src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  91     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  94     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  95     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  96     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 103     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 104     __m128i          vfitab;
 105     __m128i          ifour       = _mm_set1_epi32(4);
 106     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 107     real             *vftab;
 108     __m128           dummy_mask,cutoff_mask;
 109     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 110     __m128           one     = _mm_set1_ps(1.0);
 111     __m128           two     = _mm_set1_ps(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_ps(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     krf              = _mm_set1_ps(fr->ic->k_rf);
 126     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 127     crf              = _mm_set1_ps(fr->ic->c_rf);
 128     nvdwtype         = fr->ntype;
 129     vdwparam         = fr->nbfp;
 130     vdwtype          = mdatoms->typeA;
 131
 132     vftab            = kernel_data->table_vdw->data;
 133     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 134
 135     /* Setup water-specific parameters */
 136     inr              = nlist->iinr[0];
 137     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 138     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 139     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 140     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 141
 142     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 143     rcutoff_scalar   = fr->ic->rcoulomb;
 144     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 145     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 146
 147     /* Avoid stupid compiler warnings */
 148     jnrA = jnrB = jnrC = jnrD = 0;
 149     j_coord_offsetA = 0;
 150     j_coord_offsetB = 0;
 151     j_coord_offsetC = 0;
 152     j_coord_offsetD = 0;
 153
 154     outeriter        = 0;
 155     inneriter        = 0;
 156
 157     for(iidx=0;iidx<4*DIM;iidx++)
 158     {
 159         scratch[iidx] = 0.0;
 160     }
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 178                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 179
 180         fix0             = _mm_setzero_ps();
 181         fiy0             = _mm_setzero_ps();
 182         fiz0             = _mm_setzero_ps();
 183         fix1             = _mm_setzero_ps();
 184         fiy1             = _mm_setzero_ps();
 185         fiz1             = _mm_setzero_ps();
 186         fix2             = _mm_setzero_ps();
 187         fiy2             = _mm_setzero_ps();
 188         fiz2             = _mm_setzero_ps();
 189         fix3             = _mm_setzero_ps();
 190         fiy3             = _mm_setzero_ps();
 191         fiz3             = _mm_setzero_ps();
 192
 193         /* Reset potential sums */
 194         velecsum         = _mm_setzero_ps();
 195         vvdwsum          = _mm_setzero_ps();
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 199         {
 200
 201             /* Get j neighbor index, and coordinate index */
 202             jnrA             = jjnr[jidx];
 203             jnrB             = jjnr[jidx+1];
 204             jnrC             = jjnr[jidx+2];
 205             jnrD             = jjnr[jidx+3];
 206             j_coord_offsetA  = DIM*jnrA;
 207             j_coord_offsetB  = DIM*jnrB;
 208             j_coord_offsetC  = DIM*jnrC;
 209             j_coord_offsetD  = DIM*jnrD;
 210
 211             /* load j atom coordinates */
 212             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                               x+j_coord_offsetC,x+j_coord_offsetD,
 214                                               &jx0,&jy0,&jz0);
 215
 216             /* Calculate displacement vector */
 217             dx00             = _mm_sub_ps(ix0,jx0);
 218             dy00             = _mm_sub_ps(iy0,jy0);
 219             dz00             = _mm_sub_ps(iz0,jz0);
 220             dx10             = _mm_sub_ps(ix1,jx0);
 221             dy10             = _mm_sub_ps(iy1,jy0);
 222             dz10             = _mm_sub_ps(iz1,jz0);
 223             dx20             = _mm_sub_ps(ix2,jx0);
 224             dy20             = _mm_sub_ps(iy2,jy0);
 225             dz20             = _mm_sub_ps(iz2,jz0);
 226             dx30             = _mm_sub_ps(ix3,jx0);
 227             dy30             = _mm_sub_ps(iy3,jy0);
 228             dz30             = _mm_sub_ps(iz3,jz0);
 229
 230             /* Calculate squared distance and things based on it */
 231             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 232             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 233             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 234             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 235
 236             rinv00           = avx128fma_invsqrt_f(rsq00);
 237             rinv10           = avx128fma_invsqrt_f(rsq10);
 238             rinv20           = avx128fma_invsqrt_f(rsq20);
 239             rinv30           = avx128fma_invsqrt_f(rsq30);
 240
 241             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 242             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 243             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 244
 245             /* Load parameters for j particles */
 246             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 247                                                               charge+jnrC+0,charge+jnrD+0);
 248             vdwjidx0A        = 2*vdwtype[jnrA+0];
 249             vdwjidx0B        = 2*vdwtype[jnrB+0];
 250             vdwjidx0C        = 2*vdwtype[jnrC+0];
 251             vdwjidx0D        = 2*vdwtype[jnrD+0];
 252
 253             fjx0             = _mm_setzero_ps();
 254             fjy0             = _mm_setzero_ps();
 255             fjz0             = _mm_setzero_ps();
 256
 257             /**************************
 258              * CALCULATE INTERACTIONS *
 259              **************************/
 260
 261             r00              = _mm_mul_ps(rsq00,rinv00);
 262
 263             /* Compute parameters for interactions between i and j atoms */
 264             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 265                                          vdwparam+vdwioffset0+vdwjidx0B,
 266                                          vdwparam+vdwioffset0+vdwjidx0C,
 267                                          vdwparam+vdwioffset0+vdwjidx0D,
 268                                          &c6_00,&c12_00);
 269
 270             /* Calculate table index by multiplying r with table scale and truncate to integer */
 271             rt               = _mm_mul_ps(r00,vftabscale);
 272             vfitab           = _mm_cvttps_epi32(rt);
 273 #ifdef __XOP__
 274             vfeps            = _mm_frcz_ps(rt);
 275 #else
 276             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 277 #endif
 278             twovfeps         = _mm_add_ps(vfeps,vfeps);
 279             vfitab           = _mm_slli_epi32(vfitab,3);
 280
 281             /* CUBIC SPLINE TABLE DISPERSION */
 282             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 283             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 284             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 285             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 286             _MM_TRANSPOSE4_PS(Y,F,G,H);
 287             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 288             VV               = _mm_macc_ps(vfeps,Fp,Y);
 289             vvdw6            = _mm_mul_ps(c6_00,VV);
 290             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 291             fvdw6            = _mm_mul_ps(c6_00,FF);
 292
 293             /* CUBIC SPLINE TABLE REPULSION */
 294             vfitab           = _mm_add_epi32(vfitab,ifour);
 295             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 296             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 297             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 298             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 299             _MM_TRANSPOSE4_PS(Y,F,G,H);
 300             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 301             VV               = _mm_macc_ps(vfeps,Fp,Y);
 302             vvdw12           = _mm_mul_ps(c12_00,VV);
 303             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 304             fvdw12           = _mm_mul_ps(c12_00,FF);
 305             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 306             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 310
 311             fscal            = fvdw;
 312
 313              /* Update vectorial force */
 314             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 315             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 316             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 317
 318             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 319             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 320             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             if (gmx_mm_any_lt(rsq10,rcutoff2))
 327             {
 328
 329             /* Compute parameters for interactions between i and j atoms */
 330             qq10             = _mm_mul_ps(iq1,jq0);
 331
 332             /* REACTION-FIELD ELECTROSTATICS */
 333             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 334             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 335
 336             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 337
 338             /* Update potential sum for this i atom from the interaction with this j atom. */
 339             velec            = _mm_and_ps(velec,cutoff_mask);
 340             velecsum         = _mm_add_ps(velecsum,velec);
 341
 342             fscal            = felec;
 343
 344             fscal            = _mm_and_ps(fscal,cutoff_mask);
 345
 346              /* Update vectorial force */
 347             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 348             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 349             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 350
 351             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 352             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 353             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 354
 355             }
 356
 357             /**************************
 358              * CALCULATE INTERACTIONS *
 359              **************************/
 360
 361             if (gmx_mm_any_lt(rsq20,rcutoff2))
 362             {
 363
 364             /* Compute parameters for interactions between i and j atoms */
 365             qq20             = _mm_mul_ps(iq2,jq0);
 366
 367             /* REACTION-FIELD ELECTROSTATICS */
 368             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 369             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 370
 371             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             velec            = _mm_and_ps(velec,cutoff_mask);
 375             velecsum         = _mm_add_ps(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             fscal            = _mm_and_ps(fscal,cutoff_mask);
 380
 381              /* Update vectorial force */
 382             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 383             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 384             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 385
 386             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 387             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 388             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 389
 390             }
 391
 392             /**************************
 393              * CALCULATE INTERACTIONS *
 394              **************************/
 395
 396             if (gmx_mm_any_lt(rsq30,rcutoff2))
 397             {
 398
 399             /* Compute parameters for interactions between i and j atoms */
 400             qq30             = _mm_mul_ps(iq3,jq0);
 401
 402             /* REACTION-FIELD ELECTROSTATICS */
 403             velec            = _mm_mul_ps(qq30,_mm_sub_ps(_mm_macc_ps(krf,rsq30,rinv30),crf));
 404             felec            = _mm_mul_ps(qq30,_mm_msub_ps(rinv30,rinvsq30,krf2));
 405
 406             cutoff_mask      = _mm_cmplt_ps(rsq30,rcutoff2);
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velec            = _mm_and_ps(velec,cutoff_mask);
 410             velecsum         = _mm_add_ps(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414             fscal            = _mm_and_ps(fscal,cutoff_mask);
 415
 416              /* Update vectorial force */
 417             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 418             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 419             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 420
 421             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 422             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 423             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 424
 425             }
 426
 427             fjptrA             = f+j_coord_offsetA;
 428             fjptrB             = f+j_coord_offsetB;
 429             fjptrC             = f+j_coord_offsetC;
 430             fjptrD             = f+j_coord_offsetD;
 431
 432             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 433
 434             /* Inner loop uses 176 flops */
 435         }
 436
 437         if(jidx<j_index_end)
 438         {
 439
 440             /* Get j neighbor index, and coordinate index */
 441             jnrlistA         = jjnr[jidx];
 442             jnrlistB         = jjnr[jidx+1];
 443             jnrlistC         = jjnr[jidx+2];
 444             jnrlistD         = jjnr[jidx+3];
 445             /* Sign of each element will be negative for non-real atoms.
 446              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 447              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 448              */
 449             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 450             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 451             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 452             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 453             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 454             j_coord_offsetA  = DIM*jnrA;
 455             j_coord_offsetB  = DIM*jnrB;
 456             j_coord_offsetC  = DIM*jnrC;
 457             j_coord_offsetD  = DIM*jnrD;
 458
 459             /* load j atom coordinates */
 460             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 461                                               x+j_coord_offsetC,x+j_coord_offsetD,
 462                                               &jx0,&jy0,&jz0);
 463
 464             /* Calculate displacement vector */
 465             dx00             = _mm_sub_ps(ix0,jx0);
 466             dy00             = _mm_sub_ps(iy0,jy0);
 467             dz00             = _mm_sub_ps(iz0,jz0);
 468             dx10             = _mm_sub_ps(ix1,jx0);
 469             dy10             = _mm_sub_ps(iy1,jy0);
 470             dz10             = _mm_sub_ps(iz1,jz0);
 471             dx20             = _mm_sub_ps(ix2,jx0);
 472             dy20             = _mm_sub_ps(iy2,jy0);
 473             dz20             = _mm_sub_ps(iz2,jz0);
 474             dx30             = _mm_sub_ps(ix3,jx0);
 475             dy30             = _mm_sub_ps(iy3,jy0);
 476             dz30             = _mm_sub_ps(iz3,jz0);
 477
 478             /* Calculate squared distance and things based on it */
 479             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 480             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 481             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 482             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 483
 484             rinv00           = avx128fma_invsqrt_f(rsq00);
 485             rinv10           = avx128fma_invsqrt_f(rsq10);
 486             rinv20           = avx128fma_invsqrt_f(rsq20);
 487             rinv30           = avx128fma_invsqrt_f(rsq30);
 488
 489             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 490             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 491             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 492
 493             /* Load parameters for j particles */
 494             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 495                                                               charge+jnrC+0,charge+jnrD+0);
 496             vdwjidx0A        = 2*vdwtype[jnrA+0];
 497             vdwjidx0B        = 2*vdwtype[jnrB+0];
 498             vdwjidx0C        = 2*vdwtype[jnrC+0];
 499             vdwjidx0D        = 2*vdwtype[jnrD+0];
 500
 501             fjx0             = _mm_setzero_ps();
 502             fjy0             = _mm_setzero_ps();
 503             fjz0             = _mm_setzero_ps();
 504
 505             /**************************
 506              * CALCULATE INTERACTIONS *
 507              **************************/
 508
 509             r00              = _mm_mul_ps(rsq00,rinv00);
 510             r00              = _mm_andnot_ps(dummy_mask,r00);
 511
 512             /* Compute parameters for interactions between i and j atoms */
 513             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 514                                          vdwparam+vdwioffset0+vdwjidx0B,
 515                                          vdwparam+vdwioffset0+vdwjidx0C,
 516                                          vdwparam+vdwioffset0+vdwjidx0D,
 517                                          &c6_00,&c12_00);
 518
 519             /* Calculate table index by multiplying r with table scale and truncate to integer */
 520             rt               = _mm_mul_ps(r00,vftabscale);
 521             vfitab           = _mm_cvttps_epi32(rt);
 522 #ifdef __XOP__
 523             vfeps            = _mm_frcz_ps(rt);
 524 #else
 525             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 526 #endif
 527             twovfeps         = _mm_add_ps(vfeps,vfeps);
 528             vfitab           = _mm_slli_epi32(vfitab,3);
 529
 530             /* CUBIC SPLINE TABLE DISPERSION */
 531             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 532             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 533             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 534             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 535             _MM_TRANSPOSE4_PS(Y,F,G,H);
 536             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 537             VV               = _mm_macc_ps(vfeps,Fp,Y);
 538             vvdw6            = _mm_mul_ps(c6_00,VV);
 539             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 540             fvdw6            = _mm_mul_ps(c6_00,FF);
 541
 542             /* CUBIC SPLINE TABLE REPULSION */
 543             vfitab           = _mm_add_epi32(vfitab,ifour);
 544             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 545             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 546             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 547             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 548             _MM_TRANSPOSE4_PS(Y,F,G,H);
 549             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 550             VV               = _mm_macc_ps(vfeps,Fp,Y);
 551             vvdw12           = _mm_mul_ps(c12_00,VV);
 552             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 553             fvdw12           = _mm_mul_ps(c12_00,FF);
 554             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 555             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 556
 557             /* Update potential sum for this i atom from the interaction with this j atom. */
 558             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 559             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 560
 561             fscal            = fvdw;
 562
 563             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 564
 565              /* Update vectorial force */
 566             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 567             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 568             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 569
 570             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 571             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 572             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 573
 574             /**************************
 575              * CALCULATE INTERACTIONS *
 576              **************************/
 577
 578             if (gmx_mm_any_lt(rsq10,rcutoff2))
 579             {
 580
 581             /* Compute parameters for interactions between i and j atoms */
 582             qq10             = _mm_mul_ps(iq1,jq0);
 583
 584             /* REACTION-FIELD ELECTROSTATICS */
 585             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 586             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 587
 588             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 589
 590             /* Update potential sum for this i atom from the interaction with this j atom. */
 591             velec            = _mm_and_ps(velec,cutoff_mask);
 592             velec            = _mm_andnot_ps(dummy_mask,velec);
 593             velecsum         = _mm_add_ps(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             fscal            = _mm_and_ps(fscal,cutoff_mask);
 598
 599             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 600
 601              /* Update vectorial force */
 602             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 603             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 604             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 605
 606             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 607             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 608             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 609
 610             }
 611
 612             /**************************
 613              * CALCULATE INTERACTIONS *
 614              **************************/
 615
 616             if (gmx_mm_any_lt(rsq20,rcutoff2))
 617             {
 618
 619             /* Compute parameters for interactions between i and j atoms */
 620             qq20             = _mm_mul_ps(iq2,jq0);
 621
 622             /* REACTION-FIELD ELECTROSTATICS */
 623             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 624             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 625
 626             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _mm_and_ps(velec,cutoff_mask);
 630             velec            = _mm_andnot_ps(dummy_mask,velec);
 631             velecsum         = _mm_add_ps(velecsum,velec);
 632
 633             fscal            = felec;
 634
 635             fscal            = _mm_and_ps(fscal,cutoff_mask);
 636
 637             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 638
 639              /* Update vectorial force */
 640             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 641             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 642             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 643
 644             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 645             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 646             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 647
 648             }
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             if (gmx_mm_any_lt(rsq30,rcutoff2))
 655             {
 656
 657             /* Compute parameters for interactions between i and j atoms */
 658             qq30             = _mm_mul_ps(iq3,jq0);
 659
 660             /* REACTION-FIELD ELECTROSTATICS */
 661             velec            = _mm_mul_ps(qq30,_mm_sub_ps(_mm_macc_ps(krf,rsq30,rinv30),crf));
 662             felec            = _mm_mul_ps(qq30,_mm_msub_ps(rinv30,rinvsq30,krf2));
 663
 664             cutoff_mask      = _mm_cmplt_ps(rsq30,rcutoff2);
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _mm_and_ps(velec,cutoff_mask);
 668             velec            = _mm_andnot_ps(dummy_mask,velec);
 669             velecsum         = _mm_add_ps(velecsum,velec);
 670
 671             fscal            = felec;
 672
 673             fscal            = _mm_and_ps(fscal,cutoff_mask);
 674
 675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 676
 677              /* Update vectorial force */
 678             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 679             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 680             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 681
 682             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 683             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 684             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 685
 686             }
 687
 688             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 689             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 690             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 691             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 692
 693             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 694
 695             /* Inner loop uses 177 flops */
 696         }
 697
 698         /* End of innermost loop */
 699
 700         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 701                                               f+i_coord_offset,fshift+i_shift_offset);
 702
 703         ggid                        = gid[iidx];
 704         /* Update potential energies */
 705         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 706         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 707
 708         /* Increment number of inner iterations */
 709         inneriter                  += j_index_end - j_index_start;
 710
 711         /* Outer loop uses 26 flops */
 712     }
 713
 714     /* Increment number of outer iterations */
 715     outeriter        += nri;
 716
 717     /* Update outer/inner flops */
 718
 719     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*177);
 720 }
 721 /*
 722  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_F_avx_128_fma_single
 723  * Electrostatics interaction: ReactionField
 724  * VdW interaction:            CubicSplineTable
 725  * Geometry:                   Water4-Particle
 726  * Calculate force/pot:        Force
 727  */
 728 void
 729 nb_kernel_ElecRFCut_VdwCSTab_GeomW4P1_F_avx_128_fma_single
 730                     (t_nblist                    * gmx_restrict       nlist,
 731                      rvec                        * gmx_restrict          xx,
 732                      rvec                        * gmx_restrict          ff,
 733                      struct t_forcerec           * gmx_restrict          fr,
 734                      t_mdatoms                   * gmx_restrict     mdatoms,
 735                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 736                      t_nrnb                      * gmx_restrict        nrnb)
 737 {
 738     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 739      * just 0 for non-waters.
 740      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 741      * jnr indices corresponding to data put in the four positions in the SIMD register.
 742      */
 743     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 744     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 745     int              jnrA,jnrB,jnrC,jnrD;
 746     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 747     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 748     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 749     real             rcutoff_scalar;
 750     real             *shiftvec,*fshift,*x,*f;
 751     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 752     real             scratch[4*DIM];
 753     __m128           fscal,rcutoff,rcutoff2,jidxall;
 754     int              vdwioffset0;
 755     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 756     int              vdwioffset1;
 757     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 758     int              vdwioffset2;
 759     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 760     int              vdwioffset3;
 761     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 762     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 763     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 764     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 765     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 766     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 767     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 768     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 769     real             *charge;
 770     int              nvdwtype;
 771     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 772     int              *vdwtype;
 773     real             *vdwparam;
 774     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 775     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 776     __m128i          vfitab;
 777     __m128i          ifour       = _mm_set1_epi32(4);
 778     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 779     real             *vftab;
 780     __m128           dummy_mask,cutoff_mask;
 781     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 782     __m128           one     = _mm_set1_ps(1.0);
 783     __m128           two     = _mm_set1_ps(2.0);
 784     x                = xx[0];
 785     f                = ff[0];
 786
 787     nri              = nlist->nri;
 788     iinr             = nlist->iinr;
 789     jindex           = nlist->jindex;
 790     jjnr             = nlist->jjnr;
 791     shiftidx         = nlist->shift;
 792     gid              = nlist->gid;
 793     shiftvec         = fr->shift_vec[0];
 794     fshift           = fr->fshift[0];
 795     facel            = _mm_set1_ps(fr->ic->epsfac);
 796     charge           = mdatoms->chargeA;
 797     krf              = _mm_set1_ps(fr->ic->k_rf);
 798     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 799     crf              = _mm_set1_ps(fr->ic->c_rf);
 800     nvdwtype         = fr->ntype;
 801     vdwparam         = fr->nbfp;
 802     vdwtype          = mdatoms->typeA;
 803
 804     vftab            = kernel_data->table_vdw->data;
 805     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 806
 807     /* Setup water-specific parameters */
 808     inr              = nlist->iinr[0];
 809     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 810     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 811     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 812     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 813
 814     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 815     rcutoff_scalar   = fr->ic->rcoulomb;
 816     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 817     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 818
 819     /* Avoid stupid compiler warnings */
 820     jnrA = jnrB = jnrC = jnrD = 0;
 821     j_coord_offsetA = 0;
 822     j_coord_offsetB = 0;
 823     j_coord_offsetC = 0;
 824     j_coord_offsetD = 0;
 825
 826     outeriter        = 0;
 827     inneriter        = 0;
 828
 829     for(iidx=0;iidx<4*DIM;iidx++)
 830     {
 831         scratch[iidx] = 0.0;
 832     }
 833
 834     /* Start outer loop over neighborlists */
 835     for(iidx=0; iidx<nri; iidx++)
 836     {
 837         /* Load shift vector for this list */
 838         i_shift_offset   = DIM*shiftidx[iidx];
 839
 840         /* Load limits for loop over neighbors */
 841         j_index_start    = jindex[iidx];
 842         j_index_end      = jindex[iidx+1];
 843
 844         /* Get outer coordinate index */
 845         inr              = iinr[iidx];
 846         i_coord_offset   = DIM*inr;
 847
 848         /* Load i particle coords and add shift vector */
 849         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 850                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 851
 852         fix0             = _mm_setzero_ps();
 853         fiy0             = _mm_setzero_ps();
 854         fiz0             = _mm_setzero_ps();
 855         fix1             = _mm_setzero_ps();
 856         fiy1             = _mm_setzero_ps();
 857         fiz1             = _mm_setzero_ps();
 858         fix2             = _mm_setzero_ps();
 859         fiy2             = _mm_setzero_ps();
 860         fiz2             = _mm_setzero_ps();
 861         fix3             = _mm_setzero_ps();
 862         fiy3             = _mm_setzero_ps();
 863         fiz3             = _mm_setzero_ps();
 864
 865         /* Start inner kernel loop */
 866         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 867         {
 868
 869             /* Get j neighbor index, and coordinate index */
 870             jnrA             = jjnr[jidx];
 871             jnrB             = jjnr[jidx+1];
 872             jnrC             = jjnr[jidx+2];
 873             jnrD             = jjnr[jidx+3];
 874             j_coord_offsetA  = DIM*jnrA;
 875             j_coord_offsetB  = DIM*jnrB;
 876             j_coord_offsetC  = DIM*jnrC;
 877             j_coord_offsetD  = DIM*jnrD;
 878
 879             /* load j atom coordinates */
 880             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 881                                               x+j_coord_offsetC,x+j_coord_offsetD,
 882                                               &jx0,&jy0,&jz0);
 883
 884             /* Calculate displacement vector */
 885             dx00             = _mm_sub_ps(ix0,jx0);
 886             dy00             = _mm_sub_ps(iy0,jy0);
 887             dz00             = _mm_sub_ps(iz0,jz0);
 888             dx10             = _mm_sub_ps(ix1,jx0);
 889             dy10             = _mm_sub_ps(iy1,jy0);
 890             dz10             = _mm_sub_ps(iz1,jz0);
 891             dx20             = _mm_sub_ps(ix2,jx0);
 892             dy20             = _mm_sub_ps(iy2,jy0);
 893             dz20             = _mm_sub_ps(iz2,jz0);
 894             dx30             = _mm_sub_ps(ix3,jx0);
 895             dy30             = _mm_sub_ps(iy3,jy0);
 896             dz30             = _mm_sub_ps(iz3,jz0);
 897
 898             /* Calculate squared distance and things based on it */
 899             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 900             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 901             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 902             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 903
 904             rinv00           = avx128fma_invsqrt_f(rsq00);
 905             rinv10           = avx128fma_invsqrt_f(rsq10);
 906             rinv20           = avx128fma_invsqrt_f(rsq20);
 907             rinv30           = avx128fma_invsqrt_f(rsq30);
 908
 909             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 910             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 911             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 912
 913             /* Load parameters for j particles */
 914             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 915                                                               charge+jnrC+0,charge+jnrD+0);
 916             vdwjidx0A        = 2*vdwtype[jnrA+0];
 917             vdwjidx0B        = 2*vdwtype[jnrB+0];
 918             vdwjidx0C        = 2*vdwtype[jnrC+0];
 919             vdwjidx0D        = 2*vdwtype[jnrD+0];
 920
 921             fjx0             = _mm_setzero_ps();
 922             fjy0             = _mm_setzero_ps();
 923             fjz0             = _mm_setzero_ps();
 924
 925             /**************************
 926              * CALCULATE INTERACTIONS *
 927              **************************/
 928
 929             r00              = _mm_mul_ps(rsq00,rinv00);
 930
 931             /* Compute parameters for interactions between i and j atoms */
 932             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 933                                          vdwparam+vdwioffset0+vdwjidx0B,
 934                                          vdwparam+vdwioffset0+vdwjidx0C,
 935                                          vdwparam+vdwioffset0+vdwjidx0D,
 936                                          &c6_00,&c12_00);
 937
 938             /* Calculate table index by multiplying r with table scale and truncate to integer */
 939             rt               = _mm_mul_ps(r00,vftabscale);
 940             vfitab           = _mm_cvttps_epi32(rt);
 941 #ifdef __XOP__
 942             vfeps            = _mm_frcz_ps(rt);
 943 #else
 944             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 945 #endif
 946             twovfeps         = _mm_add_ps(vfeps,vfeps);
 947             vfitab           = _mm_slli_epi32(vfitab,3);
 948
 949             /* CUBIC SPLINE TABLE DISPERSION */
 950             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 951             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 952             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 953             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 954             _MM_TRANSPOSE4_PS(Y,F,G,H);
 955             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 956             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 957             fvdw6            = _mm_mul_ps(c6_00,FF);
 958
 959             /* CUBIC SPLINE TABLE REPULSION */
 960             vfitab           = _mm_add_epi32(vfitab,ifour);
 961             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 962             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 963             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 964             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 965             _MM_TRANSPOSE4_PS(Y,F,G,H);
 966             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 967             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 968             fvdw12           = _mm_mul_ps(c12_00,FF);
 969             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 970
 971             fscal            = fvdw;
 972
 973              /* Update vectorial force */
 974             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 975             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 976             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 977
 978             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 979             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 980             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 981
 982             /**************************
 983              * CALCULATE INTERACTIONS *
 984              **************************/
 985
 986             if (gmx_mm_any_lt(rsq10,rcutoff2))
 987             {
 988
 989             /* Compute parameters for interactions between i and j atoms */
 990             qq10             = _mm_mul_ps(iq1,jq0);
 991
 992             /* REACTION-FIELD ELECTROSTATICS */
 993             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 994
 995             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 996
 997             fscal            = felec;
 998
 999             fscal            = _mm_and_ps(fscal,cutoff_mask);
1000
1001              /* Update vectorial force */
1002             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1003             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1004             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1005
1006             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1007             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1008             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1009
1010             }
1011
1012             /**************************
1013              * CALCULATE INTERACTIONS *
1014              **************************/
1015
1016             if (gmx_mm_any_lt(rsq20,rcutoff2))
1017             {
1018
1019             /* Compute parameters for interactions between i and j atoms */
1020             qq20             = _mm_mul_ps(iq2,jq0);
1021
1022             /* REACTION-FIELD ELECTROSTATICS */
1023             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1024
1025             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1026
1027             fscal            = felec;
1028
1029             fscal            = _mm_and_ps(fscal,cutoff_mask);
1030
1031              /* Update vectorial force */
1032             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1033             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1034             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1035
1036             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1037             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1038             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1039
1040             }
1041
1042             /**************************
1043              * CALCULATE INTERACTIONS *
1044              **************************/
1045
1046             if (gmx_mm_any_lt(rsq30,rcutoff2))
1047             {
1048
1049             /* Compute parameters for interactions between i and j atoms */
1050             qq30             = _mm_mul_ps(iq3,jq0);
1051
1052             /* REACTION-FIELD ELECTROSTATICS */
1053             felec            = _mm_mul_ps(qq30,_mm_msub_ps(rinv30,rinvsq30,krf2));
1054
1055             cutoff_mask      = _mm_cmplt_ps(rsq30,rcutoff2);
1056
1057             fscal            = felec;
1058
1059             fscal            = _mm_and_ps(fscal,cutoff_mask);
1060
1061              /* Update vectorial force */
1062             fix3             = _mm_macc_ps(dx30,fscal,fix3);
1063             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
1064             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
1065
1066             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
1067             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
1068             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
1069
1070             }
1071
1072             fjptrA             = f+j_coord_offsetA;
1073             fjptrB             = f+j_coord_offsetB;
1074             fjptrC             = f+j_coord_offsetC;
1075             fjptrD             = f+j_coord_offsetD;
1076
1077             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1078
1079             /* Inner loop uses 150 flops */
1080         }
1081
1082         if(jidx<j_index_end)
1083         {
1084
1085             /* Get j neighbor index, and coordinate index */
1086             jnrlistA         = jjnr[jidx];
1087             jnrlistB         = jjnr[jidx+1];
1088             jnrlistC         = jjnr[jidx+2];
1089             jnrlistD         = jjnr[jidx+3];
1090             /* Sign of each element will be negative for non-real atoms.
1091              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1092              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1093              */
1094             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1095             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1096             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1097             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1098             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1099             j_coord_offsetA  = DIM*jnrA;
1100             j_coord_offsetB  = DIM*jnrB;
1101             j_coord_offsetC  = DIM*jnrC;
1102             j_coord_offsetD  = DIM*jnrD;
1103
1104             /* load j atom coordinates */
1105             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1106                                               x+j_coord_offsetC,x+j_coord_offsetD,
1107                                               &jx0,&jy0,&jz0);
1108
1109             /* Calculate displacement vector */
1110             dx00             = _mm_sub_ps(ix0,jx0);
1111             dy00             = _mm_sub_ps(iy0,jy0);
1112             dz00             = _mm_sub_ps(iz0,jz0);
1113             dx10             = _mm_sub_ps(ix1,jx0);
1114             dy10             = _mm_sub_ps(iy1,jy0);
1115             dz10             = _mm_sub_ps(iz1,jz0);
1116             dx20             = _mm_sub_ps(ix2,jx0);
1117             dy20             = _mm_sub_ps(iy2,jy0);
1118             dz20             = _mm_sub_ps(iz2,jz0);
1119             dx30             = _mm_sub_ps(ix3,jx0);
1120             dy30             = _mm_sub_ps(iy3,jy0);
1121             dz30             = _mm_sub_ps(iz3,jz0);
1122
1123             /* Calculate squared distance and things based on it */
1124             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1125             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1126             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1127             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
1128
1129             rinv00           = avx128fma_invsqrt_f(rsq00);
1130             rinv10           = avx128fma_invsqrt_f(rsq10);
1131             rinv20           = avx128fma_invsqrt_f(rsq20);
1132             rinv30           = avx128fma_invsqrt_f(rsq30);
1133
1134             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1135             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1136             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
1137
1138             /* Load parameters for j particles */
1139             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1140                                                               charge+jnrC+0,charge+jnrD+0);
1141             vdwjidx0A        = 2*vdwtype[jnrA+0];
1142             vdwjidx0B        = 2*vdwtype[jnrB+0];
1143             vdwjidx0C        = 2*vdwtype[jnrC+0];
1144             vdwjidx0D        = 2*vdwtype[jnrD+0];
1145
1146             fjx0             = _mm_setzero_ps();
1147             fjy0             = _mm_setzero_ps();
1148             fjz0             = _mm_setzero_ps();
1149
1150             /**************************
1151              * CALCULATE INTERACTIONS *
1152              **************************/
1153
1154             r00              = _mm_mul_ps(rsq00,rinv00);
1155             r00              = _mm_andnot_ps(dummy_mask,r00);
1156
1157             /* Compute parameters for interactions between i and j atoms */
1158             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1159                                          vdwparam+vdwioffset0+vdwjidx0B,
1160                                          vdwparam+vdwioffset0+vdwjidx0C,
1161                                          vdwparam+vdwioffset0+vdwjidx0D,
1162                                          &c6_00,&c12_00);
1163
1164             /* Calculate table index by multiplying r with table scale and truncate to integer */
1165             rt               = _mm_mul_ps(r00,vftabscale);
1166             vfitab           = _mm_cvttps_epi32(rt);
1167 #ifdef __XOP__
1168             vfeps            = _mm_frcz_ps(rt);
1169 #else
1170             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1171 #endif
1172             twovfeps         = _mm_add_ps(vfeps,vfeps);
1173             vfitab           = _mm_slli_epi32(vfitab,3);
1174
1175             /* CUBIC SPLINE TABLE DISPERSION */
1176             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1177             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1178             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1179             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1180             _MM_TRANSPOSE4_PS(Y,F,G,H);
1181             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1182             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1183             fvdw6            = _mm_mul_ps(c6_00,FF);
1184
1185             /* CUBIC SPLINE TABLE REPULSION */
1186             vfitab           = _mm_add_epi32(vfitab,ifour);
1187             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1188             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1189             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1190             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1191             _MM_TRANSPOSE4_PS(Y,F,G,H);
1192             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1193             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1194             fvdw12           = _mm_mul_ps(c12_00,FF);
1195             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1196
1197             fscal            = fvdw;
1198
1199             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1200
1201              /* Update vectorial force */
1202             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1203             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1204             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1205
1206             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1207             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1208             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1209
1210             /**************************
1211              * CALCULATE INTERACTIONS *
1212              **************************/
1213
1214             if (gmx_mm_any_lt(rsq10,rcutoff2))
1215             {
1216
1217             /* Compute parameters for interactions between i and j atoms */
1218             qq10             = _mm_mul_ps(iq1,jq0);
1219
1220             /* REACTION-FIELD ELECTROSTATICS */
1221             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1222
1223             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1224
1225             fscal            = felec;
1226
1227             fscal            = _mm_and_ps(fscal,cutoff_mask);
1228
1229             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1230
1231              /* Update vectorial force */
1232             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1233             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1234             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1235
1236             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1237             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1238             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1239
1240             }
1241
1242             /**************************
1243              * CALCULATE INTERACTIONS *
1244              **************************/
1245
1246             if (gmx_mm_any_lt(rsq20,rcutoff2))
1247             {
1248
1249             /* Compute parameters for interactions between i and j atoms */
1250             qq20             = _mm_mul_ps(iq2,jq0);
1251
1252             /* REACTION-FIELD ELECTROSTATICS */
1253             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1254
1255             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1256
1257             fscal            = felec;
1258
1259             fscal            = _mm_and_ps(fscal,cutoff_mask);
1260
1261             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1262
1263              /* Update vectorial force */
1264             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1265             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1266             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1267
1268             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1269             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1270             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1271
1272             }
1273
1274             /**************************
1275              * CALCULATE INTERACTIONS *
1276              **************************/
1277
1278             if (gmx_mm_any_lt(rsq30,rcutoff2))
1279             {
1280
1281             /* Compute parameters for interactions between i and j atoms */
1282             qq30             = _mm_mul_ps(iq3,jq0);
1283
1284             /* REACTION-FIELD ELECTROSTATICS */
1285             felec            = _mm_mul_ps(qq30,_mm_msub_ps(rinv30,rinvsq30,krf2));
1286
1287             cutoff_mask      = _mm_cmplt_ps(rsq30,rcutoff2);
1288
1289             fscal            = felec;
1290
1291             fscal            = _mm_and_ps(fscal,cutoff_mask);
1292
1293             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1294
1295              /* Update vectorial force */
1296             fix3             = _mm_macc_ps(dx30,fscal,fix3);
1297             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
1298             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
1299
1300             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
1301             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
1302             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
1303
1304             }
1305
1306             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1307             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1308             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1309             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1310
1311             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1312
1313             /* Inner loop uses 151 flops */
1314         }
1315
1316         /* End of innermost loop */
1317
1318         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1319                                               f+i_coord_offset,fshift+i_shift_offset);
1320
1321         /* Increment number of inner iterations */
1322         inneriter                  += j_index_end - j_index_start;
1323
1324         /* Outer loop uses 24 flops */
1325     }
1326
1327     /* Increment number of outer iterations */
1328     outeriter        += nri;
1329
1330     /* Update outer/inner flops */
1331
1332     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*151);
1333 }