src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRFCut_VdwLJSw_GeomW3P1_avx_256_double.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_256_double 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_256_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3P1_VF_avx_256_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSw_GeomW3P1_VF_avx_256_double
  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, 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              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     real *           vdwioffsetptr0;
  84     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     real *           vdwioffsetptr1;
  86     __m256d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     real *           vdwioffsetptr2;
  88     __m256d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m256d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  93     __m256d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  94     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  95     real             *charge;
  96     int              nvdwtype;
  97     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  98     int              *vdwtype;
  99     real             *vdwparam;
 100     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 101     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 102     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 103     real             rswitch_scalar,d_scalar;
 104     __m256d          dummy_mask,cutoff_mask;
 105     __m128           tmpmask0,tmpmask1;
 106     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 107     __m256d          one     = _mm256_set1_pd(1.0);
 108     __m256d          two     = _mm256_set1_pd(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            = _mm256_set1_pd(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm256_set1_pd(fr->ic->k_rf);
 123     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 124     crf              = _mm256_set1_pd(fr->ic->c_rf);
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 132     iq1              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
 133     iq2              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
 134     vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 135
 136     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 137     rcutoff_scalar   = fr->ic->rcoulomb;
 138     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 139     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 140
 141     rswitch_scalar   = fr->ic->rvdw_switch;
 142     rswitch          = _mm256_set1_pd(rswitch_scalar);
 143     /* Setup switch parameters */
 144     d_scalar         = rcutoff_scalar-rswitch_scalar;
 145     d                = _mm256_set1_pd(d_scalar);
 146     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 147     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 148     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 149     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 150     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 151     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 152
 153     /* Avoid stupid compiler warnings */
 154     jnrA = jnrB = jnrC = jnrD = 0;
 155     j_coord_offsetA = 0;
 156     j_coord_offsetB = 0;
 157     j_coord_offsetC = 0;
 158     j_coord_offsetD = 0;
 159
 160     outeriter        = 0;
 161     inneriter        = 0;
 162
 163     for(iidx=0;iidx<4*DIM;iidx++)
 164     {
 165         scratch[iidx] = 0.0;
 166     }
 167
 168     /* Start outer loop over neighborlists */
 169     for(iidx=0; iidx<nri; iidx++)
 170     {
 171         /* Load shift vector for this list */
 172         i_shift_offset   = DIM*shiftidx[iidx];
 173
 174         /* Load limits for loop over neighbors */
 175         j_index_start    = jindex[iidx];
 176         j_index_end      = jindex[iidx+1];
 177
 178         /* Get outer coordinate index */
 179         inr              = iinr[iidx];
 180         i_coord_offset   = DIM*inr;
 181
 182         /* Load i particle coords and add shift vector */
 183         gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 184                                                     &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 185
 186         fix0             = _mm256_setzero_pd();
 187         fiy0             = _mm256_setzero_pd();
 188         fiz0             = _mm256_setzero_pd();
 189         fix1             = _mm256_setzero_pd();
 190         fiy1             = _mm256_setzero_pd();
 191         fiz1             = _mm256_setzero_pd();
 192         fix2             = _mm256_setzero_pd();
 193         fiy2             = _mm256_setzero_pd();
 194         fiz2             = _mm256_setzero_pd();
 195
 196         /* Reset potential sums */
 197         velecsum         = _mm256_setzero_pd();
 198         vvdwsum          = _mm256_setzero_pd();
 199
 200         /* Start inner kernel loop */
 201         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 202         {
 203
 204             /* Get j neighbor index, and coordinate index */
 205             jnrA             = jjnr[jidx];
 206             jnrB             = jjnr[jidx+1];
 207             jnrC             = jjnr[jidx+2];
 208             jnrD             = jjnr[jidx+3];
 209             j_coord_offsetA  = DIM*jnrA;
 210             j_coord_offsetB  = DIM*jnrB;
 211             j_coord_offsetC  = DIM*jnrC;
 212             j_coord_offsetD  = DIM*jnrD;
 213
 214             /* load j atom coordinates */
 215             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 216                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 217                                                  &jx0,&jy0,&jz0);
 218
 219             /* Calculate displacement vector */
 220             dx00             = _mm256_sub_pd(ix0,jx0);
 221             dy00             = _mm256_sub_pd(iy0,jy0);
 222             dz00             = _mm256_sub_pd(iz0,jz0);
 223             dx10             = _mm256_sub_pd(ix1,jx0);
 224             dy10             = _mm256_sub_pd(iy1,jy0);
 225             dz10             = _mm256_sub_pd(iz1,jz0);
 226             dx20             = _mm256_sub_pd(ix2,jx0);
 227             dy20             = _mm256_sub_pd(iy2,jy0);
 228             dz20             = _mm256_sub_pd(iz2,jz0);
 229
 230             /* Calculate squared distance and things based on it */
 231             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 232             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
 233             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
 234
 235             rinv00           = avx256_invsqrt_d(rsq00);
 236             rinv10           = avx256_invsqrt_d(rsq10);
 237             rinv20           = avx256_invsqrt_d(rsq20);
 238
 239             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 240             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
 241             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
 242
 243             /* Load parameters for j particles */
 244             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 245                                                                  charge+jnrC+0,charge+jnrD+0);
 246             vdwjidx0A        = 2*vdwtype[jnrA+0];
 247             vdwjidx0B        = 2*vdwtype[jnrB+0];
 248             vdwjidx0C        = 2*vdwtype[jnrC+0];
 249             vdwjidx0D        = 2*vdwtype[jnrD+0];
 250
 251             fjx0             = _mm256_setzero_pd();
 252             fjy0             = _mm256_setzero_pd();
 253             fjz0             = _mm256_setzero_pd();
 254
 255             /**************************
 256              * CALCULATE INTERACTIONS *
 257              **************************/
 258
 259             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 260             {
 261
 262             r00              = _mm256_mul_pd(rsq00,rinv00);
 263
 264             /* Compute parameters for interactions between i and j atoms */
 265             qq00             = _mm256_mul_pd(iq0,jq0);
 266             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 267                                             vdwioffsetptr0+vdwjidx0B,
 268                                             vdwioffsetptr0+vdwjidx0C,
 269                                             vdwioffsetptr0+vdwjidx0D,
 270                                             &c6_00,&c12_00);
 271
 272             /* REACTION-FIELD ELECTROSTATICS */
 273             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 274             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 275
 276             /* LENNARD-JONES DISPERSION/REPULSION */
 277
 278             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 279             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 280             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 281             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 282             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 283
 284             d                = _mm256_sub_pd(r00,rswitch);
 285             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 286             d2               = _mm256_mul_pd(d,d);
 287             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 288
 289             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 290
 291             /* Evaluate switch function */
 292             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 293             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 294             vvdw             = _mm256_mul_pd(vvdw,sw);
 295             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velec            = _mm256_and_pd(velec,cutoff_mask);
 299             velecsum         = _mm256_add_pd(velecsum,velec);
 300             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 301             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 302
 303             fscal            = _mm256_add_pd(felec,fvdw);
 304
 305             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 306
 307             /* Calculate temporary vectorial force */
 308             tx               = _mm256_mul_pd(fscal,dx00);
 309             ty               = _mm256_mul_pd(fscal,dy00);
 310             tz               = _mm256_mul_pd(fscal,dz00);
 311
 312             /* Update vectorial force */
 313             fix0             = _mm256_add_pd(fix0,tx);
 314             fiy0             = _mm256_add_pd(fiy0,ty);
 315             fiz0             = _mm256_add_pd(fiz0,tz);
 316
 317             fjx0             = _mm256_add_pd(fjx0,tx);
 318             fjy0             = _mm256_add_pd(fjy0,ty);
 319             fjz0             = _mm256_add_pd(fjz0,tz);
 320
 321             }
 322
 323             /**************************
 324              * CALCULATE INTERACTIONS *
 325              **************************/
 326
 327             if (gmx_mm256_any_lt(rsq10,rcutoff2))
 328             {
 329
 330             /* Compute parameters for interactions between i and j atoms */
 331             qq10             = _mm256_mul_pd(iq1,jq0);
 332
 333             /* REACTION-FIELD ELECTROSTATICS */
 334             velec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
 335             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
 336
 337             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
 338
 339             /* Update potential sum for this i atom from the interaction with this j atom. */
 340             velec            = _mm256_and_pd(velec,cutoff_mask);
 341             velecsum         = _mm256_add_pd(velecsum,velec);
 342
 343             fscal            = felec;
 344
 345             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 346
 347             /* Calculate temporary vectorial force */
 348             tx               = _mm256_mul_pd(fscal,dx10);
 349             ty               = _mm256_mul_pd(fscal,dy10);
 350             tz               = _mm256_mul_pd(fscal,dz10);
 351
 352             /* Update vectorial force */
 353             fix1             = _mm256_add_pd(fix1,tx);
 354             fiy1             = _mm256_add_pd(fiy1,ty);
 355             fiz1             = _mm256_add_pd(fiz1,tz);
 356
 357             fjx0             = _mm256_add_pd(fjx0,tx);
 358             fjy0             = _mm256_add_pd(fjy0,ty);
 359             fjz0             = _mm256_add_pd(fjz0,tz);
 360
 361             }
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             if (gmx_mm256_any_lt(rsq20,rcutoff2))
 368             {
 369
 370             /* Compute parameters for interactions between i and j atoms */
 371             qq20             = _mm256_mul_pd(iq2,jq0);
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
 375             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
 376
 377             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm256_and_pd(velec,cutoff_mask);
 381             velecsum         = _mm256_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 386
 387             /* Calculate temporary vectorial force */
 388             tx               = _mm256_mul_pd(fscal,dx20);
 389             ty               = _mm256_mul_pd(fscal,dy20);
 390             tz               = _mm256_mul_pd(fscal,dz20);
 391
 392             /* Update vectorial force */
 393             fix2             = _mm256_add_pd(fix2,tx);
 394             fiy2             = _mm256_add_pd(fiy2,ty);
 395             fiz2             = _mm256_add_pd(fiz2,tz);
 396
 397             fjx0             = _mm256_add_pd(fjx0,tx);
 398             fjy0             = _mm256_add_pd(fjy0,ty);
 399             fjz0             = _mm256_add_pd(fjz0,tz);
 400
 401             }
 402
 403             fjptrA             = f+j_coord_offsetA;
 404             fjptrB             = f+j_coord_offsetB;
 405             fjptrC             = f+j_coord_offsetC;
 406             fjptrD             = f+j_coord_offsetD;
 407
 408             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 409
 410             /* Inner loop uses 145 flops */
 411         }
 412
 413         if(jidx<j_index_end)
 414         {
 415
 416             /* Get j neighbor index, and coordinate index */
 417             jnrlistA         = jjnr[jidx];
 418             jnrlistB         = jjnr[jidx+1];
 419             jnrlistC         = jjnr[jidx+2];
 420             jnrlistD         = jjnr[jidx+3];
 421             /* Sign of each element will be negative for non-real atoms.
 422              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 423              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 424              */
 425             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 426
 427             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 428             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 429             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 430
 431             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 432             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 433             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 434             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 435             j_coord_offsetA  = DIM*jnrA;
 436             j_coord_offsetB  = DIM*jnrB;
 437             j_coord_offsetC  = DIM*jnrC;
 438             j_coord_offsetD  = DIM*jnrD;
 439
 440             /* load j atom coordinates */
 441             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 442                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 443                                                  &jx0,&jy0,&jz0);
 444
 445             /* Calculate displacement vector */
 446             dx00             = _mm256_sub_pd(ix0,jx0);
 447             dy00             = _mm256_sub_pd(iy0,jy0);
 448             dz00             = _mm256_sub_pd(iz0,jz0);
 449             dx10             = _mm256_sub_pd(ix1,jx0);
 450             dy10             = _mm256_sub_pd(iy1,jy0);
 451             dz10             = _mm256_sub_pd(iz1,jz0);
 452             dx20             = _mm256_sub_pd(ix2,jx0);
 453             dy20             = _mm256_sub_pd(iy2,jy0);
 454             dz20             = _mm256_sub_pd(iz2,jz0);
 455
 456             /* Calculate squared distance and things based on it */
 457             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 458             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
 459             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
 460
 461             rinv00           = avx256_invsqrt_d(rsq00);
 462             rinv10           = avx256_invsqrt_d(rsq10);
 463             rinv20           = avx256_invsqrt_d(rsq20);
 464
 465             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 466             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
 467             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
 468
 469             /* Load parameters for j particles */
 470             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 471                                                                  charge+jnrC+0,charge+jnrD+0);
 472             vdwjidx0A        = 2*vdwtype[jnrA+0];
 473             vdwjidx0B        = 2*vdwtype[jnrB+0];
 474             vdwjidx0C        = 2*vdwtype[jnrC+0];
 475             vdwjidx0D        = 2*vdwtype[jnrD+0];
 476
 477             fjx0             = _mm256_setzero_pd();
 478             fjy0             = _mm256_setzero_pd();
 479             fjz0             = _mm256_setzero_pd();
 480
 481             /**************************
 482              * CALCULATE INTERACTIONS *
 483              **************************/
 484
 485             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 486             {
 487
 488             r00              = _mm256_mul_pd(rsq00,rinv00);
 489             r00              = _mm256_andnot_pd(dummy_mask,r00);
 490
 491             /* Compute parameters for interactions between i and j atoms */
 492             qq00             = _mm256_mul_pd(iq0,jq0);
 493             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 494                                             vdwioffsetptr0+vdwjidx0B,
 495                                             vdwioffsetptr0+vdwjidx0C,
 496                                             vdwioffsetptr0+vdwjidx0D,
 497                                             &c6_00,&c12_00);
 498
 499             /* REACTION-FIELD ELECTROSTATICS */
 500             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 501             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 502
 503             /* LENNARD-JONES DISPERSION/REPULSION */
 504
 505             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 506             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 507             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 508             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 509             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 510
 511             d                = _mm256_sub_pd(r00,rswitch);
 512             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 513             d2               = _mm256_mul_pd(d,d);
 514             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 515
 516             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 517
 518             /* Evaluate switch function */
 519             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 520             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 521             vvdw             = _mm256_mul_pd(vvdw,sw);
 522             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 523
 524             /* Update potential sum for this i atom from the interaction with this j atom. */
 525             velec            = _mm256_and_pd(velec,cutoff_mask);
 526             velec            = _mm256_andnot_pd(dummy_mask,velec);
 527             velecsum         = _mm256_add_pd(velecsum,velec);
 528             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 529             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 530             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 531
 532             fscal            = _mm256_add_pd(felec,fvdw);
 533
 534             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 535
 536             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 537
 538             /* Calculate temporary vectorial force */
 539             tx               = _mm256_mul_pd(fscal,dx00);
 540             ty               = _mm256_mul_pd(fscal,dy00);
 541             tz               = _mm256_mul_pd(fscal,dz00);
 542
 543             /* Update vectorial force */
 544             fix0             = _mm256_add_pd(fix0,tx);
 545             fiy0             = _mm256_add_pd(fiy0,ty);
 546             fiz0             = _mm256_add_pd(fiz0,tz);
 547
 548             fjx0             = _mm256_add_pd(fjx0,tx);
 549             fjy0             = _mm256_add_pd(fjy0,ty);
 550             fjz0             = _mm256_add_pd(fjz0,tz);
 551
 552             }
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             if (gmx_mm256_any_lt(rsq10,rcutoff2))
 559             {
 560
 561             /* Compute parameters for interactions between i and j atoms */
 562             qq10             = _mm256_mul_pd(iq1,jq0);
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
 566             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
 567
 568             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _mm256_and_pd(velec,cutoff_mask);
 572             velec            = _mm256_andnot_pd(dummy_mask,velec);
 573             velecsum         = _mm256_add_pd(velecsum,velec);
 574
 575             fscal            = felec;
 576
 577             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 578
 579             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 580
 581             /* Calculate temporary vectorial force */
 582             tx               = _mm256_mul_pd(fscal,dx10);
 583             ty               = _mm256_mul_pd(fscal,dy10);
 584             tz               = _mm256_mul_pd(fscal,dz10);
 585
 586             /* Update vectorial force */
 587             fix1             = _mm256_add_pd(fix1,tx);
 588             fiy1             = _mm256_add_pd(fiy1,ty);
 589             fiz1             = _mm256_add_pd(fiz1,tz);
 590
 591             fjx0             = _mm256_add_pd(fjx0,tx);
 592             fjy0             = _mm256_add_pd(fjy0,ty);
 593             fjz0             = _mm256_add_pd(fjz0,tz);
 594
 595             }
 596
 597             /**************************
 598              * CALCULATE INTERACTIONS *
 599              **************************/
 600
 601             if (gmx_mm256_any_lt(rsq20,rcutoff2))
 602             {
 603
 604             /* Compute parameters for interactions between i and j atoms */
 605             qq20             = _mm256_mul_pd(iq2,jq0);
 606
 607             /* REACTION-FIELD ELECTROSTATICS */
 608             velec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
 609             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
 610
 611             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
 612
 613             /* Update potential sum for this i atom from the interaction with this j atom. */
 614             velec            = _mm256_and_pd(velec,cutoff_mask);
 615             velec            = _mm256_andnot_pd(dummy_mask,velec);
 616             velecsum         = _mm256_add_pd(velecsum,velec);
 617
 618             fscal            = felec;
 619
 620             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 621
 622             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 623
 624             /* Calculate temporary vectorial force */
 625             tx               = _mm256_mul_pd(fscal,dx20);
 626             ty               = _mm256_mul_pd(fscal,dy20);
 627             tz               = _mm256_mul_pd(fscal,dz20);
 628
 629             /* Update vectorial force */
 630             fix2             = _mm256_add_pd(fix2,tx);
 631             fiy2             = _mm256_add_pd(fiy2,ty);
 632             fiz2             = _mm256_add_pd(fiz2,tz);
 633
 634             fjx0             = _mm256_add_pd(fjx0,tx);
 635             fjy0             = _mm256_add_pd(fjy0,ty);
 636             fjz0             = _mm256_add_pd(fjz0,tz);
 637
 638             }
 639
 640             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 641             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 642             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 643             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 644
 645             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 646
 647             /* Inner loop uses 146 flops */
 648         }
 649
 650         /* End of innermost loop */
 651
 652         gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 653                                                  f+i_coord_offset,fshift+i_shift_offset);
 654
 655         ggid                        = gid[iidx];
 656         /* Update potential energies */
 657         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 658         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 659
 660         /* Increment number of inner iterations */
 661         inneriter                  += j_index_end - j_index_start;
 662
 663         /* Outer loop uses 20 flops */
 664     }
 665
 666     /* Increment number of outer iterations */
 667     outeriter        += nri;
 668
 669     /* Update outer/inner flops */
 670
 671     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*146);
 672 }
 673 /*
 674  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3P1_F_avx_256_double
 675  * Electrostatics interaction: ReactionField
 676  * VdW interaction:            LennardJones
 677  * Geometry:                   Water3-Particle
 678  * Calculate force/pot:        Force
 679  */
 680 void
 681 nb_kernel_ElecRFCut_VdwLJSw_GeomW3P1_F_avx_256_double
 682                     (t_nblist                    * gmx_restrict       nlist,
 683                      rvec                        * gmx_restrict          xx,
 684                      rvec                        * gmx_restrict          ff,
 685                      struct t_forcerec           * gmx_restrict          fr,
 686                      t_mdatoms                   * gmx_restrict     mdatoms,
 687                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 688                      t_nrnb                      * gmx_restrict        nrnb)
 689 {
 690     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 691      * just 0 for non-waters.
 692      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 693      * jnr indices corresponding to data put in the four positions in the SIMD register.
 694      */
 695     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 696     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 697     int              jnrA,jnrB,jnrC,jnrD;
 698     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 699     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 700     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 701     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 702     real             rcutoff_scalar;
 703     real             *shiftvec,*fshift,*x,*f;
 704     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 705     real             scratch[4*DIM];
 706     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 707     real *           vdwioffsetptr0;
 708     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 709     real *           vdwioffsetptr1;
 710     __m256d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 711     real *           vdwioffsetptr2;
 712     __m256d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 713     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 714     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 715     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 716     __m256d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 717     __m256d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 718     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 719     real             *charge;
 720     int              nvdwtype;
 721     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 722     int              *vdwtype;
 723     real             *vdwparam;
 724     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 725     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 726     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 727     real             rswitch_scalar,d_scalar;
 728     __m256d          dummy_mask,cutoff_mask;
 729     __m128           tmpmask0,tmpmask1;
 730     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 731     __m256d          one     = _mm256_set1_pd(1.0);
 732     __m256d          two     = _mm256_set1_pd(2.0);
 733     x                = xx[0];
 734     f                = ff[0];
 735
 736     nri              = nlist->nri;
 737     iinr             = nlist->iinr;
 738     jindex           = nlist->jindex;
 739     jjnr             = nlist->jjnr;
 740     shiftidx         = nlist->shift;
 741     gid              = nlist->gid;
 742     shiftvec         = fr->shift_vec[0];
 743     fshift           = fr->fshift[0];
 744     facel            = _mm256_set1_pd(fr->ic->epsfac);
 745     charge           = mdatoms->chargeA;
 746     krf              = _mm256_set1_pd(fr->ic->k_rf);
 747     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 748     crf              = _mm256_set1_pd(fr->ic->c_rf);
 749     nvdwtype         = fr->ntype;
 750     vdwparam         = fr->nbfp;
 751     vdwtype          = mdatoms->typeA;
 752
 753     /* Setup water-specific parameters */
 754     inr              = nlist->iinr[0];
 755     iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 756     iq1              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
 757     iq2              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
 758     vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 759
 760     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 761     rcutoff_scalar   = fr->ic->rcoulomb;
 762     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 763     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 764
 765     rswitch_scalar   = fr->ic->rvdw_switch;
 766     rswitch          = _mm256_set1_pd(rswitch_scalar);
 767     /* Setup switch parameters */
 768     d_scalar         = rcutoff_scalar-rswitch_scalar;
 769     d                = _mm256_set1_pd(d_scalar);
 770     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 771     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 772     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 773     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 774     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 775     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 776
 777     /* Avoid stupid compiler warnings */
 778     jnrA = jnrB = jnrC = jnrD = 0;
 779     j_coord_offsetA = 0;
 780     j_coord_offsetB = 0;
 781     j_coord_offsetC = 0;
 782     j_coord_offsetD = 0;
 783
 784     outeriter        = 0;
 785     inneriter        = 0;
 786
 787     for(iidx=0;iidx<4*DIM;iidx++)
 788     {
 789         scratch[iidx] = 0.0;
 790     }
 791
 792     /* Start outer loop over neighborlists */
 793     for(iidx=0; iidx<nri; iidx++)
 794     {
 795         /* Load shift vector for this list */
 796         i_shift_offset   = DIM*shiftidx[iidx];
 797
 798         /* Load limits for loop over neighbors */
 799         j_index_start    = jindex[iidx];
 800         j_index_end      = jindex[iidx+1];
 801
 802         /* Get outer coordinate index */
 803         inr              = iinr[iidx];
 804         i_coord_offset   = DIM*inr;
 805
 806         /* Load i particle coords and add shift vector */
 807         gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 808                                                     &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 809
 810         fix0             = _mm256_setzero_pd();
 811         fiy0             = _mm256_setzero_pd();
 812         fiz0             = _mm256_setzero_pd();
 813         fix1             = _mm256_setzero_pd();
 814         fiy1             = _mm256_setzero_pd();
 815         fiz1             = _mm256_setzero_pd();
 816         fix2             = _mm256_setzero_pd();
 817         fiy2             = _mm256_setzero_pd();
 818         fiz2             = _mm256_setzero_pd();
 819
 820         /* Start inner kernel loop */
 821         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 822         {
 823
 824             /* Get j neighbor index, and coordinate index */
 825             jnrA             = jjnr[jidx];
 826             jnrB             = jjnr[jidx+1];
 827             jnrC             = jjnr[jidx+2];
 828             jnrD             = jjnr[jidx+3];
 829             j_coord_offsetA  = DIM*jnrA;
 830             j_coord_offsetB  = DIM*jnrB;
 831             j_coord_offsetC  = DIM*jnrC;
 832             j_coord_offsetD  = DIM*jnrD;
 833
 834             /* load j atom coordinates */
 835             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 836                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 837                                                  &jx0,&jy0,&jz0);
 838
 839             /* Calculate displacement vector */
 840             dx00             = _mm256_sub_pd(ix0,jx0);
 841             dy00             = _mm256_sub_pd(iy0,jy0);
 842             dz00             = _mm256_sub_pd(iz0,jz0);
 843             dx10             = _mm256_sub_pd(ix1,jx0);
 844             dy10             = _mm256_sub_pd(iy1,jy0);
 845             dz10             = _mm256_sub_pd(iz1,jz0);
 846             dx20             = _mm256_sub_pd(ix2,jx0);
 847             dy20             = _mm256_sub_pd(iy2,jy0);
 848             dz20             = _mm256_sub_pd(iz2,jz0);
 849
 850             /* Calculate squared distance and things based on it */
 851             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 852             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
 853             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
 854
 855             rinv00           = avx256_invsqrt_d(rsq00);
 856             rinv10           = avx256_invsqrt_d(rsq10);
 857             rinv20           = avx256_invsqrt_d(rsq20);
 858
 859             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 860             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
 861             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
 862
 863             /* Load parameters for j particles */
 864             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 865                                                                  charge+jnrC+0,charge+jnrD+0);
 866             vdwjidx0A        = 2*vdwtype[jnrA+0];
 867             vdwjidx0B        = 2*vdwtype[jnrB+0];
 868             vdwjidx0C        = 2*vdwtype[jnrC+0];
 869             vdwjidx0D        = 2*vdwtype[jnrD+0];
 870
 871             fjx0             = _mm256_setzero_pd();
 872             fjy0             = _mm256_setzero_pd();
 873             fjz0             = _mm256_setzero_pd();
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 880             {
 881
 882             r00              = _mm256_mul_pd(rsq00,rinv00);
 883
 884             /* Compute parameters for interactions between i and j atoms */
 885             qq00             = _mm256_mul_pd(iq0,jq0);
 886             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 887                                             vdwioffsetptr0+vdwjidx0B,
 888                                             vdwioffsetptr0+vdwjidx0C,
 889                                             vdwioffsetptr0+vdwjidx0D,
 890                                             &c6_00,&c12_00);
 891
 892             /* REACTION-FIELD ELECTROSTATICS */
 893             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 894
 895             /* LENNARD-JONES DISPERSION/REPULSION */
 896
 897             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 898             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 899             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 900             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 901             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 902
 903             d                = _mm256_sub_pd(r00,rswitch);
 904             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 905             d2               = _mm256_mul_pd(d,d);
 906             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 907
 908             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 909
 910             /* Evaluate switch function */
 911             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 912             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 913             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 914
 915             fscal            = _mm256_add_pd(felec,fvdw);
 916
 917             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 918
 919             /* Calculate temporary vectorial force */
 920             tx               = _mm256_mul_pd(fscal,dx00);
 921             ty               = _mm256_mul_pd(fscal,dy00);
 922             tz               = _mm256_mul_pd(fscal,dz00);
 923
 924             /* Update vectorial force */
 925             fix0             = _mm256_add_pd(fix0,tx);
 926             fiy0             = _mm256_add_pd(fiy0,ty);
 927             fiz0             = _mm256_add_pd(fiz0,tz);
 928
 929             fjx0             = _mm256_add_pd(fjx0,tx);
 930             fjy0             = _mm256_add_pd(fjy0,ty);
 931             fjz0             = _mm256_add_pd(fjz0,tz);
 932
 933             }
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             if (gmx_mm256_any_lt(rsq10,rcutoff2))
 940             {
 941
 942             /* Compute parameters for interactions between i and j atoms */
 943             qq10             = _mm256_mul_pd(iq1,jq0);
 944
 945             /* REACTION-FIELD ELECTROSTATICS */
 946             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
 947
 948             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
 949
 950             fscal            = felec;
 951
 952             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 953
 954             /* Calculate temporary vectorial force */
 955             tx               = _mm256_mul_pd(fscal,dx10);
 956             ty               = _mm256_mul_pd(fscal,dy10);
 957             tz               = _mm256_mul_pd(fscal,dz10);
 958
 959             /* Update vectorial force */
 960             fix1             = _mm256_add_pd(fix1,tx);
 961             fiy1             = _mm256_add_pd(fiy1,ty);
 962             fiz1             = _mm256_add_pd(fiz1,tz);
 963
 964             fjx0             = _mm256_add_pd(fjx0,tx);
 965             fjy0             = _mm256_add_pd(fjy0,ty);
 966             fjz0             = _mm256_add_pd(fjz0,tz);
 967
 968             }
 969
 970             /**************************
 971              * CALCULATE INTERACTIONS *
 972              **************************/
 973
 974             if (gmx_mm256_any_lt(rsq20,rcutoff2))
 975             {
 976
 977             /* Compute parameters for interactions between i and j atoms */
 978             qq20             = _mm256_mul_pd(iq2,jq0);
 979
 980             /* REACTION-FIELD ELECTROSTATICS */
 981             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
 982
 983             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
 984
 985             fscal            = felec;
 986
 987             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 988
 989             /* Calculate temporary vectorial force */
 990             tx               = _mm256_mul_pd(fscal,dx20);
 991             ty               = _mm256_mul_pd(fscal,dy20);
 992             tz               = _mm256_mul_pd(fscal,dz20);
 993
 994             /* Update vectorial force */
 995             fix2             = _mm256_add_pd(fix2,tx);
 996             fiy2             = _mm256_add_pd(fiy2,ty);
 997             fiz2             = _mm256_add_pd(fiz2,tz);
 998
 999             fjx0             = _mm256_add_pd(fjx0,tx);
1000             fjy0             = _mm256_add_pd(fjy0,ty);
1001             fjz0             = _mm256_add_pd(fjz0,tz);
1002
1003             }
1004
1005             fjptrA             = f+j_coord_offsetA;
1006             fjptrB             = f+j_coord_offsetB;
1007             fjptrC             = f+j_coord_offsetC;
1008             fjptrD             = f+j_coord_offsetD;
1009
1010             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1011
1012             /* Inner loop uses 124 flops */
1013         }
1014
1015         if(jidx<j_index_end)
1016         {
1017
1018             /* Get j neighbor index, and coordinate index */
1019             jnrlistA         = jjnr[jidx];
1020             jnrlistB         = jjnr[jidx+1];
1021             jnrlistC         = jjnr[jidx+2];
1022             jnrlistD         = jjnr[jidx+3];
1023             /* Sign of each element will be negative for non-real atoms.
1024              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1025              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
1026              */
1027             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1028
1029             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
1030             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
1031             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
1032
1033             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1034             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1035             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1036             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1037             j_coord_offsetA  = DIM*jnrA;
1038             j_coord_offsetB  = DIM*jnrB;
1039             j_coord_offsetC  = DIM*jnrC;
1040             j_coord_offsetD  = DIM*jnrD;
1041
1042             /* load j atom coordinates */
1043             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1044                                                  x+j_coord_offsetC,x+j_coord_offsetD,
1045                                                  &jx0,&jy0,&jz0);
1046
1047             /* Calculate displacement vector */
1048             dx00             = _mm256_sub_pd(ix0,jx0);
1049             dy00             = _mm256_sub_pd(iy0,jy0);
1050             dz00             = _mm256_sub_pd(iz0,jz0);
1051             dx10             = _mm256_sub_pd(ix1,jx0);
1052             dy10             = _mm256_sub_pd(iy1,jy0);
1053             dz10             = _mm256_sub_pd(iz1,jz0);
1054             dx20             = _mm256_sub_pd(ix2,jx0);
1055             dy20             = _mm256_sub_pd(iy2,jy0);
1056             dz20             = _mm256_sub_pd(iz2,jz0);
1057
1058             /* Calculate squared distance and things based on it */
1059             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
1060             rsq10            = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
1061             rsq20            = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
1062
1063             rinv00           = avx256_invsqrt_d(rsq00);
1064             rinv10           = avx256_invsqrt_d(rsq10);
1065             rinv20           = avx256_invsqrt_d(rsq20);
1066
1067             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
1068             rinvsq10         = _mm256_mul_pd(rinv10,rinv10);
1069             rinvsq20         = _mm256_mul_pd(rinv20,rinv20);
1070
1071             /* Load parameters for j particles */
1072             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
1073                                                                  charge+jnrC+0,charge+jnrD+0);
1074             vdwjidx0A        = 2*vdwtype[jnrA+0];
1075             vdwjidx0B        = 2*vdwtype[jnrB+0];
1076             vdwjidx0C        = 2*vdwtype[jnrC+0];
1077             vdwjidx0D        = 2*vdwtype[jnrD+0];
1078
1079             fjx0             = _mm256_setzero_pd();
1080             fjy0             = _mm256_setzero_pd();
1081             fjz0             = _mm256_setzero_pd();
1082
1083             /**************************
1084              * CALCULATE INTERACTIONS *
1085              **************************/
1086
1087             if (gmx_mm256_any_lt(rsq00,rcutoff2))
1088             {
1089
1090             r00              = _mm256_mul_pd(rsq00,rinv00);
1091             r00              = _mm256_andnot_pd(dummy_mask,r00);
1092
1093             /* Compute parameters for interactions between i and j atoms */
1094             qq00             = _mm256_mul_pd(iq0,jq0);
1095             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
1096                                             vdwioffsetptr0+vdwjidx0B,
1097                                             vdwioffsetptr0+vdwjidx0C,
1098                                             vdwioffsetptr0+vdwjidx0D,
1099                                             &c6_00,&c12_00);
1100
1101             /* REACTION-FIELD ELECTROSTATICS */
1102             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
1103
1104             /* LENNARD-JONES DISPERSION/REPULSION */
1105
1106             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1107             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
1108             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
1109             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
1110             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
1111
1112             d                = _mm256_sub_pd(r00,rswitch);
1113             d                = _mm256_max_pd(d,_mm256_setzero_pd());
1114             d2               = _mm256_mul_pd(d,d);
1115             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
1116
1117             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
1118
1119             /* Evaluate switch function */
1120             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1121             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
1122             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
1123
1124             fscal            = _mm256_add_pd(felec,fvdw);
1125
1126             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1127
1128             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
1129
1130             /* Calculate temporary vectorial force */
1131             tx               = _mm256_mul_pd(fscal,dx00);
1132             ty               = _mm256_mul_pd(fscal,dy00);
1133             tz               = _mm256_mul_pd(fscal,dz00);
1134
1135             /* Update vectorial force */
1136             fix0             = _mm256_add_pd(fix0,tx);
1137             fiy0             = _mm256_add_pd(fiy0,ty);
1138             fiz0             = _mm256_add_pd(fiz0,tz);
1139
1140             fjx0             = _mm256_add_pd(fjx0,tx);
1141             fjy0             = _mm256_add_pd(fjy0,ty);
1142             fjz0             = _mm256_add_pd(fjz0,tz);
1143
1144             }
1145
1146             /**************************
1147              * CALCULATE INTERACTIONS *
1148              **************************/
1149
1150             if (gmx_mm256_any_lt(rsq10,rcutoff2))
1151             {
1152
1153             /* Compute parameters for interactions between i and j atoms */
1154             qq10             = _mm256_mul_pd(iq1,jq0);
1155
1156             /* REACTION-FIELD ELECTROSTATICS */
1157             felec            = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
1158
1159             cutoff_mask      = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1160
1161             fscal            = felec;
1162
1163             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1164
1165             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
1166
1167             /* Calculate temporary vectorial force */
1168             tx               = _mm256_mul_pd(fscal,dx10);
1169             ty               = _mm256_mul_pd(fscal,dy10);
1170             tz               = _mm256_mul_pd(fscal,dz10);
1171
1172             /* Update vectorial force */
1173             fix1             = _mm256_add_pd(fix1,tx);
1174             fiy1             = _mm256_add_pd(fiy1,ty);
1175             fiz1             = _mm256_add_pd(fiz1,tz);
1176
1177             fjx0             = _mm256_add_pd(fjx0,tx);
1178             fjy0             = _mm256_add_pd(fjy0,ty);
1179             fjz0             = _mm256_add_pd(fjz0,tz);
1180
1181             }
1182
1183             /**************************
1184              * CALCULATE INTERACTIONS *
1185              **************************/
1186
1187             if (gmx_mm256_any_lt(rsq20,rcutoff2))
1188             {
1189
1190             /* Compute parameters for interactions between i and j atoms */
1191             qq20             = _mm256_mul_pd(iq2,jq0);
1192
1193             /* REACTION-FIELD ELECTROSTATICS */
1194             felec            = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
1195
1196             cutoff_mask      = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1197
1198             fscal            = felec;
1199
1200             fscal            = _mm256_and_pd(fscal,cutoff_mask);
1201
1202             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
1203
1204             /* Calculate temporary vectorial force */
1205             tx               = _mm256_mul_pd(fscal,dx20);
1206             ty               = _mm256_mul_pd(fscal,dy20);
1207             tz               = _mm256_mul_pd(fscal,dz20);
1208
1209             /* Update vectorial force */
1210             fix2             = _mm256_add_pd(fix2,tx);
1211             fiy2             = _mm256_add_pd(fiy2,ty);
1212             fiz2             = _mm256_add_pd(fiz2,tz);
1213
1214             fjx0             = _mm256_add_pd(fjx0,tx);
1215             fjy0             = _mm256_add_pd(fjy0,ty);
1216             fjz0             = _mm256_add_pd(fjz0,tz);
1217
1218             }
1219
1220             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1221             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1222             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1223             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1224
1225             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1226
1227             /* Inner loop uses 125 flops */
1228         }
1229
1230         /* End of innermost loop */
1231
1232         gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1233                                                  f+i_coord_offset,fshift+i_shift_offset);
1234
1235         /* Increment number of inner iterations */
1236         inneriter                  += j_index_end - j_index_start;
1237
1238         /* Outer loop uses 18 flops */
1239     }
1240
1241     /* Increment number of outer iterations */
1242     outeriter        += nri;
1243
1244     /* Update outer/inner flops */
1245
1246     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*125);
1247 }