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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  98     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  99     __m256d          dummy_mask,cutoff_mask;
 100     __m128           tmpmask0,tmpmask1;
 101     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 102     __m256d          one     = _mm256_set1_pd(1.0);
 103     __m256d          two     = _mm256_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm256_set1_pd(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     krf              = _mm256_set1_pd(fr->ic->k_rf);
 118     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 119     crf              = _mm256_set1_pd(fr->ic->c_rf);
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 125     rcutoff_scalar   = fr->rcoulomb;
 126     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 127     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 128
 129     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 130     rvdw             = _mm256_set1_pd(fr->rvdw);
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = jnrC = jnrD = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136     j_coord_offsetC = 0;
 137     j_coord_offsetD = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 163
 164         fix0             = _mm256_setzero_pd();
 165         fiy0             = _mm256_setzero_pd();
 166         fiz0             = _mm256_setzero_pd();
 167
 168         /* Load parameters for i particles */
 169         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 170         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm256_setzero_pd();
 174         vvdwsum          = _mm256_setzero_pd();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             jnrC             = jjnr[jidx+2];
 184             jnrD             = jjnr[jidx+3];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187             j_coord_offsetC  = DIM*jnrC;
 188             j_coord_offsetD  = DIM*jnrD;
 189
 190             /* load j atom coordinates */
 191             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 192                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 193                                                  &jx0,&jy0,&jz0);
 194
 195             /* Calculate displacement vector */
 196             dx00             = _mm256_sub_pd(ix0,jx0);
 197             dy00             = _mm256_sub_pd(iy0,jy0);
 198             dz00             = _mm256_sub_pd(iz0,jz0);
 199
 200             /* Calculate squared distance and things based on it */
 201             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 202
 203             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 204
 205             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 206
 207             /* Load parameters for j particles */
 208             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 209                                                                  charge+jnrC+0,charge+jnrD+0);
 210             vdwjidx0A        = 2*vdwtype[jnrA+0];
 211             vdwjidx0B        = 2*vdwtype[jnrB+0];
 212             vdwjidx0C        = 2*vdwtype[jnrC+0];
 213             vdwjidx0D        = 2*vdwtype[jnrD+0];
 214
 215             /**************************
 216              * CALCULATE INTERACTIONS *
 217              **************************/
 218
 219             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 220             {
 221
 222             /* Compute parameters for interactions between i and j atoms */
 223             qq00             = _mm256_mul_pd(iq0,jq0);
 224             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 225                                             vdwioffsetptr0+vdwjidx0B,
 226                                             vdwioffsetptr0+vdwjidx0C,
 227                                             vdwioffsetptr0+vdwjidx0D,
 228                                             &c6_00,&c12_00);
 229
 230             /* REACTION-FIELD ELECTROSTATICS */
 231             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 232             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 233
 234             /* LENNARD-JONES DISPERSION/REPULSION */
 235
 236             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 237             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 238             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 239             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) ,
 240                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 241             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 242
 243             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 244
 245             /* Update potential sum for this i atom from the interaction with this j atom. */
 246             velec            = _mm256_and_pd(velec,cutoff_mask);
 247             velecsum         = _mm256_add_pd(velecsum,velec);
 248             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 249             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 250
 251             fscal            = _mm256_add_pd(felec,fvdw);
 252
 253             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 254
 255             /* Calculate temporary vectorial force */
 256             tx               = _mm256_mul_pd(fscal,dx00);
 257             ty               = _mm256_mul_pd(fscal,dy00);
 258             tz               = _mm256_mul_pd(fscal,dz00);
 259
 260             /* Update vectorial force */
 261             fix0             = _mm256_add_pd(fix0,tx);
 262             fiy0             = _mm256_add_pd(fiy0,ty);
 263             fiz0             = _mm256_add_pd(fiz0,tz);
 264
 265             fjptrA             = f+j_coord_offsetA;
 266             fjptrB             = f+j_coord_offsetB;
 267             fjptrC             = f+j_coord_offsetC;
 268             fjptrD             = f+j_coord_offsetD;
 269             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 270
 271             }
 272
 273             /* Inner loop uses 54 flops */
 274         }
 275
 276         if(jidx<j_index_end)
 277         {
 278
 279             /* Get j neighbor index, and coordinate index */
 280             jnrlistA         = jjnr[jidx];
 281             jnrlistB         = jjnr[jidx+1];
 282             jnrlistC         = jjnr[jidx+2];
 283             jnrlistD         = jjnr[jidx+3];
 284             /* Sign of each element will be negative for non-real atoms.
 285              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 286              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 287              */
 288             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 289
 290             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 291             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 292             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 293
 294             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 295             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 296             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 297             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 298             j_coord_offsetA  = DIM*jnrA;
 299             j_coord_offsetB  = DIM*jnrB;
 300             j_coord_offsetC  = DIM*jnrC;
 301             j_coord_offsetD  = DIM*jnrD;
 302
 303             /* load j atom coordinates */
 304             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 305                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 306                                                  &jx0,&jy0,&jz0);
 307
 308             /* Calculate displacement vector */
 309             dx00             = _mm256_sub_pd(ix0,jx0);
 310             dy00             = _mm256_sub_pd(iy0,jy0);
 311             dz00             = _mm256_sub_pd(iz0,jz0);
 312
 313             /* Calculate squared distance and things based on it */
 314             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 315
 316             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 317
 318             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 319
 320             /* Load parameters for j particles */
 321             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 322                                                                  charge+jnrC+0,charge+jnrD+0);
 323             vdwjidx0A        = 2*vdwtype[jnrA+0];
 324             vdwjidx0B        = 2*vdwtype[jnrB+0];
 325             vdwjidx0C        = 2*vdwtype[jnrC+0];
 326             vdwjidx0D        = 2*vdwtype[jnrD+0];
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 333             {
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq00             = _mm256_mul_pd(iq0,jq0);
 337             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 338                                             vdwioffsetptr0+vdwjidx0B,
 339                                             vdwioffsetptr0+vdwjidx0C,
 340                                             vdwioffsetptr0+vdwjidx0D,
 341                                             &c6_00,&c12_00);
 342
 343             /* REACTION-FIELD ELECTROSTATICS */
 344             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 345             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 346
 347             /* LENNARD-JONES DISPERSION/REPULSION */
 348
 349             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 350             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 351             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 352             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) ,
 353                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 354             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 355
 356             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velec            = _mm256_and_pd(velec,cutoff_mask);
 360             velec            = _mm256_andnot_pd(dummy_mask,velec);
 361             velecsum         = _mm256_add_pd(velecsum,velec);
 362             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 363             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 364             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 365
 366             fscal            = _mm256_add_pd(felec,fvdw);
 367
 368             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 369
 370             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 371
 372             /* Calculate temporary vectorial force */
 373             tx               = _mm256_mul_pd(fscal,dx00);
 374             ty               = _mm256_mul_pd(fscal,dy00);
 375             tz               = _mm256_mul_pd(fscal,dz00);
 376
 377             /* Update vectorial force */
 378             fix0             = _mm256_add_pd(fix0,tx);
 379             fiy0             = _mm256_add_pd(fiy0,ty);
 380             fiz0             = _mm256_add_pd(fiz0,tz);
 381
 382             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 383             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 384             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 385             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 386             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 387
 388             }
 389
 390             /* Inner loop uses 54 flops */
 391         }
 392
 393         /* End of innermost loop */
 394
 395         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 396                                                  f+i_coord_offset,fshift+i_shift_offset);
 397
 398         ggid                        = gid[iidx];
 399         /* Update potential energies */
 400         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 401         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 402
 403         /* Increment number of inner iterations */
 404         inneriter                  += j_index_end - j_index_start;
 405
 406         /* Outer loop uses 9 flops */
 407     }
 408
 409     /* Increment number of outer iterations */
 410     outeriter        += nri;
 411
 412     /* Update outer/inner flops */
 413
 414     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 415 }
 416 /*
 417  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_256_double
 418  * Electrostatics interaction: ReactionField
 419  * VdW interaction:            LennardJones
 420  * Geometry:                   Particle-Particle
 421  * Calculate force/pot:        Force
 422  */
 423 void
 424 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_256_double
 425                     (t_nblist                    * gmx_restrict       nlist,
 426                      rvec                        * gmx_restrict          xx,
 427                      rvec                        * gmx_restrict          ff,
 428                      t_forcerec                  * gmx_restrict          fr,
 429                      t_mdatoms                   * gmx_restrict     mdatoms,
 430                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 431                      t_nrnb                      * gmx_restrict        nrnb)
 432 {
 433     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 434      * just 0 for non-waters.
 435      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 436      * jnr indices corresponding to data put in the four positions in the SIMD register.
 437      */
 438     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 439     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 440     int              jnrA,jnrB,jnrC,jnrD;
 441     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 442     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 443     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 444     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 445     real             rcutoff_scalar;
 446     real             *shiftvec,*fshift,*x,*f;
 447     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 448     real             scratch[4*DIM];
 449     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 450     real *           vdwioffsetptr0;
 451     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 452     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 453     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 454     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 455     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 456     real             *charge;
 457     int              nvdwtype;
 458     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 459     int              *vdwtype;
 460     real             *vdwparam;
 461     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 462     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 463     __m256d          dummy_mask,cutoff_mask;
 464     __m128           tmpmask0,tmpmask1;
 465     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 466     __m256d          one     = _mm256_set1_pd(1.0);
 467     __m256d          two     = _mm256_set1_pd(2.0);
 468     x                = xx[0];
 469     f                = ff[0];
 470
 471     nri              = nlist->nri;
 472     iinr             = nlist->iinr;
 473     jindex           = nlist->jindex;
 474     jjnr             = nlist->jjnr;
 475     shiftidx         = nlist->shift;
 476     gid              = nlist->gid;
 477     shiftvec         = fr->shift_vec[0];
 478     fshift           = fr->fshift[0];
 479     facel            = _mm256_set1_pd(fr->epsfac);
 480     charge           = mdatoms->chargeA;
 481     krf              = _mm256_set1_pd(fr->ic->k_rf);
 482     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 483     crf              = _mm256_set1_pd(fr->ic->c_rf);
 484     nvdwtype         = fr->ntype;
 485     vdwparam         = fr->nbfp;
 486     vdwtype          = mdatoms->typeA;
 487
 488     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 489     rcutoff_scalar   = fr->rcoulomb;
 490     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 491     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 492
 493     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 494     rvdw             = _mm256_set1_pd(fr->rvdw);
 495
 496     /* Avoid stupid compiler warnings */
 497     jnrA = jnrB = jnrC = jnrD = 0;
 498     j_coord_offsetA = 0;
 499     j_coord_offsetB = 0;
 500     j_coord_offsetC = 0;
 501     j_coord_offsetD = 0;
 502
 503     outeriter        = 0;
 504     inneriter        = 0;
 505
 506     for(iidx=0;iidx<4*DIM;iidx++)
 507     {
 508         scratch[iidx] = 0.0;
 509     }
 510
 511     /* Start outer loop over neighborlists */
 512     for(iidx=0; iidx<nri; iidx++)
 513     {
 514         /* Load shift vector for this list */
 515         i_shift_offset   = DIM*shiftidx[iidx];
 516
 517         /* Load limits for loop over neighbors */
 518         j_index_start    = jindex[iidx];
 519         j_index_end      = jindex[iidx+1];
 520
 521         /* Get outer coordinate index */
 522         inr              = iinr[iidx];
 523         i_coord_offset   = DIM*inr;
 524
 525         /* Load i particle coords and add shift vector */
 526         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 527
 528         fix0             = _mm256_setzero_pd();
 529         fiy0             = _mm256_setzero_pd();
 530         fiz0             = _mm256_setzero_pd();
 531
 532         /* Load parameters for i particles */
 533         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 534         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 535
 536         /* Start inner kernel loop */
 537         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 538         {
 539
 540             /* Get j neighbor index, and coordinate index */
 541             jnrA             = jjnr[jidx];
 542             jnrB             = jjnr[jidx+1];
 543             jnrC             = jjnr[jidx+2];
 544             jnrD             = jjnr[jidx+3];
 545             j_coord_offsetA  = DIM*jnrA;
 546             j_coord_offsetB  = DIM*jnrB;
 547             j_coord_offsetC  = DIM*jnrC;
 548             j_coord_offsetD  = DIM*jnrD;
 549
 550             /* load j atom coordinates */
 551             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 552                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 553                                                  &jx0,&jy0,&jz0);
 554
 555             /* Calculate displacement vector */
 556             dx00             = _mm256_sub_pd(ix0,jx0);
 557             dy00             = _mm256_sub_pd(iy0,jy0);
 558             dz00             = _mm256_sub_pd(iz0,jz0);
 559
 560             /* Calculate squared distance and things based on it */
 561             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 562
 563             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 564
 565             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 566
 567             /* Load parameters for j particles */
 568             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 569                                                                  charge+jnrC+0,charge+jnrD+0);
 570             vdwjidx0A        = 2*vdwtype[jnrA+0];
 571             vdwjidx0B        = 2*vdwtype[jnrB+0];
 572             vdwjidx0C        = 2*vdwtype[jnrC+0];
 573             vdwjidx0D        = 2*vdwtype[jnrD+0];
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 580             {
 581
 582             /* Compute parameters for interactions between i and j atoms */
 583             qq00             = _mm256_mul_pd(iq0,jq0);
 584             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 585                                             vdwioffsetptr0+vdwjidx0B,
 586                                             vdwioffsetptr0+vdwjidx0C,
 587                                             vdwioffsetptr0+vdwjidx0D,
 588                                             &c6_00,&c12_00);
 589
 590             /* REACTION-FIELD ELECTROSTATICS */
 591             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 592
 593             /* LENNARD-JONES DISPERSION/REPULSION */
 594
 595             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 596             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 597
 598             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 599
 600             fscal            = _mm256_add_pd(felec,fvdw);
 601
 602             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 603
 604             /* Calculate temporary vectorial force */
 605             tx               = _mm256_mul_pd(fscal,dx00);
 606             ty               = _mm256_mul_pd(fscal,dy00);
 607             tz               = _mm256_mul_pd(fscal,dz00);
 608
 609             /* Update vectorial force */
 610             fix0             = _mm256_add_pd(fix0,tx);
 611             fiy0             = _mm256_add_pd(fiy0,ty);
 612             fiz0             = _mm256_add_pd(fiz0,tz);
 613
 614             fjptrA             = f+j_coord_offsetA;
 615             fjptrB             = f+j_coord_offsetB;
 616             fjptrC             = f+j_coord_offsetC;
 617             fjptrD             = f+j_coord_offsetD;
 618             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 619
 620             }
 621
 622             /* Inner loop uses 37 flops */
 623         }
 624
 625         if(jidx<j_index_end)
 626         {
 627
 628             /* Get j neighbor index, and coordinate index */
 629             jnrlistA         = jjnr[jidx];
 630             jnrlistB         = jjnr[jidx+1];
 631             jnrlistC         = jjnr[jidx+2];
 632             jnrlistD         = jjnr[jidx+3];
 633             /* Sign of each element will be negative for non-real atoms.
 634              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 635              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 636              */
 637             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 638
 639             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 640             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 641             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 642
 643             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 644             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 645             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 646             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 647             j_coord_offsetA  = DIM*jnrA;
 648             j_coord_offsetB  = DIM*jnrB;
 649             j_coord_offsetC  = DIM*jnrC;
 650             j_coord_offsetD  = DIM*jnrD;
 651
 652             /* load j atom coordinates */
 653             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 654                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 655                                                  &jx0,&jy0,&jz0);
 656
 657             /* Calculate displacement vector */
 658             dx00             = _mm256_sub_pd(ix0,jx0);
 659             dy00             = _mm256_sub_pd(iy0,jy0);
 660             dz00             = _mm256_sub_pd(iz0,jz0);
 661
 662             /* Calculate squared distance and things based on it */
 663             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 664
 665             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 666
 667             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 668
 669             /* Load parameters for j particles */
 670             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 671                                                                  charge+jnrC+0,charge+jnrD+0);
 672             vdwjidx0A        = 2*vdwtype[jnrA+0];
 673             vdwjidx0B        = 2*vdwtype[jnrB+0];
 674             vdwjidx0C        = 2*vdwtype[jnrC+0];
 675             vdwjidx0D        = 2*vdwtype[jnrD+0];
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 682             {
 683
 684             /* Compute parameters for interactions between i and j atoms */
 685             qq00             = _mm256_mul_pd(iq0,jq0);
 686             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 687                                             vdwioffsetptr0+vdwjidx0B,
 688                                             vdwioffsetptr0+vdwjidx0C,
 689                                             vdwioffsetptr0+vdwjidx0D,
 690                                             &c6_00,&c12_00);
 691
 692             /* REACTION-FIELD ELECTROSTATICS */
 693             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 694
 695             /* LENNARD-JONES DISPERSION/REPULSION */
 696
 697             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 698             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 699
 700             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 701
 702             fscal            = _mm256_add_pd(felec,fvdw);
 703
 704             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 705
 706             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 707
 708             /* Calculate temporary vectorial force */
 709             tx               = _mm256_mul_pd(fscal,dx00);
 710             ty               = _mm256_mul_pd(fscal,dy00);
 711             tz               = _mm256_mul_pd(fscal,dz00);
 712
 713             /* Update vectorial force */
 714             fix0             = _mm256_add_pd(fix0,tx);
 715             fiy0             = _mm256_add_pd(fiy0,ty);
 716             fiz0             = _mm256_add_pd(fiz0,tz);
 717
 718             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 719             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 720             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 721             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 722             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 723
 724             }
 725
 726             /* Inner loop uses 37 flops */
 727         }
 728
 729         /* End of innermost loop */
 730
 731         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 732                                                  f+i_coord_offset,fshift+i_shift_offset);
 733
 734         /* Increment number of inner iterations */
 735         inneriter                  += j_index_end - j_index_start;
 736
 737         /* Outer loop uses 7 flops */
 738     }
 739
 740     /* Increment number of outer iterations */
 741     outeriter        += nri;
 742
 743     /* Update outer/inner flops */
 744
 745     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*37);
 746 }