src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  91     int              nvdwtype;
  92     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  96     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  97     __m256d          dummy_mask,cutoff_mask;
  98     __m128           tmpmask0,tmpmask1;
  99     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 100     __m256d          one     = _mm256_set1_pd(1.0);
 101     __m256d          two     = _mm256_set1_pd(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm256_set1_pd(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115     krf              = _mm256_set1_pd(fr->ic->k_rf);
 116     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 117     crf              = _mm256_set1_pd(fr->ic->c_rf);
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 123     rcutoff_scalar   = fr->rcoulomb;
 124     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 125     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 126
 127     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 128     rvdw             = _mm256_set1_pd(fr->rvdw);
 129
 130     /* Avoid stupid compiler warnings */
 131     jnrA = jnrB = jnrC = jnrD = 0;
 132     j_coord_offsetA = 0;
 133     j_coord_offsetB = 0;
 134     j_coord_offsetC = 0;
 135     j_coord_offsetD = 0;
 136
 137     outeriter        = 0;
 138     inneriter        = 0;
 139
 140     for(iidx=0;iidx<4*DIM;iidx++)
 141     {
 142         scratch[iidx] = 0.0;
 143     }
 144
 145     /* Start outer loop over neighborlists */
 146     for(iidx=0; iidx<nri; iidx++)
 147     {
 148         /* Load shift vector for this list */
 149         i_shift_offset   = DIM*shiftidx[iidx];
 150
 151         /* Load limits for loop over neighbors */
 152         j_index_start    = jindex[iidx];
 153         j_index_end      = jindex[iidx+1];
 154
 155         /* Get outer coordinate index */
 156         inr              = iinr[iidx];
 157         i_coord_offset   = DIM*inr;
 158
 159         /* Load i particle coords and add shift vector */
 160         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 161
 162         fix0             = _mm256_setzero_pd();
 163         fiy0             = _mm256_setzero_pd();
 164         fiz0             = _mm256_setzero_pd();
 165
 166         /* Load parameters for i particles */
 167         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 168         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 169
 170         /* Reset potential sums */
 171         velecsum         = _mm256_setzero_pd();
 172         vvdwsum          = _mm256_setzero_pd();
 173
 174         /* Start inner kernel loop */
 175         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 176         {
 177
 178             /* Get j neighbor index, and coordinate index */
 179             jnrA             = jjnr[jidx];
 180             jnrB             = jjnr[jidx+1];
 181             jnrC             = jjnr[jidx+2];
 182             jnrD             = jjnr[jidx+3];
 183             j_coord_offsetA  = DIM*jnrA;
 184             j_coord_offsetB  = DIM*jnrB;
 185             j_coord_offsetC  = DIM*jnrC;
 186             j_coord_offsetD  = DIM*jnrD;
 187
 188             /* load j atom coordinates */
 189             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 190                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 191                                                  &jx0,&jy0,&jz0);
 192
 193             /* Calculate displacement vector */
 194             dx00             = _mm256_sub_pd(ix0,jx0);
 195             dy00             = _mm256_sub_pd(iy0,jy0);
 196             dz00             = _mm256_sub_pd(iz0,jz0);
 197
 198             /* Calculate squared distance and things based on it */
 199             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 200
 201             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 202
 203             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 204
 205             /* Load parameters for j particles */
 206             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 207                                                                  charge+jnrC+0,charge+jnrD+0);
 208             vdwjidx0A        = 2*vdwtype[jnrA+0];
 209             vdwjidx0B        = 2*vdwtype[jnrB+0];
 210             vdwjidx0C        = 2*vdwtype[jnrC+0];
 211             vdwjidx0D        = 2*vdwtype[jnrD+0];
 212
 213             /**************************
 214              * CALCULATE INTERACTIONS *
 215              **************************/
 216
 217             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 218             {
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq00             = _mm256_mul_pd(iq0,jq0);
 222             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 223                                             vdwioffsetptr0+vdwjidx0B,
 224                                             vdwioffsetptr0+vdwjidx0C,
 225                                             vdwioffsetptr0+vdwjidx0D,
 226                                             &c6_00,&c12_00);
 227
 228             /* REACTION-FIELD ELECTROSTATICS */
 229             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 230             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 231
 232             /* LENNARD-JONES DISPERSION/REPULSION */
 233
 234             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 235             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 236             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 237             vvdw             = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_00,_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 238                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 239             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 240
 241             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 242
 243             /* Update potential sum for this i atom from the interaction with this j atom. */
 244             velec            = _mm256_and_pd(velec,cutoff_mask);
 245             velecsum         = _mm256_add_pd(velecsum,velec);
 246             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 247             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 248
 249             fscal            = _mm256_add_pd(felec,fvdw);
 250
 251             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 252
 253             /* Calculate temporary vectorial force */
 254             tx               = _mm256_mul_pd(fscal,dx00);
 255             ty               = _mm256_mul_pd(fscal,dy00);
 256             tz               = _mm256_mul_pd(fscal,dz00);
 257
 258             /* Update vectorial force */
 259             fix0             = _mm256_add_pd(fix0,tx);
 260             fiy0             = _mm256_add_pd(fiy0,ty);
 261             fiz0             = _mm256_add_pd(fiz0,tz);
 262
 263             fjptrA             = f+j_coord_offsetA;
 264             fjptrB             = f+j_coord_offsetB;
 265             fjptrC             = f+j_coord_offsetC;
 266             fjptrD             = f+j_coord_offsetD;
 267             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 268
 269             }
 270
 271             /* Inner loop uses 54 flops */
 272         }
 273
 274         if(jidx<j_index_end)
 275         {
 276
 277             /* Get j neighbor index, and coordinate index */
 278             jnrlistA         = jjnr[jidx];
 279             jnrlistB         = jjnr[jidx+1];
 280             jnrlistC         = jjnr[jidx+2];
 281             jnrlistD         = jjnr[jidx+3];
 282             /* Sign of each element will be negative for non-real atoms.
 283              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 284              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 285              */
 286             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 287
 288             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 289             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 290             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 291
 292             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 293             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 294             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 295             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 296             j_coord_offsetA  = DIM*jnrA;
 297             j_coord_offsetB  = DIM*jnrB;
 298             j_coord_offsetC  = DIM*jnrC;
 299             j_coord_offsetD  = DIM*jnrD;
 300
 301             /* load j atom coordinates */
 302             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 303                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 304                                                  &jx0,&jy0,&jz0);
 305
 306             /* Calculate displacement vector */
 307             dx00             = _mm256_sub_pd(ix0,jx0);
 308             dy00             = _mm256_sub_pd(iy0,jy0);
 309             dz00             = _mm256_sub_pd(iz0,jz0);
 310
 311             /* Calculate squared distance and things based on it */
 312             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 313
 314             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 315
 316             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 317
 318             /* Load parameters for j particles */
 319             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 320                                                                  charge+jnrC+0,charge+jnrD+0);
 321             vdwjidx0A        = 2*vdwtype[jnrA+0];
 322             vdwjidx0B        = 2*vdwtype[jnrB+0];
 323             vdwjidx0C        = 2*vdwtype[jnrC+0];
 324             vdwjidx0D        = 2*vdwtype[jnrD+0];
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 331             {
 332
 333             /* Compute parameters for interactions between i and j atoms */
 334             qq00             = _mm256_mul_pd(iq0,jq0);
 335             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 336                                             vdwioffsetptr0+vdwjidx0B,
 337                                             vdwioffsetptr0+vdwjidx0C,
 338                                             vdwioffsetptr0+vdwjidx0D,
 339                                             &c6_00,&c12_00);
 340
 341             /* REACTION-FIELD ELECTROSTATICS */
 342             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 343             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 344
 345             /* LENNARD-JONES DISPERSION/REPULSION */
 346
 347             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 348             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 349             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 350             vvdw             = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_00,_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 351                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 352             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 353
 354             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velec            = _mm256_and_pd(velec,cutoff_mask);
 358             velec            = _mm256_andnot_pd(dummy_mask,velec);
 359             velecsum         = _mm256_add_pd(velecsum,velec);
 360             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 361             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 362             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 363
 364             fscal            = _mm256_add_pd(felec,fvdw);
 365
 366             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 367
 368             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 369
 370             /* Calculate temporary vectorial force */
 371             tx               = _mm256_mul_pd(fscal,dx00);
 372             ty               = _mm256_mul_pd(fscal,dy00);
 373             tz               = _mm256_mul_pd(fscal,dz00);
 374
 375             /* Update vectorial force */
 376             fix0             = _mm256_add_pd(fix0,tx);
 377             fiy0             = _mm256_add_pd(fiy0,ty);
 378             fiz0             = _mm256_add_pd(fiz0,tz);
 379
 380             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 381             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 382             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 383             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 384             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 385
 386             }
 387
 388             /* Inner loop uses 54 flops */
 389         }
 390
 391         /* End of innermost loop */
 392
 393         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 394                                                  f+i_coord_offset,fshift+i_shift_offset);
 395
 396         ggid                        = gid[iidx];
 397         /* Update potential energies */
 398         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 399         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 400
 401         /* Increment number of inner iterations */
 402         inneriter                  += j_index_end - j_index_start;
 403
 404         /* Outer loop uses 9 flops */
 405     }
 406
 407     /* Increment number of outer iterations */
 408     outeriter        += nri;
 409
 410     /* Update outer/inner flops */
 411
 412     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 413 }
 414 /*
 415  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_256_double
 416  * Electrostatics interaction: ReactionField
 417  * VdW interaction:            LennardJones
 418  * Geometry:                   Particle-Particle
 419  * Calculate force/pot:        Force
 420  */
 421 void
 422 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_256_double
 423                     (t_nblist                    * gmx_restrict       nlist,
 424                      rvec                        * gmx_restrict          xx,
 425                      rvec                        * gmx_restrict          ff,
 426                      t_forcerec                  * gmx_restrict          fr,
 427                      t_mdatoms                   * gmx_restrict     mdatoms,
 428                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 429                      t_nrnb                      * gmx_restrict        nrnb)
 430 {
 431     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 432      * just 0 for non-waters.
 433      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 434      * jnr indices corresponding to data put in the four positions in the SIMD register.
 435      */
 436     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 437     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 438     int              jnrA,jnrB,jnrC,jnrD;
 439     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 440     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 441     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 442     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 443     real             rcutoff_scalar;
 444     real             *shiftvec,*fshift,*x,*f;
 445     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 446     real             scratch[4*DIM];
 447     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 448     real *           vdwioffsetptr0;
 449     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 450     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 451     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 452     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 453     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 454     real             *charge;
 455     int              nvdwtype;
 456     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 457     int              *vdwtype;
 458     real             *vdwparam;
 459     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 460     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 461     __m256d          dummy_mask,cutoff_mask;
 462     __m128           tmpmask0,tmpmask1;
 463     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 464     __m256d          one     = _mm256_set1_pd(1.0);
 465     __m256d          two     = _mm256_set1_pd(2.0);
 466     x                = xx[0];
 467     f                = ff[0];
 468
 469     nri              = nlist->nri;
 470     iinr             = nlist->iinr;
 471     jindex           = nlist->jindex;
 472     jjnr             = nlist->jjnr;
 473     shiftidx         = nlist->shift;
 474     gid              = nlist->gid;
 475     shiftvec         = fr->shift_vec[0];
 476     fshift           = fr->fshift[0];
 477     facel            = _mm256_set1_pd(fr->epsfac);
 478     charge           = mdatoms->chargeA;
 479     krf              = _mm256_set1_pd(fr->ic->k_rf);
 480     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 481     crf              = _mm256_set1_pd(fr->ic->c_rf);
 482     nvdwtype         = fr->ntype;
 483     vdwparam         = fr->nbfp;
 484     vdwtype          = mdatoms->typeA;
 485
 486     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 487     rcutoff_scalar   = fr->rcoulomb;
 488     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 489     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 490
 491     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 492     rvdw             = _mm256_set1_pd(fr->rvdw);
 493
 494     /* Avoid stupid compiler warnings */
 495     jnrA = jnrB = jnrC = jnrD = 0;
 496     j_coord_offsetA = 0;
 497     j_coord_offsetB = 0;
 498     j_coord_offsetC = 0;
 499     j_coord_offsetD = 0;
 500
 501     outeriter        = 0;
 502     inneriter        = 0;
 503
 504     for(iidx=0;iidx<4*DIM;iidx++)
 505     {
 506         scratch[iidx] = 0.0;
 507     }
 508
 509     /* Start outer loop over neighborlists */
 510     for(iidx=0; iidx<nri; iidx++)
 511     {
 512         /* Load shift vector for this list */
 513         i_shift_offset   = DIM*shiftidx[iidx];
 514
 515         /* Load limits for loop over neighbors */
 516         j_index_start    = jindex[iidx];
 517         j_index_end      = jindex[iidx+1];
 518
 519         /* Get outer coordinate index */
 520         inr              = iinr[iidx];
 521         i_coord_offset   = DIM*inr;
 522
 523         /* Load i particle coords and add shift vector */
 524         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 525
 526         fix0             = _mm256_setzero_pd();
 527         fiy0             = _mm256_setzero_pd();
 528         fiz0             = _mm256_setzero_pd();
 529
 530         /* Load parameters for i particles */
 531         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 532         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 533
 534         /* Start inner kernel loop */
 535         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 536         {
 537
 538             /* Get j neighbor index, and coordinate index */
 539             jnrA             = jjnr[jidx];
 540             jnrB             = jjnr[jidx+1];
 541             jnrC             = jjnr[jidx+2];
 542             jnrD             = jjnr[jidx+3];
 543             j_coord_offsetA  = DIM*jnrA;
 544             j_coord_offsetB  = DIM*jnrB;
 545             j_coord_offsetC  = DIM*jnrC;
 546             j_coord_offsetD  = DIM*jnrD;
 547
 548             /* load j atom coordinates */
 549             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 550                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 551                                                  &jx0,&jy0,&jz0);
 552
 553             /* Calculate displacement vector */
 554             dx00             = _mm256_sub_pd(ix0,jx0);
 555             dy00             = _mm256_sub_pd(iy0,jy0);
 556             dz00             = _mm256_sub_pd(iz0,jz0);
 557
 558             /* Calculate squared distance and things based on it */
 559             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 560
 561             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 562
 563             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 564
 565             /* Load parameters for j particles */
 566             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 567                                                                  charge+jnrC+0,charge+jnrD+0);
 568             vdwjidx0A        = 2*vdwtype[jnrA+0];
 569             vdwjidx0B        = 2*vdwtype[jnrB+0];
 570             vdwjidx0C        = 2*vdwtype[jnrC+0];
 571             vdwjidx0D        = 2*vdwtype[jnrD+0];
 572
 573             /**************************
 574              * CALCULATE INTERACTIONS *
 575              **************************/
 576
 577             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 578             {
 579
 580             /* Compute parameters for interactions between i and j atoms */
 581             qq00             = _mm256_mul_pd(iq0,jq0);
 582             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 583                                             vdwioffsetptr0+vdwjidx0B,
 584                                             vdwioffsetptr0+vdwjidx0C,
 585                                             vdwioffsetptr0+vdwjidx0D,
 586                                             &c6_00,&c12_00);
 587
 588             /* REACTION-FIELD ELECTROSTATICS */
 589             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 590
 591             /* LENNARD-JONES DISPERSION/REPULSION */
 592
 593             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 594             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 595
 596             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 597
 598             fscal            = _mm256_add_pd(felec,fvdw);
 599
 600             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 601
 602             /* Calculate temporary vectorial force */
 603             tx               = _mm256_mul_pd(fscal,dx00);
 604             ty               = _mm256_mul_pd(fscal,dy00);
 605             tz               = _mm256_mul_pd(fscal,dz00);
 606
 607             /* Update vectorial force */
 608             fix0             = _mm256_add_pd(fix0,tx);
 609             fiy0             = _mm256_add_pd(fiy0,ty);
 610             fiz0             = _mm256_add_pd(fiz0,tz);
 611
 612             fjptrA             = f+j_coord_offsetA;
 613             fjptrB             = f+j_coord_offsetB;
 614             fjptrC             = f+j_coord_offsetC;
 615             fjptrD             = f+j_coord_offsetD;
 616             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 617
 618             }
 619
 620             /* Inner loop uses 37 flops */
 621         }
 622
 623         if(jidx<j_index_end)
 624         {
 625
 626             /* Get j neighbor index, and coordinate index */
 627             jnrlistA         = jjnr[jidx];
 628             jnrlistB         = jjnr[jidx+1];
 629             jnrlistC         = jjnr[jidx+2];
 630             jnrlistD         = jjnr[jidx+3];
 631             /* Sign of each element will be negative for non-real atoms.
 632              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 633              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 634              */
 635             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 636
 637             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 638             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 639             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 640
 641             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 642             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 643             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 644             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 645             j_coord_offsetA  = DIM*jnrA;
 646             j_coord_offsetB  = DIM*jnrB;
 647             j_coord_offsetC  = DIM*jnrC;
 648             j_coord_offsetD  = DIM*jnrD;
 649
 650             /* load j atom coordinates */
 651             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 652                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 653                                                  &jx0,&jy0,&jz0);
 654
 655             /* Calculate displacement vector */
 656             dx00             = _mm256_sub_pd(ix0,jx0);
 657             dy00             = _mm256_sub_pd(iy0,jy0);
 658             dz00             = _mm256_sub_pd(iz0,jz0);
 659
 660             /* Calculate squared distance and things based on it */
 661             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 662
 663             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 664
 665             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 666
 667             /* Load parameters for j particles */
 668             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 669                                                                  charge+jnrC+0,charge+jnrD+0);
 670             vdwjidx0A        = 2*vdwtype[jnrA+0];
 671             vdwjidx0B        = 2*vdwtype[jnrB+0];
 672             vdwjidx0C        = 2*vdwtype[jnrC+0];
 673             vdwjidx0D        = 2*vdwtype[jnrD+0];
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 680             {
 681
 682             /* Compute parameters for interactions between i and j atoms */
 683             qq00             = _mm256_mul_pd(iq0,jq0);
 684             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 685                                             vdwioffsetptr0+vdwjidx0B,
 686                                             vdwioffsetptr0+vdwjidx0C,
 687                                             vdwioffsetptr0+vdwjidx0D,
 688                                             &c6_00,&c12_00);
 689
 690             /* REACTION-FIELD ELECTROSTATICS */
 691             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 692
 693             /* LENNARD-JONES DISPERSION/REPULSION */
 694
 695             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 696             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 697
 698             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 699
 700             fscal            = _mm256_add_pd(felec,fvdw);
 701
 702             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 703
 704             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 705
 706             /* Calculate temporary vectorial force */
 707             tx               = _mm256_mul_pd(fscal,dx00);
 708             ty               = _mm256_mul_pd(fscal,dy00);
 709             tz               = _mm256_mul_pd(fscal,dz00);
 710
 711             /* Update vectorial force */
 712             fix0             = _mm256_add_pd(fix0,tx);
 713             fiy0             = _mm256_add_pd(fiy0,ty);
 714             fiz0             = _mm256_add_pd(fiz0,tz);
 715
 716             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 717             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 718             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 719             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 720             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 721
 722             }
 723
 724             /* Inner loop uses 37 flops */
 725         }
 726
 727         /* End of innermost loop */
 728
 729         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 730                                                  f+i_coord_offset,fshift+i_shift_offset);
 731
 732         /* Increment number of inner iterations */
 733         inneriter                  += j_index_end - j_index_start;
 734
 735         /* Outer loop uses 7 flops */
 736     }
 737
 738     /* Increment number of outer iterations */
 739     outeriter        += nri;
 740
 741     /* Update outer/inner flops */
 742
 743     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*37);
 744 }