src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecNone_VdwLJSh_GeomP1P1_avx_256_single.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_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_single.h"
  48 #include "kernelutil_x86_avx_256_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_256_single
  52  * Electrostatics interaction: None
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_256_single
  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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  76     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  77     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  84     real             scratch[4*DIM];
  85     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  89     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     int              nvdwtype;
  92     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  96     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  97     __m256           dummy_mask,cutoff_mask;
  98     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  99     __m256           one     = _mm256_set1_ps(1.0);
 100     __m256           two     = _mm256_set1_ps(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     nvdwtype         = fr->ntype;
 113     vdwparam         = fr->nbfp;
 114     vdwtype          = mdatoms->typeA;
 115
 116     rcutoff_scalar   = fr->rvdw;
 117     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 118     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 119
 120     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 121     rvdw             = _mm256_set1_ps(fr->rvdw);
 122
 123     /* Avoid stupid compiler warnings */
 124     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 125     j_coord_offsetA = 0;
 126     j_coord_offsetB = 0;
 127     j_coord_offsetC = 0;
 128     j_coord_offsetD = 0;
 129     j_coord_offsetE = 0;
 130     j_coord_offsetF = 0;
 131     j_coord_offsetG = 0;
 132     j_coord_offsetH = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     for(iidx=0;iidx<4*DIM;iidx++)
 138     {
 139         scratch[iidx] = 0.0;
 140     }
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 158
 159         fix0             = _mm256_setzero_ps();
 160         fiy0             = _mm256_setzero_ps();
 161         fiz0             = _mm256_setzero_ps();
 162
 163         /* Load parameters for i particles */
 164         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 165
 166         /* Reset potential sums */
 167         vvdwsum          = _mm256_setzero_ps();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             jnrC             = jjnr[jidx+2];
 177             jnrD             = jjnr[jidx+3];
 178             jnrE             = jjnr[jidx+4];
 179             jnrF             = jjnr[jidx+5];
 180             jnrG             = jjnr[jidx+6];
 181             jnrH             = jjnr[jidx+7];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184             j_coord_offsetC  = DIM*jnrC;
 185             j_coord_offsetD  = DIM*jnrD;
 186             j_coord_offsetE  = DIM*jnrE;
 187             j_coord_offsetF  = DIM*jnrF;
 188             j_coord_offsetG  = DIM*jnrG;
 189             j_coord_offsetH  = DIM*jnrH;
 190
 191             /* load j atom coordinates */
 192             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 193                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 194                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 195                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 196                                                  &jx0,&jy0,&jz0);
 197
 198             /* Calculate displacement vector */
 199             dx00             = _mm256_sub_ps(ix0,jx0);
 200             dy00             = _mm256_sub_ps(iy0,jy0);
 201             dz00             = _mm256_sub_ps(iz0,jz0);
 202
 203             /* Calculate squared distance and things based on it */
 204             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 205
 206             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 207
 208             /* Load parameters for j particles */
 209             vdwjidx0A        = 2*vdwtype[jnrA+0];
 210             vdwjidx0B        = 2*vdwtype[jnrB+0];
 211             vdwjidx0C        = 2*vdwtype[jnrC+0];
 212             vdwjidx0D        = 2*vdwtype[jnrD+0];
 213             vdwjidx0E        = 2*vdwtype[jnrE+0];
 214             vdwjidx0F        = 2*vdwtype[jnrF+0];
 215             vdwjidx0G        = 2*vdwtype[jnrG+0];
 216             vdwjidx0H        = 2*vdwtype[jnrH+0];
 217
 218             /**************************
 219              * CALCULATE INTERACTIONS *
 220              **************************/
 221
 222             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 223             {
 224
 225             /* Compute parameters for interactions between i and j atoms */
 226             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 227                                             vdwioffsetptr0+vdwjidx0B,
 228                                             vdwioffsetptr0+vdwjidx0C,
 229                                             vdwioffsetptr0+vdwjidx0D,
 230                                             vdwioffsetptr0+vdwjidx0E,
 231                                             vdwioffsetptr0+vdwjidx0F,
 232                                             vdwioffsetptr0+vdwjidx0G,
 233                                             vdwioffsetptr0+vdwjidx0H,
 234                                             &c6_00,&c12_00);
 235
 236             /* LENNARD-JONES DISPERSION/REPULSION */
 237
 238             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 239             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 240             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 241             vvdw             = _mm256_sub_ps(_mm256_mul_ps( _mm256_sub_ps(vvdw12 , _mm256_mul_ps(c12_00,_mm256_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 242                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 243             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 244
 245             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 249             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 250
 251             fscal            = fvdw;
 252
 253             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 254
 255             /* Calculate temporary vectorial force */
 256             tx               = _mm256_mul_ps(fscal,dx00);
 257             ty               = _mm256_mul_ps(fscal,dy00);
 258             tz               = _mm256_mul_ps(fscal,dz00);
 259
 260             /* Update vectorial force */
 261             fix0             = _mm256_add_ps(fix0,tx);
 262             fiy0             = _mm256_add_ps(fiy0,ty);
 263             fiz0             = _mm256_add_ps(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             fjptrE             = f+j_coord_offsetE;
 270             fjptrF             = f+j_coord_offsetF;
 271             fjptrG             = f+j_coord_offsetG;
 272             fjptrH             = f+j_coord_offsetH;
 273             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 274
 275             }
 276
 277             /* Inner loop uses 41 flops */
 278         }
 279
 280         if(jidx<j_index_end)
 281         {
 282
 283             /* Get j neighbor index, and coordinate index */
 284             jnrlistA         = jjnr[jidx];
 285             jnrlistB         = jjnr[jidx+1];
 286             jnrlistC         = jjnr[jidx+2];
 287             jnrlistD         = jjnr[jidx+3];
 288             jnrlistE         = jjnr[jidx+4];
 289             jnrlistF         = jjnr[jidx+5];
 290             jnrlistG         = jjnr[jidx+6];
 291             jnrlistH         = jjnr[jidx+7];
 292             /* Sign of each element will be negative for non-real atoms.
 293              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 294              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 295              */
 296             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 297                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 298
 299             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 300             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 301             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 302             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 303             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 304             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 305             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 306             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 307             j_coord_offsetA  = DIM*jnrA;
 308             j_coord_offsetB  = DIM*jnrB;
 309             j_coord_offsetC  = DIM*jnrC;
 310             j_coord_offsetD  = DIM*jnrD;
 311             j_coord_offsetE  = DIM*jnrE;
 312             j_coord_offsetF  = DIM*jnrF;
 313             j_coord_offsetG  = DIM*jnrG;
 314             j_coord_offsetH  = DIM*jnrH;
 315
 316             /* load j atom coordinates */
 317             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 318                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 319                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 320                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 321                                                  &jx0,&jy0,&jz0);
 322
 323             /* Calculate displacement vector */
 324             dx00             = _mm256_sub_ps(ix0,jx0);
 325             dy00             = _mm256_sub_ps(iy0,jy0);
 326             dz00             = _mm256_sub_ps(iz0,jz0);
 327
 328             /* Calculate squared distance and things based on it */
 329             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 330
 331             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 332
 333             /* Load parameters for j particles */
 334             vdwjidx0A        = 2*vdwtype[jnrA+0];
 335             vdwjidx0B        = 2*vdwtype[jnrB+0];
 336             vdwjidx0C        = 2*vdwtype[jnrC+0];
 337             vdwjidx0D        = 2*vdwtype[jnrD+0];
 338             vdwjidx0E        = 2*vdwtype[jnrE+0];
 339             vdwjidx0F        = 2*vdwtype[jnrF+0];
 340             vdwjidx0G        = 2*vdwtype[jnrG+0];
 341             vdwjidx0H        = 2*vdwtype[jnrH+0];
 342
 343             /**************************
 344              * CALCULATE INTERACTIONS *
 345              **************************/
 346
 347             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 348             {
 349
 350             /* Compute parameters for interactions between i and j atoms */
 351             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 352                                             vdwioffsetptr0+vdwjidx0B,
 353                                             vdwioffsetptr0+vdwjidx0C,
 354                                             vdwioffsetptr0+vdwjidx0D,
 355                                             vdwioffsetptr0+vdwjidx0E,
 356                                             vdwioffsetptr0+vdwjidx0F,
 357                                             vdwioffsetptr0+vdwjidx0G,
 358                                             vdwioffsetptr0+vdwjidx0H,
 359                                             &c6_00,&c12_00);
 360
 361             /* LENNARD-JONES DISPERSION/REPULSION */
 362
 363             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 364             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 365             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 366             vvdw             = _mm256_sub_ps(_mm256_mul_ps( _mm256_sub_ps(vvdw12 , _mm256_mul_ps(c12_00,_mm256_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 367                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 368             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 369
 370             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 374             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 375             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 376
 377             fscal            = fvdw;
 378
 379             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 380
 381             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 382
 383             /* Calculate temporary vectorial force */
 384             tx               = _mm256_mul_ps(fscal,dx00);
 385             ty               = _mm256_mul_ps(fscal,dy00);
 386             tz               = _mm256_mul_ps(fscal,dz00);
 387
 388             /* Update vectorial force */
 389             fix0             = _mm256_add_ps(fix0,tx);
 390             fiy0             = _mm256_add_ps(fiy0,ty);
 391             fiz0             = _mm256_add_ps(fiz0,tz);
 392
 393             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 394             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 395             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 396             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 397             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 398             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 399             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 400             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 401             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 402
 403             }
 404
 405             /* Inner loop uses 41 flops */
 406         }
 407
 408         /* End of innermost loop */
 409
 410         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 411                                                  f+i_coord_offset,fshift+i_shift_offset);
 412
 413         ggid                        = gid[iidx];
 414         /* Update potential energies */
 415         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 416
 417         /* Increment number of inner iterations */
 418         inneriter                  += j_index_end - j_index_start;
 419
 420         /* Outer loop uses 7 flops */
 421     }
 422
 423     /* Increment number of outer iterations */
 424     outeriter        += nri;
 425
 426     /* Update outer/inner flops */
 427
 428     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*41);
 429 }
 430 /*
 431  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_256_single
 432  * Electrostatics interaction: None
 433  * VdW interaction:            LennardJones
 434  * Geometry:                   Particle-Particle
 435  * Calculate force/pot:        Force
 436  */
 437 void
 438 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_256_single
 439                     (t_nblist                    * gmx_restrict       nlist,
 440                      rvec                        * gmx_restrict          xx,
 441                      rvec                        * gmx_restrict          ff,
 442                      t_forcerec                  * gmx_restrict          fr,
 443                      t_mdatoms                   * gmx_restrict     mdatoms,
 444                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 445                      t_nrnb                      * gmx_restrict        nrnb)
 446 {
 447     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 448      * just 0 for non-waters.
 449      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 450      * jnr indices corresponding to data put in the four positions in the SIMD register.
 451      */
 452     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 453     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 454     int              jnrA,jnrB,jnrC,jnrD;
 455     int              jnrE,jnrF,jnrG,jnrH;
 456     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 457     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 458     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 459     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 460     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 461     real             rcutoff_scalar;
 462     real             *shiftvec,*fshift,*x,*f;
 463     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 464     real             scratch[4*DIM];
 465     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 466     real *           vdwioffsetptr0;
 467     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 468     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 469     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 470     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 471     int              nvdwtype;
 472     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 473     int              *vdwtype;
 474     real             *vdwparam;
 475     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 476     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 477     __m256           dummy_mask,cutoff_mask;
 478     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 479     __m256           one     = _mm256_set1_ps(1.0);
 480     __m256           two     = _mm256_set1_ps(2.0);
 481     x                = xx[0];
 482     f                = ff[0];
 483
 484     nri              = nlist->nri;
 485     iinr             = nlist->iinr;
 486     jindex           = nlist->jindex;
 487     jjnr             = nlist->jjnr;
 488     shiftidx         = nlist->shift;
 489     gid              = nlist->gid;
 490     shiftvec         = fr->shift_vec[0];
 491     fshift           = fr->fshift[0];
 492     nvdwtype         = fr->ntype;
 493     vdwparam         = fr->nbfp;
 494     vdwtype          = mdatoms->typeA;
 495
 496     rcutoff_scalar   = fr->rvdw;
 497     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 498     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 499
 500     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 501     rvdw             = _mm256_set1_ps(fr->rvdw);
 502
 503     /* Avoid stupid compiler warnings */
 504     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 505     j_coord_offsetA = 0;
 506     j_coord_offsetB = 0;
 507     j_coord_offsetC = 0;
 508     j_coord_offsetD = 0;
 509     j_coord_offsetE = 0;
 510     j_coord_offsetF = 0;
 511     j_coord_offsetG = 0;
 512     j_coord_offsetH = 0;
 513
 514     outeriter        = 0;
 515     inneriter        = 0;
 516
 517     for(iidx=0;iidx<4*DIM;iidx++)
 518     {
 519         scratch[iidx] = 0.0;
 520     }
 521
 522     /* Start outer loop over neighborlists */
 523     for(iidx=0; iidx<nri; iidx++)
 524     {
 525         /* Load shift vector for this list */
 526         i_shift_offset   = DIM*shiftidx[iidx];
 527
 528         /* Load limits for loop over neighbors */
 529         j_index_start    = jindex[iidx];
 530         j_index_end      = jindex[iidx+1];
 531
 532         /* Get outer coordinate index */
 533         inr              = iinr[iidx];
 534         i_coord_offset   = DIM*inr;
 535
 536         /* Load i particle coords and add shift vector */
 537         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 538
 539         fix0             = _mm256_setzero_ps();
 540         fiy0             = _mm256_setzero_ps();
 541         fiz0             = _mm256_setzero_ps();
 542
 543         /* Load parameters for i particles */
 544         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 545
 546         /* Start inner kernel loop */
 547         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 548         {
 549
 550             /* Get j neighbor index, and coordinate index */
 551             jnrA             = jjnr[jidx];
 552             jnrB             = jjnr[jidx+1];
 553             jnrC             = jjnr[jidx+2];
 554             jnrD             = jjnr[jidx+3];
 555             jnrE             = jjnr[jidx+4];
 556             jnrF             = jjnr[jidx+5];
 557             jnrG             = jjnr[jidx+6];
 558             jnrH             = jjnr[jidx+7];
 559             j_coord_offsetA  = DIM*jnrA;
 560             j_coord_offsetB  = DIM*jnrB;
 561             j_coord_offsetC  = DIM*jnrC;
 562             j_coord_offsetD  = DIM*jnrD;
 563             j_coord_offsetE  = DIM*jnrE;
 564             j_coord_offsetF  = DIM*jnrF;
 565             j_coord_offsetG  = DIM*jnrG;
 566             j_coord_offsetH  = DIM*jnrH;
 567
 568             /* load j atom coordinates */
 569             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 570                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 571                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 572                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 573                                                  &jx0,&jy0,&jz0);
 574
 575             /* Calculate displacement vector */
 576             dx00             = _mm256_sub_ps(ix0,jx0);
 577             dy00             = _mm256_sub_ps(iy0,jy0);
 578             dz00             = _mm256_sub_ps(iz0,jz0);
 579
 580             /* Calculate squared distance and things based on it */
 581             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 582
 583             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 584
 585             /* Load parameters for j particles */
 586             vdwjidx0A        = 2*vdwtype[jnrA+0];
 587             vdwjidx0B        = 2*vdwtype[jnrB+0];
 588             vdwjidx0C        = 2*vdwtype[jnrC+0];
 589             vdwjidx0D        = 2*vdwtype[jnrD+0];
 590             vdwjidx0E        = 2*vdwtype[jnrE+0];
 591             vdwjidx0F        = 2*vdwtype[jnrF+0];
 592             vdwjidx0G        = 2*vdwtype[jnrG+0];
 593             vdwjidx0H        = 2*vdwtype[jnrH+0];
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 600             {
 601
 602             /* Compute parameters for interactions between i and j atoms */
 603             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 604                                             vdwioffsetptr0+vdwjidx0B,
 605                                             vdwioffsetptr0+vdwjidx0C,
 606                                             vdwioffsetptr0+vdwjidx0D,
 607                                             vdwioffsetptr0+vdwjidx0E,
 608                                             vdwioffsetptr0+vdwjidx0F,
 609                                             vdwioffsetptr0+vdwjidx0G,
 610                                             vdwioffsetptr0+vdwjidx0H,
 611                                             &c6_00,&c12_00);
 612
 613             /* LENNARD-JONES DISPERSION/REPULSION */
 614
 615             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 616             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 617
 618             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 619
 620             fscal            = fvdw;
 621
 622             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 623
 624             /* Calculate temporary vectorial force */
 625             tx               = _mm256_mul_ps(fscal,dx00);
 626             ty               = _mm256_mul_ps(fscal,dy00);
 627             tz               = _mm256_mul_ps(fscal,dz00);
 628
 629             /* Update vectorial force */
 630             fix0             = _mm256_add_ps(fix0,tx);
 631             fiy0             = _mm256_add_ps(fiy0,ty);
 632             fiz0             = _mm256_add_ps(fiz0,tz);
 633
 634             fjptrA             = f+j_coord_offsetA;
 635             fjptrB             = f+j_coord_offsetB;
 636             fjptrC             = f+j_coord_offsetC;
 637             fjptrD             = f+j_coord_offsetD;
 638             fjptrE             = f+j_coord_offsetE;
 639             fjptrF             = f+j_coord_offsetF;
 640             fjptrG             = f+j_coord_offsetG;
 641             fjptrH             = f+j_coord_offsetH;
 642             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 643
 644             }
 645
 646             /* Inner loop uses 30 flops */
 647         }
 648
 649         if(jidx<j_index_end)
 650         {
 651
 652             /* Get j neighbor index, and coordinate index */
 653             jnrlistA         = jjnr[jidx];
 654             jnrlistB         = jjnr[jidx+1];
 655             jnrlistC         = jjnr[jidx+2];
 656             jnrlistD         = jjnr[jidx+3];
 657             jnrlistE         = jjnr[jidx+4];
 658             jnrlistF         = jjnr[jidx+5];
 659             jnrlistG         = jjnr[jidx+6];
 660             jnrlistH         = jjnr[jidx+7];
 661             /* Sign of each element will be negative for non-real atoms.
 662              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 663              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 664              */
 665             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 666                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 667
 668             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 669             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 670             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 671             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 672             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 673             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 674             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 675             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 676             j_coord_offsetA  = DIM*jnrA;
 677             j_coord_offsetB  = DIM*jnrB;
 678             j_coord_offsetC  = DIM*jnrC;
 679             j_coord_offsetD  = DIM*jnrD;
 680             j_coord_offsetE  = DIM*jnrE;
 681             j_coord_offsetF  = DIM*jnrF;
 682             j_coord_offsetG  = DIM*jnrG;
 683             j_coord_offsetH  = DIM*jnrH;
 684
 685             /* load j atom coordinates */
 686             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 687                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 688                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 689                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 690                                                  &jx0,&jy0,&jz0);
 691
 692             /* Calculate displacement vector */
 693             dx00             = _mm256_sub_ps(ix0,jx0);
 694             dy00             = _mm256_sub_ps(iy0,jy0);
 695             dz00             = _mm256_sub_ps(iz0,jz0);
 696
 697             /* Calculate squared distance and things based on it */
 698             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 699
 700             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 701
 702             /* Load parameters for j particles */
 703             vdwjidx0A        = 2*vdwtype[jnrA+0];
 704             vdwjidx0B        = 2*vdwtype[jnrB+0];
 705             vdwjidx0C        = 2*vdwtype[jnrC+0];
 706             vdwjidx0D        = 2*vdwtype[jnrD+0];
 707             vdwjidx0E        = 2*vdwtype[jnrE+0];
 708             vdwjidx0F        = 2*vdwtype[jnrF+0];
 709             vdwjidx0G        = 2*vdwtype[jnrG+0];
 710             vdwjidx0H        = 2*vdwtype[jnrH+0];
 711
 712             /**************************
 713              * CALCULATE INTERACTIONS *
 714              **************************/
 715
 716             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 717             {
 718
 719             /* Compute parameters for interactions between i and j atoms */
 720             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 721                                             vdwioffsetptr0+vdwjidx0B,
 722                                             vdwioffsetptr0+vdwjidx0C,
 723                                             vdwioffsetptr0+vdwjidx0D,
 724                                             vdwioffsetptr0+vdwjidx0E,
 725                                             vdwioffsetptr0+vdwjidx0F,
 726                                             vdwioffsetptr0+vdwjidx0G,
 727                                             vdwioffsetptr0+vdwjidx0H,
 728                                             &c6_00,&c12_00);
 729
 730             /* LENNARD-JONES DISPERSION/REPULSION */
 731
 732             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 733             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 734
 735             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 736
 737             fscal            = fvdw;
 738
 739             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 740
 741             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 742
 743             /* Calculate temporary vectorial force */
 744             tx               = _mm256_mul_ps(fscal,dx00);
 745             ty               = _mm256_mul_ps(fscal,dy00);
 746             tz               = _mm256_mul_ps(fscal,dz00);
 747
 748             /* Update vectorial force */
 749             fix0             = _mm256_add_ps(fix0,tx);
 750             fiy0             = _mm256_add_ps(fiy0,ty);
 751             fiz0             = _mm256_add_ps(fiz0,tz);
 752
 753             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 754             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 755             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 756             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 757             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 758             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 759             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 760             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 761             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 762
 763             }
 764
 765             /* Inner loop uses 30 flops */
 766         }
 767
 768         /* End of innermost loop */
 769
 770         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 771                                                  f+i_coord_offset,fshift+i_shift_offset);
 772
 773         /* Increment number of inner iterations */
 774         inneriter                  += j_index_end - j_index_start;
 775
 776         /* Outer loop uses 6 flops */
 777     }
 778
 779     /* Increment number of outer iterations */
 780     outeriter        += nri;
 781
 782     /* Update outer/inner flops */
 783
 784     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*30);
 785 }