffset0+vdwjidx0B,
 860                                          vdwparam+vdwioffset0+vdwjidx0C,
 861                                          vdwparam+vdwioffset0+vdwjidx0D,
 862                                          &c6_00,&c12_00);
 863
 864             /* Calculate table index by multiplying r with table scale and truncate to integer */
 865             rt               = _mm_mul_ps(r00,vftabscale);
 866             vfitab           = _mm_cvttps_epi32(rt);
 867 #ifdef __XOP__
 868             vfeps            = _mm_frcz_ps(rt);
 869 #else
 870             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 871 #endif
 872             twovfeps         = _mm_add_ps(vfeps,vfeps);
 873             vfitab           = _mm_slli_epi32(vfitab,3);
 874
 875             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 876             isaprod          = _mm_mul_ps(isai0,isaj0);
 877             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 878             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 879
 880             /* Calculate generalized born table index - this is a separate table from the normal one,
 881              * but we use the same procedure by multiplying r with scale and truncating to integer.
 882              */
 883             rt               = _mm_mul_ps(r00,gbscale);
 884             gbitab           = _mm_cvttps_epi32(rt);
 885 #ifdef __XOP__
 886             gbeps            = _mm_frcz_ps(rt);
 887 #else
 888             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 889 #endif
 890             gbitab           = _mm_slli_epi32(gbitab,2);
 891
 892             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 893             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 894             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 895             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 896             _MM_TRANSPOSE4_PS(Y,F,G,H);
 897             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 898             VV               = _mm_macc_ps(gbeps,Fp,Y);
 899             vgb              = _mm_mul_ps(gbqqfactor,VV);
 900
 901             twogbeps         = _mm_add_ps(gbeps,gbeps);
 902             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 903             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 904             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 905             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 906             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 907             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 908             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 909             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 910             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 911             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 912             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 913             velec            = _mm_mul_ps(qq00,rinv00);
 914             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 915
 916             /* CUBIC SPLINE TABLE DISPERSION */
 917             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 918             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 919             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 920             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 921             _MM_TRANSPOSE4_PS(Y,F,G,H);
 922             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 923             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 924             fvdw6            = _mm_mul_ps(c6_00,FF);
 925
 926             /* CUBIC SPLINE TABLE REPULSION */
 927             vfitab           = _mm_add_epi32(vfitab,ifour);
 928             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 929             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 930             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 931             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 932             _MM_TRANSPOSE4_PS(Y,F,G,H);
 933             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 934             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 935             fvdw12           = _mm_mul_ps(c12_00,FF);
 936             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 937
 938             fscal            = _mm_add_ps(felec,fvdw);
 939
 940             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 941
 942              /* Update vectorial force */
 943             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 944             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 945             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 946
 947             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 948             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 949             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 950             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 951             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 952                                                    _mm_mul_ps(dx00,fscal),
 953                                                    _mm_mul_ps(dy00,fscal),
 954                                                    _mm_mul_ps(dz00,fscal));
 955
 956             /* Inner loop uses 86 flops */
 957         }
 958
 959         /* End of innermost loop */
 960
 961         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 962                                               f+i_coord_offset,fshift+i_shift_offset);
 963
 964         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 965         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 966
 967         /* Increment number of inner iterations */
 968         inneriter                  += j_index_end - j_index_start;
 969
 970         /* Outer loop uses 7 flops */
 971     }
 972
 973     /* Increment number of outer iterations */
 974     outeriter        += nri;
 975
 976     /* Update outer/inner flops */
 977
 978     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*86);
 979 }