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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_256_single.h"
  50 #include "kernelutil_x86_avx_256_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_256_single
  54  * Electrostatics interaction: None
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_256_single
  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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  78     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  79     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  80     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  81     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  82     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  83     real             rcutoff_scalar;
  84     real             *shiftvec,*fshift,*x,*f;
  85     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  86     real             scratch[4*DIM];
  87     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  88     real *           vdwioffsetptr0;
  89     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  91     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     int              nvdwtype;
  94     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  98     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  99     __m256           dummy_mask,cutoff_mask;
 100     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 101     __m256           one     = _mm256_set1_ps(1.0);
 102     __m256           two     = _mm256_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117
 118     rcutoff_scalar   = fr->rvdw;
 119     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 120     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 121
 122     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 123     rvdw             = _mm256_set1_ps(fr->rvdw);
 124
 125     /* Avoid stupid compiler warnings */
 126     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 127     j_coord_offsetA = 0;
 128     j_coord_offsetB = 0;
 129     j_coord_offsetC = 0;
 130     j_coord_offsetD = 0;
 131     j_coord_offsetE = 0;
 132     j_coord_offsetF = 0;
 133     j_coord_offsetG = 0;
 134     j_coord_offsetH = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 160
 161         fix0             = _mm256_setzero_ps();
 162         fiy0             = _mm256_setzero_ps();
 163         fiz0             = _mm256_setzero_ps();
 164
 165         /* Load parameters for i particles */
 166         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 167
 168         /* Reset potential sums */
 169         vvdwsum          = _mm256_setzero_ps();
 170
 171         /* Start inner kernel loop */
 172         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 173         {
 174
 175             /* Get j neighbor index, and coordinate index */
 176             jnrA             = jjnr[jidx];
 177             jnrB             = jjnr[jidx+1];
 178             jnrC             = jjnr[jidx+2];
 179             jnrD             = jjnr[jidx+3];
 180             jnrE             = jjnr[jidx+4];
 181             jnrF             = jjnr[jidx+5];
 182             jnrG             = jjnr[jidx+6];
 183             jnrH             = jjnr[jidx+7];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188             j_coord_offsetE  = DIM*jnrE;
 189             j_coord_offsetF  = DIM*jnrF;
 190             j_coord_offsetG  = DIM*jnrG;
 191             j_coord_offsetH  = DIM*jnrH;
 192
 193             /* load j atom coordinates */
 194             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 195                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 196                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 197                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 198                                                  &jx0,&jy0,&jz0);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm256_sub_ps(ix0,jx0);
 202             dy00             = _mm256_sub_ps(iy0,jy0);
 203             dz00             = _mm256_sub_ps(iz0,jz0);
 204
 205             /* Calculate squared distance and things based on it */
 206             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 207
 208             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 209
 210             /* Load parameters for j particles */
 211             vdwjidx0A        = 2*vdwtype[jnrA+0];
 212             vdwjidx0B        = 2*vdwtype[jnrB+0];
 213             vdwjidx0C        = 2*vdwtype[jnrC+0];
 214             vdwjidx0D        = 2*vdwtype[jnrD+0];
 215             vdwjidx0E        = 2*vdwtype[jnrE+0];
 216             vdwjidx0F        = 2*vdwtype[jnrF+0];
 217             vdwjidx0G        = 2*vdwtype[jnrG+0];
 218             vdwjidx0H        = 2*vdwtype[jnrH+0];
 219
 220             /**************************
 221              * CALCULATE INTERACTIONS *
 222              **************************/
 223
 224             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 225             {
 226
 227             /* Compute parameters for interactions between i and j atoms */
 228             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 229                                             vdwioffsetptr0+vdwjidx0B,
 230                                             vdwioffsetptr0+vdwjidx0C,
 231                                             vdwioffsetptr0+vdwjidx0D,
 232                                             vdwioffsetptr0+vdwjidx0E,
 233                                             vdwioffsetptr0+vdwjidx0F,
 234                                             vdwioffsetptr0+vdwjidx0G,
 235                                             vdwioffsetptr0+vdwjidx0H,
 236                                             &c6_00,&c12_00);
 237
 238             /* LENNARD-JONES DISPERSION/REPULSION */
 239
 240             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 241             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 242             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 243             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) ,
 244                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 245             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 246
 247             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 251             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 252
 253             fscal            = fvdw;
 254
 255             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 256
 257             /* Calculate temporary vectorial force */
 258             tx               = _mm256_mul_ps(fscal,dx00);
 259             ty               = _mm256_mul_ps(fscal,dy00);
 260             tz               = _mm256_mul_ps(fscal,dz00);
 261
 262             /* Update vectorial force */
 263             fix0             = _mm256_add_ps(fix0,tx);
 264             fiy0             = _mm256_add_ps(fiy0,ty);
 265             fiz0             = _mm256_add_ps(fiz0,tz);
 266
 267             fjptrA             = f+j_coord_offsetA;
 268             fjptrB             = f+j_coord_offsetB;
 269             fjptrC             = f+j_coord_offsetC;
 270             fjptrD             = f+j_coord_offsetD;
 271             fjptrE             = f+j_coord_offsetE;
 272             fjptrF             = f+j_coord_offsetF;
 273             fjptrG             = f+j_coord_offsetG;
 274             fjptrH             = f+j_coord_offsetH;
 275             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 276
 277             }
 278
 279             /* Inner loop uses 41 flops */
 280         }
 281
 282         if(jidx<j_index_end)
 283         {
 284
 285             /* Get j neighbor index, and coordinate index */
 286             jnrlistA         = jjnr[jidx];
 287             jnrlistB         = jjnr[jidx+1];
 288             jnrlistC         = jjnr[jidx+2];
 289             jnrlistD         = jjnr[jidx+3];
 290             jnrlistE         = jjnr[jidx+4];
 291             jnrlistF         = jjnr[jidx+5];
 292             jnrlistG         = jjnr[jidx+6];
 293             jnrlistH         = jjnr[jidx+7];
 294             /* Sign of each element will be negative for non-real atoms.
 295              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 296              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 297              */
 298             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 299                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 300
 301             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 302             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 303             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 304             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 305             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 306             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 307             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 308             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 309             j_coord_offsetA  = DIM*jnrA;
 310             j_coord_offsetB  = DIM*jnrB;
 311             j_coord_offsetC  = DIM*jnrC;
 312             j_coord_offsetD  = DIM*jnrD;
 313             j_coord_offsetE  = DIM*jnrE;
 314             j_coord_offsetF  = DIM*jnrF;
 315             j_coord_offsetG  = DIM*jnrG;
 316             j_coord_offsetH  = DIM*jnrH;
 317
 318             /* load j atom coordinates */
 319             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 320                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 321                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 322                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 323                                                  &jx0,&jy0,&jz0);
 324
 325             /* Calculate displacement vector */
 326             dx00             = _mm256_sub_ps(ix0,jx0);
 327             dy00             = _mm256_sub_ps(iy0,jy0);
 328             dz00             = _mm256_sub_ps(iz0,jz0);
 329
 330             /* Calculate squared distance and things based on it */
 331             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 332
 333             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 334
 335             /* Load parameters for j particles */
 336             vdwjidx0A        = 2*vdwtype[jnrA+0];
 337             vdwjidx0B        = 2*vdwtype[jnrB+0];
 338             vdwjidx0C        = 2*vdwtype[jnrC+0];
 339             vdwjidx0D        = 2*vdwtype[jnrD+0];
 340             vdwjidx0E        = 2*vdwtype[jnrE+0];
 341             vdwjidx0F        = 2*vdwtype[jnrF+0];
 342             vdwjidx0G        = 2*vdwtype[jnrG+0];
 343             vdwjidx0H        = 2*vdwtype[jnrH+0];
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 350             {
 351
 352             /* Compute parameters for interactions between i and j atoms */
 353             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 354                                             vdwioffsetptr0+vdwjidx0B,
 355                                             vdwioffsetptr0+vdwjidx0C,
 356                                             vdwioffsetptr0+vdwjidx0D,
 357                                             vdwioffsetptr0+vdwjidx0E,
 358                                             vdwioffsetptr0+vdwjidx0F,
 359                                             vdwioffsetptr0+vdwjidx0G,
 360                                             vdwioffsetptr0+vdwjidx0H,
 361                                             &c6_00,&c12_00);
 362
 363             /* LENNARD-JONES DISPERSION/REPULSION */
 364
 365             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 366             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 367             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 368             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) ,
 369                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 370             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 371
 372             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 376             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 377             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 378
 379             fscal            = fvdw;
 380
 381             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 382
 383             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 384
 385             /* Calculate temporary vectorial force */
 386             tx               = _mm256_mul_ps(fscal,dx00);
 387             ty               = _mm256_mul_ps(fscal,dy00);
 388             tz               = _mm256_mul_ps(fscal,dz00);
 389
 390             /* Update vectorial force */
 391             fix0             = _mm256_add_ps(fix0,tx);
 392             fiy0             = _mm256_add_ps(fiy0,ty);
 393             fiz0             = _mm256_add_ps(fiz0,tz);
 394
 395             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 396             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 397             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 398             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 399             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 400             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 401             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 402             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 403             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 404
 405             }
 406
 407             /* Inner loop uses 41 flops */
 408         }
 409
 410         /* End of innermost loop */
 411
 412         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 413                                                  f+i_coord_offset,fshift+i_shift_offset);
 414
 415         ggid                        = gid[iidx];
 416         /* Update potential energies */
 417         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 418
 419         /* Increment number of inner iterations */
 420         inneriter                  += j_index_end - j_index_start;
 421
 422         /* Outer loop uses 7 flops */
 423     }
 424
 425     /* Increment number of outer iterations */
 426     outeriter        += nri;
 427
 428     /* Update outer/inner flops */
 429
 430     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*41);
 431 }
 432 /*
 433  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_256_single
 434  * Electrostatics interaction: None
 435  * VdW interaction:            LennardJones
 436  * Geometry:                   Particle-Particle
 437  * Calculate force/pot:        Force
 438  */
 439 void
 440 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_256_single
 441                     (t_nblist                    * gmx_restrict       nlist,
 442                      rvec                        * gmx_restrict          xx,
 443                      rvec                        * gmx_restrict          ff,
 444                      t_forcerec                  * gmx_restrict          fr,
 445                      t_mdatoms                   * gmx_restrict     mdatoms,
 446                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 447                      t_nrnb                      * gmx_restrict        nrnb)
 448 {
 449     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 450      * just 0 for non-waters.
 451      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 452      * jnr indices corresponding to data put in the four positions in the SIMD register.
 453      */
 454     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 455     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 456     int              jnrA,jnrB,jnrC,jnrD;
 457     int              jnrE,jnrF,jnrG,jnrH;
 458     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 459     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 460     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 461     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 462     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 463     real             rcutoff_scalar;
 464     real             *shiftvec,*fshift,*x,*f;
 465     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 466     real             scratch[4*DIM];
 467     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 468     real *           vdwioffsetptr0;
 469     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 470     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 471     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 472     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 473     int              nvdwtype;
 474     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 475     int              *vdwtype;
 476     real             *vdwparam;
 477     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 478     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 479     __m256           dummy_mask,cutoff_mask;
 480     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 481     __m256           one     = _mm256_set1_ps(1.0);
 482     __m256           two     = _mm256_set1_ps(2.0);
 483     x                = xx[0];
 484     f                = ff[0];
 485
 486     nri              = nlist->nri;
 487     iinr             = nlist->iinr;
 488     jindex           = nlist->jindex;
 489     jjnr             = nlist->jjnr;
 490     shiftidx         = nlist->shift;
 491     gid              = nlist->gid;
 492     shiftvec         = fr->shift_vec[0];
 493     fshift           = fr->fshift[0];
 494     nvdwtype         = fr->ntype;
 495     vdwparam         = fr->nbfp;
 496     vdwtype          = mdatoms->typeA;
 497
 498     rcutoff_scalar   = fr->rvdw;
 499     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 500     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 501
 502     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 503     rvdw             = _mm256_set1_ps(fr->rvdw);
 504
 505     /* Avoid stupid compiler warnings */
 506     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 507     j_coord_offsetA = 0;
 508     j_coord_offsetB = 0;
 509     j_coord_offsetC = 0;
 510     j_coord_offsetD = 0;
 511     j_coord_offsetE = 0;
 512     j_coord_offsetF = 0;
 513     j_coord_offsetG = 0;
 514     j_coord_offsetH = 0;
 515
 516     outeriter        = 0;
 517     inneriter        = 0;
 518
 519     for(iidx=0;iidx<4*DIM;iidx++)
 520     {
 521         scratch[iidx] = 0.0;
 522     }
 523
 524     /* Start outer loop over neighborlists */
 525     for(iidx=0; iidx<nri; iidx++)
 526     {
 527         /* Load shift vector for this list */
 528         i_shift_offset   = DIM*shiftidx[iidx];
 529
 530         /* Load limits for loop over neighbors */
 531         j_index_start    = jindex[iidx];
 532         j_index_end      = jindex[iidx+1];
 533
 534         /* Get outer coordinate index */
 535         inr              = iinr[iidx];
 536         i_coord_offset   = DIM*inr;
 537
 538         /* Load i particle coords and add shift vector */
 539         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 540
 541         fix0             = _mm256_setzero_ps();
 542         fiy0             = _mm256_setzero_ps();
 543         fiz0             = _mm256_setzero_ps();
 544
 545         /* Load parameters for i particles */
 546         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 547
 548         /* Start inner kernel loop */
 549         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 550         {
 551
 552             /* Get j neighbor index, and coordinate index */
 553             jnrA             = jjnr[jidx];
 554             jnrB             = jjnr[jidx+1];
 555             jnrC             = jjnr[jidx+2];
 556             jnrD             = jjnr[jidx+3];
 557             jnrE             = jjnr[jidx+4];
 558             jnrF             = jjnr[jidx+5];
 559             jnrG             = jjnr[jidx+6];
 560             jnrH             = jjnr[jidx+7];
 561             j_coord_offsetA  = DIM*jnrA;
 562             j_coord_offsetB  = DIM*jnrB;
 563             j_coord_offsetC  = DIM*jnrC;
 564             j_coord_offsetD  = DIM*jnrD;
 565             j_coord_offsetE  = DIM*jnrE;
 566             j_coord_offsetF  = DIM*jnrF;
 567             j_coord_offsetG  = DIM*jnrG;
 568             j_coord_offsetH  = DIM*jnrH;
 569
 570             /* load j atom coordinates */
 571             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 572                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 573                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 574                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 575                                                  &jx0,&jy0,&jz0);
 576
 577             /* Calculate displacement vector */
 578             dx00             = _mm256_sub_ps(ix0,jx0);
 579             dy00             = _mm256_sub_ps(iy0,jy0);
 580             dz00             = _mm256_sub_ps(iz0,jz0);
 581
 582             /* Calculate squared distance and things based on it */
 583             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 584
 585             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 586
 587             /* Load parameters for j particles */
 588             vdwjidx0A        = 2*vdwtype[jnrA+0];
 589             vdwjidx0B        = 2*vdwtype[jnrB+0];
 590             vdwjidx0C        = 2*vdwtype[jnrC+0];
 591             vdwjidx0D        = 2*vdwtype[jnrD+0];
 592             vdwjidx0E        = 2*vdwtype[jnrE+0];
 593             vdwjidx0F        = 2*vdwtype[jnrF+0];
 594             vdwjidx0G        = 2*vdwtype[jnrG+0];
 595             vdwjidx0H        = 2*vdwtype[jnrH+0];
 596
 597             /**************************
 598              * CALCULATE INTERACTIONS *
 599              **************************/
 600
 601             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 602             {
 603
 604             /* Compute parameters for interactions between i and j atoms */
 605             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 606                                             vdwioffsetptr0+vdwjidx0B,
 607                                             vdwioffsetptr0+vdwjidx0C,
 608                                             vdwioffsetptr0+vdwjidx0D,
 609                                             vdwioffsetptr0+vdwjidx0E,
 610                                             vdwioffsetptr0+vdwjidx0F,
 611                                             vdwioffsetptr0+vdwjidx0G,
 612                                             vdwioffsetptr0+vdwjidx0H,
 613                                             &c6_00,&c12_00);
 614
 615             /* LENNARD-JONES DISPERSION/REPULSION */
 616
 617             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 618             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 619
 620             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 621
 622             fscal            = fvdw;
 623
 624             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 625
 626             /* Calculate temporary vectorial force */
 627             tx               = _mm256_mul_ps(fscal,dx00);
 628             ty               = _mm256_mul_ps(fscal,dy00);
 629             tz               = _mm256_mul_ps(fscal,dz00);
 630
 631             /* Update vectorial force */
 632             fix0             = _mm256_add_ps(fix0,tx);
 633             fiy0             = _mm256_add_ps(fiy0,ty);
 634             fiz0             = _mm256_add_ps(fiz0,tz);
 635
 636             fjptrA             = f+j_coord_offsetA;
 637             fjptrB             = f+j_coord_offsetB;
 638             fjptrC             = f+j_coord_offsetC;
 639             fjptrD             = f+j_coord_offsetD;
 640             fjptrE             = f+j_coord_offsetE;
 641             fjptrF             = f+j_coord_offsetF;
 642             fjptrG             = f+j_coord_offsetG;
 643             fjptrH             = f+j_coord_offsetH;
 644             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 645
 646             }
 647
 648             /* Inner loop uses 30 flops */
 649         }
 650
 651         if(jidx<j_index_end)
 652         {
 653
 654             /* Get j neighbor index, and coordinate index */
 655             jnrlistA         = jjnr[jidx];
 656             jnrlistB         = jjnr[jidx+1];
 657             jnrlistC         = jjnr[jidx+2];
 658             jnrlistD         = jjnr[jidx+3];
 659             jnrlistE         = jjnr[jidx+4];
 660             jnrlistF         = jjnr[jidx+5];
 661             jnrlistG         = jjnr[jidx+6];
 662             jnrlistH         = jjnr[jidx+7];
 663             /* Sign of each element will be negative for non-real atoms.
 664              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 665              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 666              */
 667             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 668                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 669
 670             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 671             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 672             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 673             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 674             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 675             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 676             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 677             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 678             j_coord_offsetA  = DIM*jnrA;
 679             j_coord_offsetB  = DIM*jnrB;
 680             j_coord_offsetC  = DIM*jnrC;
 681             j_coord_offsetD  = DIM*jnrD;
 682             j_coord_offsetE  = DIM*jnrE;
 683             j_coord_offsetF  = DIM*jnrF;
 684             j_coord_offsetG  = DIM*jnrG;
 685             j_coord_offsetH  = DIM*jnrH;
 686
 687             /* load j atom coordinates */
 688             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 689                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 690                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 691                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 692                                                  &jx0,&jy0,&jz0);
 693
 694             /* Calculate displacement vector */
 695             dx00             = _mm256_sub_ps(ix0,jx0);
 696             dy00             = _mm256_sub_ps(iy0,jy0);
 697             dz00             = _mm256_sub_ps(iz0,jz0);
 698
 699             /* Calculate squared distance and things based on it */
 700             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 701
 702             rinvsq00         = gmx_mm256_inv_ps(rsq00);
 703
 704             /* Load parameters for j particles */
 705             vdwjidx0A        = 2*vdwtype[jnrA+0];
 706             vdwjidx0B        = 2*vdwtype[jnrB+0];
 707             vdwjidx0C        = 2*vdwtype[jnrC+0];
 708             vdwjidx0D        = 2*vdwtype[jnrD+0];
 709             vdwjidx0E        = 2*vdwtype[jnrE+0];
 710             vdwjidx0F        = 2*vdwtype[jnrF+0];
 711             vdwjidx0G        = 2*vdwtype[jnrG+0];
 712             vdwjidx0H        = 2*vdwtype[jnrH+0];
 713
 714             /**************************
 715              * CALCULATE INTERACTIONS *
 716              **************************/
 717
 718             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 719             {
 720
 721             /* Compute parameters for interactions between i and j atoms */
 722             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 723                                             vdwioffsetptr0+vdwjidx0B,
 724                                             vdwioffsetptr0+vdwjidx0C,
 725                                             vdwioffsetptr0+vdwjidx0D,
 726                                             vdwioffsetptr0+vdwjidx0E,
 727                                             vdwioffsetptr0+vdwjidx0F,
 728                                             vdwioffsetptr0+vdwjidx0G,
 729                                             vdwioffsetptr0+vdwjidx0H,
 730                                             &c6_00,&c12_00);
 731
 732             /* LENNARD-JONES DISPERSION/REPULSION */
 733
 734             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 735             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 736
 737             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 738
 739             fscal            = fvdw;
 740
 741             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 742
 743             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 744
 745             /* Calculate temporary vectorial force */
 746             tx               = _mm256_mul_ps(fscal,dx00);
 747             ty               = _mm256_mul_ps(fscal,dy00);
 748             tz               = _mm256_mul_ps(fscal,dz00);
 749
 750             /* Update vectorial force */
 751             fix0             = _mm256_add_ps(fix0,tx);
 752             fiy0             = _mm256_add_ps(fiy0,ty);
 753             fiz0             = _mm256_add_ps(fiz0,tz);
 754
 755             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 756             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 757             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 758             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 759             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 760             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 761             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 762             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 763             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 764
 765             }
 766
 767             /* Inner loop uses 30 flops */
 768         }
 769
 770         /* End of innermost loop */
 771
 772         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 773                                                  f+i_coord_offset,fshift+i_shift_offset);
 774
 775         /* Increment number of inner iterations */
 776         inneriter                  += j_index_end - j_index_start;
 777
 778         /* Outer loop uses 6 flops */
 779     }
 780
 781     /* Increment number of outer iterations */
 782     outeriter        += nri;
 783
 784     /* Update outer/inner flops */
 785
 786     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*30);
 787 }