src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_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_GeomW3P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_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              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  92     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  93     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 100     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 101     __m128i          vfitab;
 102     __m128i          ifour       = _mm_set1_epi32(4);
 103     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 104     real             *vftab;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_ps(fr->ic->k_rf);
 123     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_ps(fr->ic->c_rf);
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     vftab            = kernel_data->table_vdw->data;
 130     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 131
 132     /* Setup water-specific parameters */
 133     inr              = nlist->iinr[0];
 134     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 135     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 136     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 137     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 138
 139     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 140     rcutoff_scalar   = fr->ic->rcoulomb;
 141     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 142     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 143
 144     /* Avoid stupid compiler warnings */
 145     jnrA = jnrB = jnrC = jnrD = 0;
 146     j_coord_offsetA = 0;
 147     j_coord_offsetB = 0;
 148     j_coord_offsetC = 0;
 149     j_coord_offsetD = 0;
 150
 151     outeriter        = 0;
 152     inneriter        = 0;
 153
 154     for(iidx=0;iidx<4*DIM;iidx++)
 155     {
 156         scratch[iidx] = 0.0;
 157     }
 158
 159     /* Start outer loop over neighborlists */
 160     for(iidx=0; iidx<nri; iidx++)
 161     {
 162         /* Load shift vector for this list */
 163         i_shift_offset   = DIM*shiftidx[iidx];
 164
 165         /* Load limits for loop over neighbors */
 166         j_index_start    = jindex[iidx];
 167         j_index_end      = jindex[iidx+1];
 168
 169         /* Get outer coordinate index */
 170         inr              = iinr[iidx];
 171         i_coord_offset   = DIM*inr;
 172
 173         /* Load i particle coords and add shift vector */
 174         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 175                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 176
 177         fix0             = _mm_setzero_ps();
 178         fiy0             = _mm_setzero_ps();
 179         fiz0             = _mm_setzero_ps();
 180         fix1             = _mm_setzero_ps();
 181         fiy1             = _mm_setzero_ps();
 182         fiz1             = _mm_setzero_ps();
 183         fix2             = _mm_setzero_ps();
 184         fiy2             = _mm_setzero_ps();
 185         fiz2             = _mm_setzero_ps();
 186
 187         /* Reset potential sums */
 188         velecsum         = _mm_setzero_ps();
 189         vvdwsum          = _mm_setzero_ps();
 190
 191         /* Start inner kernel loop */
 192         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 193         {
 194
 195             /* Get j neighbor index, and coordinate index */
 196             jnrA             = jjnr[jidx];
 197             jnrB             = jjnr[jidx+1];
 198             jnrC             = jjnr[jidx+2];
 199             jnrD             = jjnr[jidx+3];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202             j_coord_offsetC  = DIM*jnrC;
 203             j_coord_offsetD  = DIM*jnrD;
 204
 205             /* load j atom coordinates */
 206             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 207                                               x+j_coord_offsetC,x+j_coord_offsetD,
 208                                               &jx0,&jy0,&jz0);
 209
 210             /* Calculate displacement vector */
 211             dx00             = _mm_sub_ps(ix0,jx0);
 212             dy00             = _mm_sub_ps(iy0,jy0);
 213             dz00             = _mm_sub_ps(iz0,jz0);
 214             dx10             = _mm_sub_ps(ix1,jx0);
 215             dy10             = _mm_sub_ps(iy1,jy0);
 216             dz10             = _mm_sub_ps(iz1,jz0);
 217             dx20             = _mm_sub_ps(ix2,jx0);
 218             dy20             = _mm_sub_ps(iy2,jy0);
 219             dz20             = _mm_sub_ps(iz2,jz0);
 220
 221             /* Calculate squared distance and things based on it */
 222             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 223             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 224             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 225
 226             rinv00           = avx128fma_invsqrt_f(rsq00);
 227             rinv10           = avx128fma_invsqrt_f(rsq10);
 228             rinv20           = avx128fma_invsqrt_f(rsq20);
 229
 230             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 231             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 232             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 233
 234             /* Load parameters for j particles */
 235             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 236                                                               charge+jnrC+0,charge+jnrD+0);
 237             vdwjidx0A        = 2*vdwtype[jnrA+0];
 238             vdwjidx0B        = 2*vdwtype[jnrB+0];
 239             vdwjidx0C        = 2*vdwtype[jnrC+0];
 240             vdwjidx0D        = 2*vdwtype[jnrD+0];
 241
 242             fjx0             = _mm_setzero_ps();
 243             fjy0             = _mm_setzero_ps();
 244             fjz0             = _mm_setzero_ps();
 245
 246             /**************************
 247              * CALCULATE INTERACTIONS *
 248              **************************/
 249
 250             if (gmx_mm_any_lt(rsq00,rcutoff2))
 251             {
 252
 253             r00              = _mm_mul_ps(rsq00,rinv00);
 254
 255             /* Compute parameters for interactions between i and j atoms */
 256             qq00             = _mm_mul_ps(iq0,jq0);
 257             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 258                                          vdwparam+vdwioffset0+vdwjidx0B,
 259                                          vdwparam+vdwioffset0+vdwjidx0C,
 260                                          vdwparam+vdwioffset0+vdwjidx0D,
 261                                          &c6_00,&c12_00);
 262
 263             /* Calculate table index by multiplying r with table scale and truncate to integer */
 264             rt               = _mm_mul_ps(r00,vftabscale);
 265             vfitab           = _mm_cvttps_epi32(rt);
 266 #ifdef __XOP__
 267             vfeps            = _mm_frcz_ps(rt);
 268 #else
 269             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 270 #endif
 271             twovfeps         = _mm_add_ps(vfeps,vfeps);
 272             vfitab           = _mm_slli_epi32(vfitab,3);
 273
 274             /* REACTION-FIELD ELECTROSTATICS */
 275             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 276             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 277
 278             /* CUBIC SPLINE TABLE DISPERSION */
 279             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 280             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 281             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 282             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 283             _MM_TRANSPOSE4_PS(Y,F,G,H);
 284             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 285             VV               = _mm_macc_ps(vfeps,Fp,Y);
 286             vvdw6            = _mm_mul_ps(c6_00,VV);
 287             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 288             fvdw6            = _mm_mul_ps(c6_00,FF);
 289
 290             /* CUBIC SPLINE TABLE REPULSION */
 291             vfitab           = _mm_add_epi32(vfitab,ifour);
 292             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 293             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 294             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 295             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 296             _MM_TRANSPOSE4_PS(Y,F,G,H);
 297             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 298             VV               = _mm_macc_ps(vfeps,Fp,Y);
 299             vvdw12           = _mm_mul_ps(c12_00,VV);
 300             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 301             fvdw12           = _mm_mul_ps(c12_00,FF);
 302             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 303             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 304
 305             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 306
 307             /* Update potential sum for this i atom from the interaction with this j atom. */
 308             velec            = _mm_and_ps(velec,cutoff_mask);
 309             velecsum         = _mm_add_ps(velecsum,velec);
 310             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 311             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 312
 313             fscal            = _mm_add_ps(felec,fvdw);
 314
 315             fscal            = _mm_and_ps(fscal,cutoff_mask);
 316
 317              /* Update vectorial force */
 318             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 319             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 320             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 321
 322             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 323             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 324             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 325
 326             }
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             if (gmx_mm_any_lt(rsq10,rcutoff2))
 333             {
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq10             = _mm_mul_ps(iq1,jq0);
 337
 338             /* REACTION-FIELD ELECTROSTATICS */
 339             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 340             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 341
 342             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_and_ps(velec,cutoff_mask);
 346             velecsum         = _mm_add_ps(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             fscal            = _mm_and_ps(fscal,cutoff_mask);
 351
 352              /* Update vectorial force */
 353             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 354             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 355             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 356
 357             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 358             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 359             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 360
 361             }
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             if (gmx_mm_any_lt(rsq20,rcutoff2))
 368             {
 369
 370             /* Compute parameters for interactions between i and j atoms */
 371             qq20             = _mm_mul_ps(iq2,jq0);
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 375             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 376
 377             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm_and_ps(velec,cutoff_mask);
 381             velecsum         = _mm_add_ps(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             fscal            = _mm_and_ps(fscal,cutoff_mask);
 386
 387              /* Update vectorial force */
 388             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 389             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 390             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 391
 392             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 393             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 394             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 395
 396             }
 397
 398             fjptrA             = f+j_coord_offsetA;
 399             fjptrB             = f+j_coord_offsetB;
 400             fjptrC             = f+j_coord_offsetC;
 401             fjptrD             = f+j_coord_offsetD;
 402
 403             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 404
 405             /* Inner loop uses 153 flops */
 406         }
 407
 408         if(jidx<j_index_end)
 409         {
 410
 411             /* Get j neighbor index, and coordinate index */
 412             jnrlistA         = jjnr[jidx];
 413             jnrlistB         = jjnr[jidx+1];
 414             jnrlistC         = jjnr[jidx+2];
 415             jnrlistD         = jjnr[jidx+3];
 416             /* Sign of each element will be negative for non-real atoms.
 417              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 418              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 419              */
 420             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 421             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 422             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 423             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 424             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 425             j_coord_offsetA  = DIM*jnrA;
 426             j_coord_offsetB  = DIM*jnrB;
 427             j_coord_offsetC  = DIM*jnrC;
 428             j_coord_offsetD  = DIM*jnrD;
 429
 430             /* load j atom coordinates */
 431             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 432                                               x+j_coord_offsetC,x+j_coord_offsetD,
 433                                               &jx0,&jy0,&jz0);
 434
 435             /* Calculate displacement vector */
 436             dx00             = _mm_sub_ps(ix0,jx0);
 437             dy00             = _mm_sub_ps(iy0,jy0);
 438             dz00             = _mm_sub_ps(iz0,jz0);
 439             dx10             = _mm_sub_ps(ix1,jx0);
 440             dy10             = _mm_sub_ps(iy1,jy0);
 441             dz10             = _mm_sub_ps(iz1,jz0);
 442             dx20             = _mm_sub_ps(ix2,jx0);
 443             dy20             = _mm_sub_ps(iy2,jy0);
 444             dz20             = _mm_sub_ps(iz2,jz0);
 445
 446             /* Calculate squared distance and things based on it */
 447             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 448             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 449             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 450
 451             rinv00           = avx128fma_invsqrt_f(rsq00);
 452             rinv10           = avx128fma_invsqrt_f(rsq10);
 453             rinv20           = avx128fma_invsqrt_f(rsq20);
 454
 455             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 456             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 457             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 458
 459             /* Load parameters for j particles */
 460             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 461                                                               charge+jnrC+0,charge+jnrD+0);
 462             vdwjidx0A        = 2*vdwtype[jnrA+0];
 463             vdwjidx0B        = 2*vdwtype[jnrB+0];
 464             vdwjidx0C        = 2*vdwtype[jnrC+0];
 465             vdwjidx0D        = 2*vdwtype[jnrD+0];
 466
 467             fjx0             = _mm_setzero_ps();
 468             fjy0             = _mm_setzero_ps();
 469             fjz0             = _mm_setzero_ps();
 470
 471             /**************************
 472              * CALCULATE INTERACTIONS *
 473              **************************/
 474
 475             if (gmx_mm_any_lt(rsq00,rcutoff2))
 476             {
 477
 478             r00              = _mm_mul_ps(rsq00,rinv00);
 479             r00              = _mm_andnot_ps(dummy_mask,r00);
 480
 481             /* Compute parameters for interactions between i and j atoms */
 482             qq00             = _mm_mul_ps(iq0,jq0);
 483             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 484                                          vdwparam+vdwioffset0+vdwjidx0B,
 485                                          vdwparam+vdwioffset0+vdwjidx0C,
 486                                          vdwparam+vdwioffset0+vdwjidx0D,
 487                                          &c6_00,&c12_00);
 488
 489             /* Calculate table index by multiplying r with table scale and truncate to integer */
 490             rt               = _mm_mul_ps(r00,vftabscale);
 491             vfitab           = _mm_cvttps_epi32(rt);
 492 #ifdef __XOP__
 493             vfeps            = _mm_frcz_ps(rt);
 494 #else
 495             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 496 #endif
 497             twovfeps         = _mm_add_ps(vfeps,vfeps);
 498             vfitab           = _mm_slli_epi32(vfitab,3);
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 502             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 503
 504             /* CUBIC SPLINE TABLE DISPERSION */
 505             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 506             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 507             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 508             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 509             _MM_TRANSPOSE4_PS(Y,F,G,H);
 510             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 511             VV               = _mm_macc_ps(vfeps,Fp,Y);
 512             vvdw6            = _mm_mul_ps(c6_00,VV);
 513             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 514             fvdw6            = _mm_mul_ps(c6_00,FF);
 515
 516             /* CUBIC SPLINE TABLE REPULSION */
 517             vfitab           = _mm_add_epi32(vfitab,ifour);
 518             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 519             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 520             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 521             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 522             _MM_TRANSPOSE4_PS(Y,F,G,H);
 523             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 524             VV               = _mm_macc_ps(vfeps,Fp,Y);
 525             vvdw12           = _mm_mul_ps(c12_00,VV);
 526             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 527             fvdw12           = _mm_mul_ps(c12_00,FF);
 528             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 529             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 530
 531             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 532
 533             /* Update potential sum for this i atom from the interaction with this j atom. */
 534             velec            = _mm_and_ps(velec,cutoff_mask);
 535             velec            = _mm_andnot_ps(dummy_mask,velec);
 536             velecsum         = _mm_add_ps(velecsum,velec);
 537             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 538             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 539             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 540
 541             fscal            = _mm_add_ps(felec,fvdw);
 542
 543             fscal            = _mm_and_ps(fscal,cutoff_mask);
 544
 545             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 546
 547              /* Update vectorial force */
 548             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 549             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 550             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 551
 552             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 553             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 554             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 555
 556             }
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             if (gmx_mm_any_lt(rsq10,rcutoff2))
 563             {
 564
 565             /* Compute parameters for interactions between i and j atoms */
 566             qq10             = _mm_mul_ps(iq1,jq0);
 567
 568             /* REACTION-FIELD ELECTROSTATICS */
 569             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 570             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 571
 572             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 573
 574             /* Update potential sum for this i atom from the interaction with this j atom. */
 575             velec            = _mm_and_ps(velec,cutoff_mask);
 576             velec            = _mm_andnot_ps(dummy_mask,velec);
 577             velecsum         = _mm_add_ps(velecsum,velec);
 578
 579             fscal            = felec;
 580
 581             fscal            = _mm_and_ps(fscal,cutoff_mask);
 582
 583             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 584
 585              /* Update vectorial force */
 586             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 587             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 588             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 589
 590             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 591             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 592             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 593
 594             }
 595
 596             /**************************
 597              * CALCULATE INTERACTIONS *
 598              **************************/
 599
 600             if (gmx_mm_any_lt(rsq20,rcutoff2))
 601             {
 602
 603             /* Compute parameters for interactions between i and j atoms */
 604             qq20             = _mm_mul_ps(iq2,jq0);
 605
 606             /* REACTION-FIELD ELECTROSTATICS */
 607             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 608             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 609
 610             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 611
 612             /* Update potential sum for this i atom from the interaction with this j atom. */
 613             velec            = _mm_and_ps(velec,cutoff_mask);
 614             velec            = _mm_andnot_ps(dummy_mask,velec);
 615             velecsum         = _mm_add_ps(velecsum,velec);
 616
 617             fscal            = felec;
 618
 619             fscal            = _mm_and_ps(fscal,cutoff_mask);
 620
 621             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 622
 623              /* Update vectorial force */
 624             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 625             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 626             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 627
 628             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 629             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 630             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 631
 632             }
 633
 634             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 635             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 636             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 637             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 638
 639             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 640
 641             /* Inner loop uses 154 flops */
 642         }
 643
 644         /* End of innermost loop */
 645
 646         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 647                                               f+i_coord_offset,fshift+i_shift_offset);
 648
 649         ggid                        = gid[iidx];
 650         /* Update potential energies */
 651         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 652         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 653
 654         /* Increment number of inner iterations */
 655         inneriter                  += j_index_end - j_index_start;
 656
 657         /* Outer loop uses 20 flops */
 658     }
 659
 660     /* Increment number of outer iterations */
 661     outeriter        += nri;
 662
 663     /* Update outer/inner flops */
 664
 665     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*154);
 666 }
 667 /*
 668  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_F_avx_128_fma_single
 669  * Electrostatics interaction: ReactionField
 670  * VdW interaction:            CubicSplineTable
 671  * Geometry:                   Water3-Particle
 672  * Calculate force/pot:        Force
 673  */
 674 void
 675 nb_kernel_ElecRFCut_VdwCSTab_GeomW3P1_F_avx_128_fma_single
 676                     (t_nblist                    * gmx_restrict       nlist,
 677                      rvec                        * gmx_restrict          xx,
 678                      rvec                        * gmx_restrict          ff,
 679                      struct t_forcerec           * gmx_restrict          fr,
 680                      t_mdatoms                   * gmx_restrict     mdatoms,
 681                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 682                      t_nrnb                      * gmx_restrict        nrnb)
 683 {
 684     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 685      * just 0 for non-waters.
 686      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 687      * jnr indices corresponding to data put in the four positions in the SIMD register.
 688      */
 689     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 690     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 691     int              jnrA,jnrB,jnrC,jnrD;
 692     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 693     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 694     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 695     real             rcutoff_scalar;
 696     real             *shiftvec,*fshift,*x,*f;
 697     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 698     real             scratch[4*DIM];
 699     __m128           fscal,rcutoff,rcutoff2,jidxall;
 700     int              vdwioffset0;
 701     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 702     int              vdwioffset1;
 703     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 704     int              vdwioffset2;
 705     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 706     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 707     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 708     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 709     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 710     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 711     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 712     real             *charge;
 713     int              nvdwtype;
 714     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 715     int              *vdwtype;
 716     real             *vdwparam;
 717     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 718     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 719     __m128i          vfitab;
 720     __m128i          ifour       = _mm_set1_epi32(4);
 721     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 722     real             *vftab;
 723     __m128           dummy_mask,cutoff_mask;
 724     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 725     __m128           one     = _mm_set1_ps(1.0);
 726     __m128           two     = _mm_set1_ps(2.0);
 727     x                = xx[0];
 728     f                = ff[0];
 729
 730     nri              = nlist->nri;
 731     iinr             = nlist->iinr;
 732     jindex           = nlist->jindex;
 733     jjnr             = nlist->jjnr;
 734     shiftidx         = nlist->shift;
 735     gid              = nlist->gid;
 736     shiftvec         = fr->shift_vec[0];
 737     fshift           = fr->fshift[0];
 738     facel            = _mm_set1_ps(fr->ic->epsfac);
 739     charge           = mdatoms->chargeA;
 740     krf              = _mm_set1_ps(fr->ic->k_rf);
 741     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 742     crf              = _mm_set1_ps(fr->ic->c_rf);
 743     nvdwtype         = fr->ntype;
 744     vdwparam         = fr->nbfp;
 745     vdwtype          = mdatoms->typeA;
 746
 747     vftab            = kernel_data->table_vdw->data;
 748     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 749
 750     /* Setup water-specific parameters */
 751     inr              = nlist->iinr[0];
 752     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 753     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 754     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 755     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 756
 757     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 758     rcutoff_scalar   = fr->ic->rcoulomb;
 759     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 760     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 761
 762     /* Avoid stupid compiler warnings */
 763     jnrA = jnrB = jnrC = jnrD = 0;
 764     j_coord_offsetA = 0;
 765     j_coord_offsetB = 0;
 766     j_coord_offsetC = 0;
 767     j_coord_offsetD = 0;
 768
 769     outeriter        = 0;
 770     inneriter        = 0;
 771
 772     for(iidx=0;iidx<4*DIM;iidx++)
 773     {
 774         scratch[iidx] = 0.0;
 775     }
 776
 777     /* Start outer loop over neighborlists */
 778     for(iidx=0; iidx<nri; iidx++)
 779     {
 780         /* Load shift vector for this list */
 781         i_shift_offset   = DIM*shiftidx[iidx];
 782
 783         /* Load limits for loop over neighbors */
 784         j_index_start    = jindex[iidx];
 785         j_index_end      = jindex[iidx+1];
 786
 787         /* Get outer coordinate index */
 788         inr              = iinr[iidx];
 789         i_coord_offset   = DIM*inr;
 790
 791         /* Load i particle coords and add shift vector */
 792         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 793                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 794
 795         fix0             = _mm_setzero_ps();
 796         fiy0             = _mm_setzero_ps();
 797         fiz0             = _mm_setzero_ps();
 798         fix1             = _mm_setzero_ps();
 799         fiy1             = _mm_setzero_ps();
 800         fiz1             = _mm_setzero_ps();
 801         fix2             = _mm_setzero_ps();
 802         fiy2             = _mm_setzero_ps();
 803         fiz2             = _mm_setzero_ps();
 804
 805         /* Start inner kernel loop */
 806         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 807         {
 808
 809             /* Get j neighbor index, and coordinate index */
 810             jnrA             = jjnr[jidx];
 811             jnrB             = jjnr[jidx+1];
 812             jnrC             = jjnr[jidx+2];
 813             jnrD             = jjnr[jidx+3];
 814             j_coord_offsetA  = DIM*jnrA;
 815             j_coord_offsetB  = DIM*jnrB;
 816             j_coord_offsetC  = DIM*jnrC;
 817             j_coord_offsetD  = DIM*jnrD;
 818
 819             /* load j atom coordinates */
 820             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 821                                               x+j_coord_offsetC,x+j_coord_offsetD,
 822                                               &jx0,&jy0,&jz0);
 823
 824             /* Calculate displacement vector */
 825             dx00             = _mm_sub_ps(ix0,jx0);
 826             dy00             = _mm_sub_ps(iy0,jy0);
 827             dz00             = _mm_sub_ps(iz0,jz0);
 828             dx10             = _mm_sub_ps(ix1,jx0);
 829             dy10             = _mm_sub_ps(iy1,jy0);
 830             dz10             = _mm_sub_ps(iz1,jz0);
 831             dx20             = _mm_sub_ps(ix2,jx0);
 832             dy20             = _mm_sub_ps(iy2,jy0);
 833             dz20             = _mm_sub_ps(iz2,jz0);
 834
 835             /* Calculate squared distance and things based on it */
 836             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 837             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 838             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 839
 840             rinv00           = avx128fma_invsqrt_f(rsq00);
 841             rinv10           = avx128fma_invsqrt_f(rsq10);
 842             rinv20           = avx128fma_invsqrt_f(rsq20);
 843
 844             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 845             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 846             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 847
 848             /* Load parameters for j particles */
 849             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 850                                                               charge+jnrC+0,charge+jnrD+0);
 851             vdwjidx0A        = 2*vdwtype[jnrA+0];
 852             vdwjidx0B        = 2*vdwtype[jnrB+0];
 853             vdwjidx0C        = 2*vdwtype[jnrC+0];
 854             vdwjidx0D        = 2*vdwtype[jnrD+0];
 855
 856             fjx0             = _mm_setzero_ps();
 857             fjy0             = _mm_setzero_ps();
 858             fjz0             = _mm_setzero_ps();
 859
 860             /**************************
 861              * CALCULATE INTERACTIONS *
 862              **************************/
 863
 864             if (gmx_mm_any_lt(rsq00,rcutoff2))
 865             {
 866
 867             r00              = _mm_mul_ps(rsq00,rinv00);
 868
 869             /* Compute parameters for interactions between i and j atoms */
 870             qq00             = _mm_mul_ps(iq0,jq0);
 871             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 872                                          vdwparam+vdwioffset0+vdwjidx0B,
 873                                          vdwparam+vdwioffset0+vdwjidx0C,
 874                                          vdwparam+vdwioffset0+vdwjidx0D,
 875                                          &c6_00,&c12_00);
 876
 877             /* Calculate table index by multiplying r with table scale and truncate to integer */
 878             rt               = _mm_mul_ps(r00,vftabscale);
 879             vfitab           = _mm_cvttps_epi32(rt);
 880 #ifdef __XOP__
 881             vfeps            = _mm_frcz_ps(rt);
 882 #else
 883             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 884 #endif
 885             twovfeps         = _mm_add_ps(vfeps,vfeps);
 886             vfitab           = _mm_slli_epi32(vfitab,3);
 887
 888             /* REACTION-FIELD ELECTROSTATICS */
 889             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 890
 891             /* CUBIC SPLINE TABLE DISPERSION */
 892             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 893             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 894             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 895             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 896             _MM_TRANSPOSE4_PS(Y,F,G,H);
 897             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 898             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 899             fvdw6            = _mm_mul_ps(c6_00,FF);
 900
 901             /* CUBIC SPLINE TABLE REPULSION */
 902             vfitab           = _mm_add_epi32(vfitab,ifour);
 903             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 904             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 905             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 906             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 907             _MM_TRANSPOSE4_PS(Y,F,G,H);
 908             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 909             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 910             fvdw12           = _mm_mul_ps(c12_00,FF);
 911             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 912
 913             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 914
 915             fscal            = _mm_add_ps(felec,fvdw);
 916
 917             fscal            = _mm_and_ps(fscal,cutoff_mask);
 918
 919              /* Update vectorial force */
 920             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 921             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 922             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 923
 924             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 925             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 926             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 927
 928             }
 929
 930             /**************************
 931              * CALCULATE INTERACTIONS *
 932              **************************/
 933
 934             if (gmx_mm_any_lt(rsq10,rcutoff2))
 935             {
 936
 937             /* Compute parameters for interactions between i and j atoms */
 938             qq10             = _mm_mul_ps(iq1,jq0);
 939
 940             /* REACTION-FIELD ELECTROSTATICS */
 941             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 942
 943             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 944
 945             fscal            = felec;
 946
 947             fscal            = _mm_and_ps(fscal,cutoff_mask);
 948
 949              /* Update vectorial force */
 950             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 951             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 952             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 953
 954             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 955             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 956             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 957
 958             }
 959
 960             /**************************
 961              * CALCULATE INTERACTIONS *
 962              **************************/
 963
 964             if (gmx_mm_any_lt(rsq20,rcutoff2))
 965             {
 966
 967             /* Compute parameters for interactions between i and j atoms */
 968             qq20             = _mm_mul_ps(iq2,jq0);
 969
 970             /* REACTION-FIELD ELECTROSTATICS */
 971             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 972
 973             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 974
 975             fscal            = felec;
 976
 977             fscal            = _mm_and_ps(fscal,cutoff_mask);
 978
 979              /* Update vectorial force */
 980             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 981             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 982             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 983
 984             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 985             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 986             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 987
 988             }
 989
 990             fjptrA             = f+j_coord_offsetA;
 991             fjptrB             = f+j_coord_offsetB;
 992             fjptrC             = f+j_coord_offsetC;
 993             fjptrD             = f+j_coord_offsetD;
 994
 995             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 996
 997             /* Inner loop uses 126 flops */
 998         }
 999
1000         if(jidx<j_index_end)
1001         {
1002
1003             /* Get j neighbor index, and coordinate index */
1004             jnrlistA         = jjnr[jidx];
1005             jnrlistB         = jjnr[jidx+1];
1006             jnrlistC         = jjnr[jidx+2];
1007             jnrlistD         = jjnr[jidx+3];
1008             /* Sign of each element will be negative for non-real atoms.
1009              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1010              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1011              */
1012             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1013             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1014             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1015             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1016             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1017             j_coord_offsetA  = DIM*jnrA;
1018             j_coord_offsetB  = DIM*jnrB;
1019             j_coord_offsetC  = DIM*jnrC;
1020             j_coord_offsetD  = DIM*jnrD;
1021
1022             /* load j atom coordinates */
1023             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1024                                               x+j_coord_offsetC,x+j_coord_offsetD,
1025                                               &jx0,&jy0,&jz0);
1026
1027             /* Calculate displacement vector */
1028             dx00             = _mm_sub_ps(ix0,jx0);
1029             dy00             = _mm_sub_ps(iy0,jy0);
1030             dz00             = _mm_sub_ps(iz0,jz0);
1031             dx10             = _mm_sub_ps(ix1,jx0);
1032             dy10             = _mm_sub_ps(iy1,jy0);
1033             dz10             = _mm_sub_ps(iz1,jz0);
1034             dx20             = _mm_sub_ps(ix2,jx0);
1035             dy20             = _mm_sub_ps(iy2,jy0);
1036             dz20             = _mm_sub_ps(iz2,jz0);
1037
1038             /* Calculate squared distance and things based on it */
1039             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1040             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1041             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1042
1043             rinv00           = avx128fma_invsqrt_f(rsq00);
1044             rinv10           = avx128fma_invsqrt_f(rsq10);
1045             rinv20           = avx128fma_invsqrt_f(rsq20);
1046
1047             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1048             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1049             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1050
1051             /* Load parameters for j particles */
1052             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1053                                                               charge+jnrC+0,charge+jnrD+0);
1054             vdwjidx0A        = 2*vdwtype[jnrA+0];
1055             vdwjidx0B        = 2*vdwtype[jnrB+0];
1056             vdwjidx0C        = 2*vdwtype[jnrC+0];
1057             vdwjidx0D        = 2*vdwtype[jnrD+0];
1058
1059             fjx0             = _mm_setzero_ps();
1060             fjy0             = _mm_setzero_ps();
1061             fjz0             = _mm_setzero_ps();
1062
1063             /**************************
1064              * CALCULATE INTERACTIONS *
1065              **************************/
1066
1067             if (gmx_mm_any_lt(rsq00,rcutoff2))
1068             {
1069
1070             r00              = _mm_mul_ps(rsq00,rinv00);
1071             r00              = _mm_andnot_ps(dummy_mask,r00);
1072
1073             /* Compute parameters for interactions between i and j atoms */
1074             qq00             = _mm_mul_ps(iq0,jq0);
1075             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1076                                          vdwparam+vdwioffset0+vdwjidx0B,
1077                                          vdwparam+vdwioffset0+vdwjidx0C,
1078                                          vdwparam+vdwioffset0+vdwjidx0D,
1079                                          &c6_00,&c12_00);
1080
1081             /* Calculate table index by multiplying r with table scale and truncate to integer */
1082             rt               = _mm_mul_ps(r00,vftabscale);
1083             vfitab           = _mm_cvttps_epi32(rt);
1084 #ifdef __XOP__
1085             vfeps            = _mm_frcz_ps(rt);
1086 #else
1087             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1088 #endif
1089             twovfeps         = _mm_add_ps(vfeps,vfeps);
1090             vfitab           = _mm_slli_epi32(vfitab,3);
1091
1092             /* REACTION-FIELD ELECTROSTATICS */
1093             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1094
1095             /* CUBIC SPLINE TABLE DISPERSION */
1096             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1097             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1098             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1099             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1100             _MM_TRANSPOSE4_PS(Y,F,G,H);
1101             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1102             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1103             fvdw6            = _mm_mul_ps(c6_00,FF);
1104
1105             /* CUBIC SPLINE TABLE REPULSION */
1106             vfitab           = _mm_add_epi32(vfitab,ifour);
1107             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1108             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1109             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1110             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1111             _MM_TRANSPOSE4_PS(Y,F,G,H);
1112             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1113             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1114             fvdw12           = _mm_mul_ps(c12_00,FF);
1115             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1116
1117             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1118
1119             fscal            = _mm_add_ps(felec,fvdw);
1120
1121             fscal            = _mm_and_ps(fscal,cutoff_mask);
1122
1123             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1124
1125              /* Update vectorial force */
1126             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1127             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1128             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1129
1130             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1131             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1132             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1133
1134             }
1135
1136             /**************************
1137              * CALCULATE INTERACTIONS *
1138              **************************/
1139
1140             if (gmx_mm_any_lt(rsq10,rcutoff2))
1141             {
1142
1143             /* Compute parameters for interactions between i and j atoms */
1144             qq10             = _mm_mul_ps(iq1,jq0);
1145
1146             /* REACTION-FIELD ELECTROSTATICS */
1147             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1148
1149             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1150
1151             fscal            = felec;
1152
1153             fscal            = _mm_and_ps(fscal,cutoff_mask);
1154
1155             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1156
1157              /* Update vectorial force */
1158             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1159             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1160             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1161
1162             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1163             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1164             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1165
1166             }
1167
1168             /**************************
1169              * CALCULATE INTERACTIONS *
1170              **************************/
1171
1172             if (gmx_mm_any_lt(rsq20,rcutoff2))
1173             {
1174
1175             /* Compute parameters for interactions between i and j atoms */
1176             qq20             = _mm_mul_ps(iq2,jq0);
1177
1178             /* REACTION-FIELD ELECTROSTATICS */
1179             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1180
1181             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1182
1183             fscal            = felec;
1184
1185             fscal            = _mm_and_ps(fscal,cutoff_mask);
1186
1187             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1188
1189              /* Update vectorial force */
1190             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1191             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1192             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1193
1194             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1195             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1196             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1197
1198             }
1199
1200             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1201             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1202             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1203             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1204
1205             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1206
1207             /* Inner loop uses 127 flops */
1208         }
1209
1210         /* End of innermost loop */
1211
1212         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1213                                               f+i_coord_offset,fshift+i_shift_offset);
1214
1215         /* Increment number of inner iterations */
1216         inneriter                  += j_index_end - j_index_start;
1217
1218         /* Outer loop uses 18 flops */
1219     }
1220
1221     /* Increment number of outer iterations */
1222     outeriter        += nri;
1223
1224     /* Update outer/inner flops */
1225
1226     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*127);
1227 }