src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     __m128i          gbitab;
  77     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  78     __m128           minushalf = _mm_set1_ps(-0.5);
  79     real             *invsqrta,*dvda,*gbtab;
  80     int              nvdwtype;
  81     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  82     int              *vdwtype;
  83     real             *vdwparam;
  84     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  85     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  86     __m128i          vfitab;
  87     __m128i          ifour       = _mm_set1_epi32(4);
  88     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  89     real             *vftab;
  90     __m128           dummy_mask,cutoff_mask;
  91     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  92     __m128           one     = _mm_set1_ps(1.0);
  93     __m128           two     = _mm_set1_ps(2.0);
  94     x                = xx[0];
  95     f                = ff[0];
  96
  97     nri              = nlist->nri;
  98     iinr             = nlist->iinr;
  99     jindex           = nlist->jindex;
 100     jjnr             = nlist->jjnr;
 101     shiftidx         = nlist->shift;
 102     gid              = nlist->gid;
 103     shiftvec         = fr->shift_vec[0];
 104     fshift           = fr->fshift[0];
 105     facel            = _mm_set1_ps(fr->epsfac);
 106     charge           = mdatoms->chargeA;
 107     nvdwtype         = fr->ntype;
 108     vdwparam         = fr->nbfp;
 109     vdwtype          = mdatoms->typeA;
 110
 111     vftab            = kernel_data->table_vdw->data;
 112     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 113
 114     invsqrta         = fr->invsqrta;
 115     dvda             = fr->dvda;
 116     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 117     gbtab            = fr->gbtab.data;
 118     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 119
 120     /* Avoid stupid compiler warnings */
 121     jnrA = jnrB = jnrC = jnrD = 0;
 122     j_coord_offsetA = 0;
 123     j_coord_offsetB = 0;
 124     j_coord_offsetC = 0;
 125     j_coord_offsetD = 0;
 126
 127     outeriter        = 0;
 128     inneriter        = 0;
 129
 130     for(iidx=0;iidx<4*DIM;iidx++)
 131     {
 132         scratch[iidx] = 0.0;
 133     }
 134
 135     /* Start outer loop over neighborlists */
 136     for(iidx=0; iidx<nri; iidx++)
 137     {
 138         /* Load shift vector for this list */
 139         i_shift_offset   = DIM*shiftidx[iidx];
 140
 141         /* Load limits for loop over neighbors */
 142         j_index_start    = jindex[iidx];
 143         j_index_end      = jindex[iidx+1];
 144
 145         /* Get outer coordinate index */
 146         inr              = iinr[iidx];
 147         i_coord_offset   = DIM*inr;
 148
 149         /* Load i particle coords and add shift vector */
 150         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 151
 152         fix0             = _mm_setzero_ps();
 153         fiy0             = _mm_setzero_ps();
 154         fiz0             = _mm_setzero_ps();
 155
 156         /* Load parameters for i particles */
 157         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 158         isai0            = _mm_load1_ps(invsqrta+inr+0);
 159         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm_setzero_ps();
 163         vgbsum           = _mm_setzero_ps();
 164         vvdwsum          = _mm_setzero_ps();
 165         dvdasum          = _mm_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                               x+j_coord_offsetC,x+j_coord_offsetD,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_ps(ix0,jx0);
 188             dy00             = _mm_sub_ps(iy0,jy0);
 189             dz00             = _mm_sub_ps(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 193
 194             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 195
 196             /* Load parameters for j particles */
 197             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 198                                                               charge+jnrC+0,charge+jnrD+0);
 199             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 200                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 201             vdwjidx0A        = 2*vdwtype[jnrA+0];
 202             vdwjidx0B        = 2*vdwtype[jnrB+0];
 203             vdwjidx0C        = 2*vdwtype[jnrC+0];
 204             vdwjidx0D        = 2*vdwtype[jnrD+0];
 205
 206             /**************************
 207              * CALCULATE INTERACTIONS *
 208              **************************/
 209
 210             r00              = _mm_mul_ps(rsq00,rinv00);
 211
 212             /* Compute parameters for interactions between i and j atoms */
 213             qq00             = _mm_mul_ps(iq0,jq0);
 214             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 215                                          vdwparam+vdwioffset0+vdwjidx0B,
 216                                          vdwparam+vdwioffset0+vdwjidx0C,
 217                                          vdwparam+vdwioffset0+vdwjidx0D,
 218                                          &c6_00,&c12_00);
 219
 220             /* Calculate table index by multiplying r with table scale and truncate to integer */
 221             rt               = _mm_mul_ps(r00,vftabscale);
 222             vfitab           = _mm_cvttps_epi32(rt);
 223 #ifdef __XOP__
 224             vfeps            = _mm_frcz_ps(rt);
 225 #else
 226             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 227 #endif
 228             twovfeps         = _mm_add_ps(vfeps,vfeps);
 229             vfitab           = _mm_slli_epi32(vfitab,3);
 230
 231             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 232             isaprod          = _mm_mul_ps(isai0,isaj0);
 233             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 234             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 235
 236             /* Calculate generalized born table index - this is a separate table from the normal one,
 237              * but we use the same procedure by multiplying r with scale and truncating to integer.
 238              */
 239             rt               = _mm_mul_ps(r00,gbscale);
 240             gbitab           = _mm_cvttps_epi32(rt);
 241 #ifdef __XOP__
 242             gbeps            = _mm_frcz_ps(rt);
 243 #else
 244             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 245 #endif
 246             gbitab           = _mm_slli_epi32(gbitab,2);
 247
 248             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 249             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 250             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 251             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 252             _MM_TRANSPOSE4_PS(Y,F,G,H);
 253             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 254             VV               = _mm_macc_ps(gbeps,Fp,Y);
 255             vgb              = _mm_mul_ps(gbqqfactor,VV);
 256
 257             twogbeps         = _mm_add_ps(gbeps,gbeps);
 258             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 259             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 260             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 261             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 262             fjptrA           = dvda+jnrA;
 263             fjptrB           = dvda+jnrB;
 264             fjptrC           = dvda+jnrC;
 265             fjptrD           = dvda+jnrD;
 266             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 267             velec            = _mm_mul_ps(qq00,rinv00);
 268             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 269
 270             /* CUBIC SPLINE TABLE DISPERSION */
 271             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 272             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 273             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 274             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 275             _MM_TRANSPOSE4_PS(Y,F,G,H);
 276             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 277             VV               = _mm_macc_ps(vfeps,Fp,Y);
 278             vvdw6            = _mm_mul_ps(c6_00,VV);
 279             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 280             fvdw6            = _mm_mul_ps(c6_00,FF);
 281
 282             /* CUBIC SPLINE TABLE REPULSION */
 283             vfitab           = _mm_add_epi32(vfitab,ifour);
 284             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 285             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 286             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 287             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 288             _MM_TRANSPOSE4_PS(Y,F,G,H);
 289             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 290             VV               = _mm_macc_ps(vfeps,Fp,Y);
 291             vvdw12           = _mm_mul_ps(c12_00,VV);
 292             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 293             fvdw12           = _mm_mul_ps(c12_00,FF);
 294             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 295             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _mm_add_ps(velecsum,velec);
 299             vgbsum           = _mm_add_ps(vgbsum,vgb);
 300             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 301
 302             fscal            = _mm_add_ps(felec,fvdw);
 303
 304              /* Update vectorial force */
 305             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 306             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 307             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 308
 309             fjptrA             = f+j_coord_offsetA;
 310             fjptrB             = f+j_coord_offsetB;
 311             fjptrC             = f+j_coord_offsetC;
 312             fjptrD             = f+j_coord_offsetD;
 313             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 314                                                    _mm_mul_ps(dx00,fscal),
 315                                                    _mm_mul_ps(dy00,fscal),
 316                                                    _mm_mul_ps(dz00,fscal));
 317
 318             /* Inner loop uses 95 flops */
 319         }
 320
 321         if(jidx<j_index_end)
 322         {
 323
 324             /* Get j neighbor index, and coordinate index */
 325             jnrlistA         = jjnr[jidx];
 326             jnrlistB         = jjnr[jidx+1];
 327             jnrlistC         = jjnr[jidx+2];
 328             jnrlistD         = jjnr[jidx+3];
 329             /* Sign of each element will be negative for non-real atoms.
 330              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 331              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 332              */
 333             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 334             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 335             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 336             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 337             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 338             j_coord_offsetA  = DIM*jnrA;
 339             j_coord_offsetB  = DIM*jnrB;
 340             j_coord_offsetC  = DIM*jnrC;
 341             j_coord_offsetD  = DIM*jnrD;
 342
 343             /* load j atom coordinates */
 344             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 345                                               x+j_coord_offsetC,x+j_coord_offsetD,
 346                                               &jx0,&jy0,&jz0);
 347
 348             /* Calculate displacement vector */
 349             dx00             = _mm_sub_ps(ix0,jx0);
 350             dy00             = _mm_sub_ps(iy0,jy0);
 351             dz00             = _mm_sub_ps(iz0,jz0);
 352
 353             /* Calculate squared distance and things based on it */
 354             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 355
 356             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 357
 358             /* Load parameters for j particles */
 359             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 360                                                               charge+jnrC+0,charge+jnrD+0);
 361             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 362                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 363             vdwjidx0A        = 2*vdwtype[jnrA+0];
 364             vdwjidx0B        = 2*vdwtype[jnrB+0];
 365             vdwjidx0C        = 2*vdwtype[jnrC+0];
 366             vdwjidx0D        = 2*vdwtype[jnrD+0];
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             r00              = _mm_mul_ps(rsq00,rinv00);
 373             r00              = _mm_andnot_ps(dummy_mask,r00);
 374
 375             /* Compute parameters for interactions between i and j atoms */
 376             qq00             = _mm_mul_ps(iq0,jq0);
 377             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 378                                          vdwparam+vdwioffset0+vdwjidx0B,
 379                                          vdwparam+vdwioffset0+vdwjidx0C,
 380                                          vdwparam+vdwioffset0+vdwjidx0D,
 381                                          &c6_00,&c12_00);
 382
 383             /* Calculate table index by multiplying r with table scale and truncate to integer */
 384             rt               = _mm_mul_ps(r00,vftabscale);
 385             vfitab           = _mm_cvttps_epi32(rt);
 386 #ifdef __XOP__
 387             vfeps            = _mm_frcz_ps(rt);
 388 #else
 389             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 390 #endif
 391             twovfeps         = _mm_add_ps(vfeps,vfeps);
 392             vfitab           = _mm_slli_epi32(vfitab,3);
 393
 394             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 395             isaprod          = _mm_mul_ps(isai0,isaj0);
 396             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 397             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 398
 399             /* Calculate generalized born table index - this is a separate table from the normal one,
 400              * but we use the same procedure by multiplying r with scale and truncating to integer.
 401              */
 402             rt               = _mm_mul_ps(r00,gbscale);
 403             gbitab           = _mm_cvttps_epi32(rt);
 404 #ifdef __XOP__
 405             gbeps            = _mm_frcz_ps(rt);
 406 #else
 407             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 408 #endif
 409             gbitab           = _mm_slli_epi32(gbitab,2);
 410
 411             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 412             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 413             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 414             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 415             _MM_TRANSPOSE4_PS(Y,F,G,H);
 416             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 417             VV               = _mm_macc_ps(gbeps,Fp,Y);
 418             vgb              = _mm_mul_ps(gbqqfactor,VV);
 419
 420             twogbeps         = _mm_add_ps(gbeps,gbeps);
 421             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 422             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 423             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 424             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 425             /* 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. */
 426             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 427             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 428             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 429             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 430             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 431             velec            = _mm_mul_ps(qq00,rinv00);
 432             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 433
 434             /* CUBIC SPLINE TABLE DISPERSION */
 435             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 436             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 437             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 438             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 439             _MM_TRANSPOSE4_PS(Y,F,G,H);
 440             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 441             VV               = _mm_macc_ps(vfeps,Fp,Y);
 442             vvdw6            = _mm_mul_ps(c6_00,VV);
 443             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 444             fvdw6            = _mm_mul_ps(c6_00,FF);
 445
 446             /* CUBIC SPLINE TABLE REPULSION */
 447             vfitab           = _mm_add_epi32(vfitab,ifour);
 448             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 449             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 450             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 451             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 452             _MM_TRANSPOSE4_PS(Y,F,G,H);
 453             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 454             VV               = _mm_macc_ps(vfeps,Fp,Y);
 455             vvdw12           = _mm_mul_ps(c12_00,VV);
 456             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 457             fvdw12           = _mm_mul_ps(c12_00,FF);
 458             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 459             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 460
 461             /* Update potential sum for this i atom from the interaction with this j atom. */
 462             velec            = _mm_andnot_ps(dummy_mask,velec);
 463             velecsum         = _mm_add_ps(velecsum,velec);
 464             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 465             vgbsum           = _mm_add_ps(vgbsum,vgb);
 466             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 467             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 468
 469             fscal            = _mm_add_ps(felec,fvdw);
 470
 471             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 472
 473              /* Update vectorial force */
 474             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 475             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 476             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 477
 478             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 479             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 480             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 481             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 482             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 483                                                    _mm_mul_ps(dx00,fscal),
 484                                                    _mm_mul_ps(dy00,fscal),
 485                                                    _mm_mul_ps(dz00,fscal));
 486
 487             /* Inner loop uses 96 flops */
 488         }
 489
 490         /* End of innermost loop */
 491
 492         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 493                                               f+i_coord_offset,fshift+i_shift_offset);
 494
 495         ggid                        = gid[iidx];
 496         /* Update potential energies */
 497         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 498         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 499         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 500         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 501         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 502
 503         /* Increment number of inner iterations */
 504         inneriter                  += j_index_end - j_index_start;
 505
 506         /* Outer loop uses 10 flops */
 507     }
 508
 509     /* Increment number of outer iterations */
 510     outeriter        += nri;
 511
 512     /* Update outer/inner flops */
 513
 514     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*96);
 515 }
 516 /*
 517  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 518  * Electrostatics interaction: GeneralizedBorn
 519  * VdW interaction:            CubicSplineTable
 520  * Geometry:                   Particle-Particle
 521  * Calculate force/pot:        Force
 522  */
 523 void
 524 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 525                     (t_nblist * gmx_restrict                nlist,
 526                      rvec * gmx_restrict                    xx,
 527                      rvec * gmx_restrict                    ff,
 528                      t_forcerec * gmx_restrict              fr,
 529                      t_mdatoms * gmx_restrict               mdatoms,
 530                      nb_kernel_data_t * gmx_restrict        kernel_data,
 531                      t_nrnb * gmx_restrict                  nrnb)
 532 {
 533     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 534      * just 0 for non-waters.
 535      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 536      * jnr indices corresponding to data put in the four positions in the SIMD register.
 537      */
 538     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 539     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 540     int              jnrA,jnrB,jnrC,jnrD;
 541     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 542     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 543     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 544     real             rcutoff_scalar;
 545     real             *shiftvec,*fshift,*x,*f;
 546     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 547     real             scratch[4*DIM];
 548     __m128           fscal,rcutoff,rcutoff2,jidxall;
 549     int              vdwioffset0;
 550     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 551     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 552     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 553     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 554     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 555     real             *charge;
 556     __m128i          gbitab;
 557     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 558     __m128           minushalf = _mm_set1_ps(-0.5);
 559     real             *invsqrta,*dvda,*gbtab;
 560     int              nvdwtype;
 561     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 562     int              *vdwtype;
 563     real             *vdwparam;
 564     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 565     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 566     __m128i          vfitab;
 567     __m128i          ifour       = _mm_set1_epi32(4);
 568     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 569     real             *vftab;
 570     __m128           dummy_mask,cutoff_mask;
 571     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 572     __m128           one     = _mm_set1_ps(1.0);
 573     __m128           two     = _mm_set1_ps(2.0);
 574     x                = xx[0];
 575     f                = ff[0];
 576
 577     nri              = nlist->nri;
 578     iinr             = nlist->iinr;
 579     jindex           = nlist->jindex;
 580     jjnr             = nlist->jjnr;
 581     shiftidx         = nlist->shift;
 582     gid              = nlist->gid;
 583     shiftvec         = fr->shift_vec[0];
 584     fshift           = fr->fshift[0];
 585     facel            = _mm_set1_ps(fr->epsfac);
 586     charge           = mdatoms->chargeA;
 587     nvdwtype         = fr->ntype;
 588     vdwparam         = fr->nbfp;
 589     vdwtype          = mdatoms->typeA;
 590
 591     vftab            = kernel_data->table_vdw->data;
 592     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 593
 594     invsqrta         = fr->invsqrta;
 595     dvda             = fr->dvda;
 596     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 597     gbtab            = fr->gbtab.data;
 598     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 599
 600     /* Avoid stupid compiler warnings */
 601     jnrA = jnrB = jnrC = jnrD = 0;
 602     j_coord_offsetA = 0;
 603     j_coord_offsetB = 0;
 604     j_coord_offsetC = 0;
 605     j_coord_offsetD = 0;
 606
 607     outeriter        = 0;
 608     inneriter        = 0;
 609
 610     for(iidx=0;iidx<4*DIM;iidx++)
 611     {
 612         scratch[iidx] = 0.0;
 613     }
 614
 615     /* Start outer loop over neighborlists */
 616     for(iidx=0; iidx<nri; iidx++)
 617     {
 618         /* Load shift vector for this list */
 619         i_shift_offset   = DIM*shiftidx[iidx];
 620
 621         /* Load limits for loop over neighbors */
 622         j_index_start    = jindex[iidx];
 623         j_index_end      = jindex[iidx+1];
 624
 625         /* Get outer coordinate index */
 626         inr              = iinr[iidx];
 627         i_coord_offset   = DIM*inr;
 628
 629         /* Load i particle coords and add shift vector */
 630         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 631
 632         fix0             = _mm_setzero_ps();
 633         fiy0             = _mm_setzero_ps();
 634         fiz0             = _mm_setzero_ps();
 635
 636         /* Load parameters for i particles */
 637         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 638         isai0            = _mm_load1_ps(invsqrta+inr+0);
 639         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 640
 641         dvdasum          = _mm_setzero_ps();
 642
 643         /* Start inner kernel loop */
 644         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 645         {
 646
 647             /* Get j neighbor index, and coordinate index */
 648             jnrA             = jjnr[jidx];
 649             jnrB             = jjnr[jidx+1];
 650             jnrC             = jjnr[jidx+2];
 651             jnrD             = jjnr[jidx+3];
 652             j_coord_offsetA  = DIM*jnrA;
 653             j_coord_offsetB  = DIM*jnrB;
 654             j_coord_offsetC  = DIM*jnrC;
 655             j_coord_offsetD  = DIM*jnrD;
 656
 657             /* load j atom coordinates */
 658             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 659                                               x+j_coord_offsetC,x+j_coord_offsetD,
 660                                               &jx0,&jy0,&jz0);
 661
 662             /* Calculate displacement vector */
 663             dx00             = _mm_sub_ps(ix0,jx0);
 664             dy00             = _mm_sub_ps(iy0,jy0);
 665             dz00             = _mm_sub_ps(iz0,jz0);
 666
 667             /* Calculate squared distance and things based on it */
 668             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 669
 670             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 671
 672             /* Load parameters for j particles */
 673             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 674                                                               charge+jnrC+0,charge+jnrD+0);
 675             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 676                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 677             vdwjidx0A        = 2*vdwtype[jnrA+0];
 678             vdwjidx0B        = 2*vdwtype[jnrB+0];
 679             vdwjidx0C        = 2*vdwtype[jnrC+0];
 680             vdwjidx0D        = 2*vdwtype[jnrD+0];
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             r00              = _mm_mul_ps(rsq00,rinv00);
 687
 688             /* Compute parameters for interactions between i and j atoms */
 689             qq00             = _mm_mul_ps(iq0,jq0);
 690             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 691                                          vdwparam+vdwioffset0+vdwjidx0B,
 692                                          vdwparam+vdwioffset0+vdwjidx0C,
 693                                          vdwparam+vdwioffset0+vdwjidx0D,
 694                                          &c6_00,&c12_00);
 695
 696             /* Calculate table index by multiplying r with table scale and truncate to integer */
 697             rt               = _mm_mul_ps(r00,vftabscale);
 698             vfitab           = _mm_cvttps_epi32(rt);
 699 #ifdef __XOP__
 700             vfeps            = _mm_frcz_ps(rt);
 701 #else
 702             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 703 #endif
 704             twovfeps         = _mm_add_ps(vfeps,vfeps);
 705             vfitab           = _mm_slli_epi32(vfitab,3);
 706
 707             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 708             isaprod          = _mm_mul_ps(isai0,isaj0);
 709             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 710             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 711
 712             /* Calculate generalized born table index - this is a separate table from the normal one,
 713              * but we use the same procedure by multiplying r with scale and truncating to integer.
 714              */
 715             rt               = _mm_mul_ps(r00,gbscale);
 716             gbitab           = _mm_cvttps_epi32(rt);
 717 #ifdef __XOP__
 718             gbeps            = _mm_frcz_ps(rt);
 719 #else
 720             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 721 #endif
 722             gbitab           = _mm_slli_epi32(gbitab,2);
 723
 724             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 725             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 726             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 727             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 728             _MM_TRANSPOSE4_PS(Y,F,G,H);
 729             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 730             VV               = _mm_macc_ps(gbeps,Fp,Y);
 731             vgb              = _mm_mul_ps(gbqqfactor,VV);
 732
 733             twogbeps         = _mm_add_ps(gbeps,gbeps);
 734             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 735             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 736             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 737             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 738             fjptrA           = dvda+jnrA;
 739             fjptrB           = dvda+jnrB;
 740             fjptrC           = dvda+jnrC;
 741             fjptrD           = dvda+jnrD;
 742             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 743             velec            = _mm_mul_ps(qq00,rinv00);
 744             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 745
 746             /* CUBIC SPLINE TABLE DISPERSION */
 747             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 748             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 749             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 750             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 751             _MM_TRANSPOSE4_PS(Y,F,G,H);
 752             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 753             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 754             fvdw6            = _mm_mul_ps(c6_00,FF);
 755
 756             /* CUBIC SPLINE TABLE REPULSION */
 757             vfitab           = _mm_add_epi32(vfitab,ifour);
 758             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 759             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 760             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 761             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 762             _MM_TRANSPOSE4_PS(Y,F,G,H);
 763             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 764             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 765             fvdw12           = _mm_mul_ps(c12_00,FF);
 766             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 767
 768             fscal            = _mm_add_ps(felec,fvdw);
 769
 770              /* Update vectorial force */
 771             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 772             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 773             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 774
 775             fjptrA             = f+j_coord_offsetA;
 776             fjptrB             = f+j_coord_offsetB;
 777             fjptrC             = f+j_coord_offsetC;
 778             fjptrD             = f+j_coord_offsetD;
 779             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 780                                                    _mm_mul_ps(dx00,fscal),
 781                                                    _mm_mul_ps(dy00,fscal),
 782                                                    _mm_mul_ps(dz00,fscal));
 783
 784             /* Inner loop uses 85 flops */
 785         }
 786
 787         if(jidx<j_index_end)
 788         {
 789
 790             /* Get j neighbor index, and coordinate index */
 791             jnrlistA         = jjnr[jidx];
 792             jnrlistB         = jjnr[jidx+1];
 793             jnrlistC         = jjnr[jidx+2];
 794             jnrlistD         = jjnr[jidx+3];
 795             /* Sign of each element will be negative for non-real atoms.
 796              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 797              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 798              */
 799             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 800             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 801             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 802             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 803             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 804             j_coord_offsetA  = DIM*jnrA;
 805             j_coord_offsetB  = DIM*jnrB;
 806             j_coord_offsetC  = DIM*jnrC;
 807             j_coord_offsetD  = DIM*jnrD;
 808
 809             /* load j atom coordinates */
 810             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 811                                               x+j_coord_offsetC,x+j_coord_offsetD,
 812                                               &jx0,&jy0,&jz0);
 813
 814             /* Calculate displacement vector */
 815             dx00             = _mm_sub_ps(ix0,jx0);
 816             dy00             = _mm_sub_ps(iy0,jy0);
 817             dz00             = _mm_sub_ps(iz0,jz0);
 818
 819             /* Calculate squared distance and things based on it */
 820             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 821
 822             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 823
 824             /* Load parameters for j particles */
 825             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 826                                                               charge+jnrC+0,charge+jnrD+0);
 827             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 828                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 829             vdwjidx0A        = 2*vdwtype[jnrA+0];
 830             vdwjidx0B        = 2*vdwtype[jnrB+0];
 831             vdwjidx0C        = 2*vdwtype[jnrC+0];
 832             vdwjidx0D        = 2*vdwtype[jnrD+0];
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             r00              = _mm_mul_ps(rsq00,rinv00);
 839             r00              = _mm_andnot_ps(dummy_mask,r00);
 840
 841             /* Compute parameters for interactions between i and j atoms */
 842             qq00             = _mm_mul_ps(iq0,jq0);
 843             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 844                                          vdwparam+vdwioffset0+vdwjidx0B,
 845                                          vdwparam+vdwioffset0+vdwjidx0C,
 846                                          vdwparam+vdwioffset0+vdwjidx0D,
 847                                          &c6_00,&c12_00);
 848
 849             /* Calculate table index by multiplying r with table scale and truncate to integer */
 850             rt               = _mm_mul_ps(r00,vftabscale);
 851             vfitab           = _mm_cvttps_epi32(rt);
 852 #ifdef __XOP__
 853             vfeps            = _mm_frcz_ps(rt);
 854 #else
 855             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 856 #endif
 857             twovfeps         = _mm_add_ps(vfeps,vfeps);
 858             vfitab           = _mm_slli_epi32(vfitab,3);
 859
 860             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 861             isaprod          = _mm_mul_ps(isai0,isaj0);
 862             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 863             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 864
 865             /* Calculate generalized born table index - this is a separate table from the normal one,
 866              * but we use the same procedure by multiplying r with scale and truncating to integer.
 867              */
 868             rt               = _mm_mul_ps(r00,gbscale);
 869             gbitab           = _mm_cvttps_epi32(rt);
 870 #ifdef __XOP__
 871             gbeps            = _mm_frcz_ps(rt);
 872 #else
 873             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 874 #endif
 875             gbitab           = _mm_slli_epi32(gbitab,2);
 876
 877             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 878             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 879             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 880             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 881             _MM_TRANSPOSE4_PS(Y,F,G,H);
 882             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 883             VV               = _mm_macc_ps(gbeps,Fp,Y);
 884             vgb              = _mm_mul_ps(gbqqfactor,VV);
 885
 886             twogbeps         = _mm_add_ps(gbeps,gbeps);
 887             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 888             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 889             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 890             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 891             /* 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. */
 892             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 893             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 894             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 895             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 896             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 897             velec            = _mm_mul_ps(qq00,rinv00);
 898             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 899
 900             /* CUBIC SPLINE TABLE DISPERSION */
 901             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 902             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 903             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 904             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 905             _MM_TRANSPOSE4_PS(Y,F,G,H);
 906             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 907             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 908             fvdw6            = _mm_mul_ps(c6_00,FF);
 909
 910             /* CUBIC SPLINE TABLE REPULSION */
 911             vfitab           = _mm_add_epi32(vfitab,ifour);
 912             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 913             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 914             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 915             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 916             _MM_TRANSPOSE4_PS(Y,F,G,H);
 917             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 918             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 919             fvdw12           = _mm_mul_ps(c12_00,FF);
 920             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 921
 922             fscal            = _mm_add_ps(felec,fvdw);
 923
 924             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 925
 926              /* Update vectorial force */
 927             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 928             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 929             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 930
 931             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 932             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 933             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 934             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 935             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 936                                                    _mm_mul_ps(dx00,fscal),
 937                                                    _mm_mul_ps(dy00,fscal),
 938                                                    _mm_mul_ps(dz00,fscal));
 939
 940             /* Inner loop uses 86 flops */
 941         }
 942
 943         /* End of innermost loop */
 944
 945         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 946                                               f+i_coord_offset,fshift+i_shift_offset);
 947
 948         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 949         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 950
 951         /* Increment number of inner iterations */
 952         inneriter                  += j_index_end - j_index_start;
 953
 954         /* Outer loop uses 7 flops */
 955     }
 956
 957     /* Increment number of outer iterations */
 958     outeriter        += nri;
 959
 960     /* Update outer/inner flops */
 961
 962     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*86);
 963 }