src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecNone_VdwLJSh_GeomP1P1_avx_256_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_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/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_256_single
  51  * Electrostatics interaction: None
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_256_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrE,jnrF,jnrG,jnrH;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  83     real             scratch[4*DIM];
  84     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     real *           vdwioffsetptr0;
  86     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  88     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     int              nvdwtype;
  91     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
  95     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
  96     __m256           dummy_mask,cutoff_mask;
  97     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  98     __m256           one     = _mm256_set1_ps(1.0);
  99     __m256           two     = _mm256_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     nvdwtype         = fr->ntype;
 112     vdwparam         = fr->nbfp;
 113     vdwtype          = mdatoms->typeA;
 114
 115     rcutoff_scalar   = fr->ic->rvdw;
 116     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 117     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 118
 119     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 120     rvdw             = _mm256_set1_ps(fr->ic->rvdw);
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128     j_coord_offsetE = 0;
 129     j_coord_offsetF = 0;
 130     j_coord_offsetG = 0;
 131     j_coord_offsetH = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 157
 158         fix0             = _mm256_setzero_ps();
 159         fiy0             = _mm256_setzero_ps();
 160         fiz0             = _mm256_setzero_ps();
 161
 162         /* Load parameters for i particles */
 163         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 164
 165         /* Reset potential sums */
 166         vvdwsum          = _mm256_setzero_ps();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             jnrE             = jjnr[jidx+4];
 178             jnrF             = jjnr[jidx+5];
 179             jnrG             = jjnr[jidx+6];
 180             jnrH             = jjnr[jidx+7];
 181             j_coord_offsetA  = DIM*jnrA;
 182             j_coord_offsetB  = DIM*jnrB;
 183             j_coord_offsetC  = DIM*jnrC;
 184             j_coord_offsetD  = DIM*jnrD;
 185             j_coord_offsetE  = DIM*jnrE;
 186             j_coord_offsetF  = DIM*jnrF;
 187             j_coord_offsetG  = DIM*jnrG;
 188             j_coord_offsetH  = DIM*jnrH;
 189
 190             /* load j atom coordinates */
 191             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 192                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 193                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 194                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 195                                                  &jx0,&jy0,&jz0);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm256_sub_ps(ix0,jx0);
 199             dy00             = _mm256_sub_ps(iy0,jy0);
 200             dz00             = _mm256_sub_ps(iz0,jz0);
 201
 202             /* Calculate squared distance and things based on it */
 203             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 204
 205             rinvsq00         = avx256_inv_f(rsq00);
 206
 207             /* Load parameters for j particles */
 208             vdwjidx0A        = 2*vdwtype[jnrA+0];
 209             vdwjidx0B        = 2*vdwtype[jnrB+0];
 210             vdwjidx0C        = 2*vdwtype[jnrC+0];
 211             vdwjidx0D        = 2*vdwtype[jnrD+0];
 212             vdwjidx0E        = 2*vdwtype[jnrE+0];
 213             vdwjidx0F        = 2*vdwtype[jnrF+0];
 214             vdwjidx0G        = 2*vdwtype[jnrG+0];
 215             vdwjidx0H        = 2*vdwtype[jnrH+0];
 216
 217             /**************************
 218              * CALCULATE INTERACTIONS *
 219              **************************/
 220
 221             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 222             {
 223
 224             /* Compute parameters for interactions between i and j atoms */
 225             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 226                                             vdwioffsetptr0+vdwjidx0B,
 227                                             vdwioffsetptr0+vdwjidx0C,
 228                                             vdwioffsetptr0+vdwjidx0D,
 229                                             vdwioffsetptr0+vdwjidx0E,
 230                                             vdwioffsetptr0+vdwjidx0F,
 231                                             vdwioffsetptr0+vdwjidx0G,
 232                                             vdwioffsetptr0+vdwjidx0H,
 233                                             &c6_00,&c12_00);
 234
 235             /* LENNARD-JONES DISPERSION/REPULSION */
 236
 237             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 238             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 239             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 240             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) ,
 241                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 242             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 243
 244             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 245
 246             /* Update potential sum for this i atom from the interaction with this j atom. */
 247             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 248             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 249
 250             fscal            = fvdw;
 251
 252             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 253
 254             /* Calculate temporary vectorial force */
 255             tx               = _mm256_mul_ps(fscal,dx00);
 256             ty               = _mm256_mul_ps(fscal,dy00);
 257             tz               = _mm256_mul_ps(fscal,dz00);
 258
 259             /* Update vectorial force */
 260             fix0             = _mm256_add_ps(fix0,tx);
 261             fiy0             = _mm256_add_ps(fiy0,ty);
 262             fiz0             = _mm256_add_ps(fiz0,tz);
 263
 264             fjptrA             = f+j_coord_offsetA;
 265             fjptrB             = f+j_coord_offsetB;
 266             fjptrC             = f+j_coord_offsetC;
 267             fjptrD             = f+j_coord_offsetD;
 268             fjptrE             = f+j_coord_offsetE;
 269             fjptrF             = f+j_coord_offsetF;
 270             fjptrG             = f+j_coord_offsetG;
 271             fjptrH             = f+j_coord_offsetH;
 272             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 273
 274             }
 275
 276             /* Inner loop uses 41 flops */
 277         }
 278
 279         if(jidx<j_index_end)
 280         {
 281
 282             /* Get j neighbor index, and coordinate index */
 283             jnrlistA         = jjnr[jidx];
 284             jnrlistB         = jjnr[jidx+1];
 285             jnrlistC         = jjnr[jidx+2];
 286             jnrlistD         = jjnr[jidx+3];
 287             jnrlistE         = jjnr[jidx+4];
 288             jnrlistF         = jjnr[jidx+5];
 289             jnrlistG         = jjnr[jidx+6];
 290             jnrlistH         = jjnr[jidx+7];
 291             /* Sign of each element will be negative for non-real atoms.
 292              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 293              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 294              */
 295             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 296                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 297
 298             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 299             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 300             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 301             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 302             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 303             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 304             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 305             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 306             j_coord_offsetA  = DIM*jnrA;
 307             j_coord_offsetB  = DIM*jnrB;
 308             j_coord_offsetC  = DIM*jnrC;
 309             j_coord_offsetD  = DIM*jnrD;
 310             j_coord_offsetE  = DIM*jnrE;
 311             j_coord_offsetF  = DIM*jnrF;
 312             j_coord_offsetG  = DIM*jnrG;
 313             j_coord_offsetH  = DIM*jnrH;
 314
 315             /* load j atom coordinates */
 316             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 317                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 318                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 319                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 320                                                  &jx0,&jy0,&jz0);
 321
 322             /* Calculate displacement vector */
 323             dx00             = _mm256_sub_ps(ix0,jx0);
 324             dy00             = _mm256_sub_ps(iy0,jy0);
 325             dz00             = _mm256_sub_ps(iz0,jz0);
 326
 327             /* Calculate squared distance and things based on it */
 328             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 329
 330             rinvsq00         = avx256_inv_f(rsq00);
 331
 332             /* Load parameters for j particles */
 333             vdwjidx0A        = 2*vdwtype[jnrA+0];
 334             vdwjidx0B        = 2*vdwtype[jnrB+0];
 335             vdwjidx0C        = 2*vdwtype[jnrC+0];
 336             vdwjidx0D        = 2*vdwtype[jnrD+0];
 337             vdwjidx0E        = 2*vdwtype[jnrE+0];
 338             vdwjidx0F        = 2*vdwtype[jnrF+0];
 339             vdwjidx0G        = 2*vdwtype[jnrG+0];
 340             vdwjidx0H        = 2*vdwtype[jnrH+0];
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 347             {
 348
 349             /* Compute parameters for interactions between i and j atoms */
 350             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 351                                             vdwioffsetptr0+vdwjidx0B,
 352                                             vdwioffsetptr0+vdwjidx0C,
 353                                             vdwioffsetptr0+vdwjidx0D,
 354                                             vdwioffsetptr0+vdwjidx0E,
 355                                             vdwioffsetptr0+vdwjidx0F,
 356                                             vdwioffsetptr0+vdwjidx0G,
 357                                             vdwioffsetptr0+vdwjidx0H,
 358                                             &c6_00,&c12_00);
 359
 360             /* LENNARD-JONES DISPERSION/REPULSION */
 361
 362             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 363             vvdw6            = _mm256_mul_ps(c6_00,rinvsix);
 364             vvdw12           = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
 365             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) ,
 366                                           _mm256_mul_ps( _mm256_sub_ps(vvdw6,_mm256_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 367             fvdw             = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
 368
 369             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             vvdw             = _mm256_and_ps(vvdw,cutoff_mask);
 373             vvdw             = _mm256_andnot_ps(dummy_mask,vvdw);
 374             vvdwsum          = _mm256_add_ps(vvdwsum,vvdw);
 375
 376             fscal            = fvdw;
 377
 378             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 379
 380             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 381
 382             /* Calculate temporary vectorial force */
 383             tx               = _mm256_mul_ps(fscal,dx00);
 384             ty               = _mm256_mul_ps(fscal,dy00);
 385             tz               = _mm256_mul_ps(fscal,dz00);
 386
 387             /* Update vectorial force */
 388             fix0             = _mm256_add_ps(fix0,tx);
 389             fiy0             = _mm256_add_ps(fiy0,ty);
 390             fiz0             = _mm256_add_ps(fiz0,tz);
 391
 392             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 393             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 394             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 395             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 396             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 397             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 398             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 399             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 400             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 401
 402             }
 403
 404             /* Inner loop uses 41 flops */
 405         }
 406
 407         /* End of innermost loop */
 408
 409         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 410                                                  f+i_coord_offset,fshift+i_shift_offset);
 411
 412         ggid                        = gid[iidx];
 413         /* Update potential energies */
 414         gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 415
 416         /* Increment number of inner iterations */
 417         inneriter                  += j_index_end - j_index_start;
 418
 419         /* Outer loop uses 7 flops */
 420     }
 421
 422     /* Increment number of outer iterations */
 423     outeriter        += nri;
 424
 425     /* Update outer/inner flops */
 426
 427     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*41);
 428 }
 429 /*
 430  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_256_single
 431  * Electrostatics interaction: None
 432  * VdW interaction:            LennardJones
 433  * Geometry:                   Particle-Particle
 434  * Calculate force/pot:        Force
 435  */
 436 void
 437 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_256_single
 438                     (t_nblist                    * gmx_restrict       nlist,
 439                      rvec                        * gmx_restrict          xx,
 440                      rvec                        * gmx_restrict          ff,
 441                      struct t_forcerec           * gmx_restrict          fr,
 442                      t_mdatoms                   * gmx_restrict     mdatoms,
 443                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 444                      t_nrnb                      * gmx_restrict        nrnb)
 445 {
 446     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 447      * just 0 for non-waters.
 448      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 449      * jnr indices corresponding to data put in the four positions in the SIMD register.
 450      */
 451     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 452     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 453     int              jnrA,jnrB,jnrC,jnrD;
 454     int              jnrE,jnrF,jnrG,jnrH;
 455     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 456     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 457     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 458     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 459     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 460     real             rcutoff_scalar;
 461     real             *shiftvec,*fshift,*x,*f;
 462     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 463     real             scratch[4*DIM];
 464     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 465     real *           vdwioffsetptr0;
 466     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 467     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 468     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 469     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 470     int              nvdwtype;
 471     __m256           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 472     int              *vdwtype;
 473     real             *vdwparam;
 474     __m256           one_sixth   = _mm256_set1_ps(1.0/6.0);
 475     __m256           one_twelfth = _mm256_set1_ps(1.0/12.0);
 476     __m256           dummy_mask,cutoff_mask;
 477     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 478     __m256           one     = _mm256_set1_ps(1.0);
 479     __m256           two     = _mm256_set1_ps(2.0);
 480     x                = xx[0];
 481     f                = ff[0];
 482
 483     nri              = nlist->nri;
 484     iinr             = nlist->iinr;
 485     jindex           = nlist->jindex;
 486     jjnr             = nlist->jjnr;
 487     shiftidx         = nlist->shift;
 488     gid              = nlist->gid;
 489     shiftvec         = fr->shift_vec[0];
 490     fshift           = fr->fshift[0];
 491     nvdwtype         = fr->ntype;
 492     vdwparam         = fr->nbfp;
 493     vdwtype          = mdatoms->typeA;
 494
 495     rcutoff_scalar   = fr->ic->rvdw;
 496     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 497     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 498
 499     sh_vdw_invrcut6  = _mm256_set1_ps(fr->ic->sh_invrc6);
 500     rvdw             = _mm256_set1_ps(fr->ic->rvdw);
 501
 502     /* Avoid stupid compiler warnings */
 503     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 504     j_coord_offsetA = 0;
 505     j_coord_offsetB = 0;
 506     j_coord_offsetC = 0;
 507     j_coord_offsetD = 0;
 508     j_coord_offsetE = 0;
 509     j_coord_offsetF = 0;
 510     j_coord_offsetG = 0;
 511     j_coord_offsetH = 0;
 512
 513     outeriter        = 0;
 514     inneriter        = 0;
 515
 516     for(iidx=0;iidx<4*DIM;iidx++)
 517     {
 518         scratch[iidx] = 0.0;
 519     }
 520
 521     /* Start outer loop over neighborlists */
 522     for(iidx=0; iidx<nri; iidx++)
 523     {
 524         /* Load shift vector for this list */
 525         i_shift_offset   = DIM*shiftidx[iidx];
 526
 527         /* Load limits for loop over neighbors */
 528         j_index_start    = jindex[iidx];
 529         j_index_end      = jindex[iidx+1];
 530
 531         /* Get outer coordinate index */
 532         inr              = iinr[iidx];
 533         i_coord_offset   = DIM*inr;
 534
 535         /* Load i particle coords and add shift vector */
 536         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 537
 538         fix0             = _mm256_setzero_ps();
 539         fiy0             = _mm256_setzero_ps();
 540         fiz0             = _mm256_setzero_ps();
 541
 542         /* Load parameters for i particles */
 543         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 544
 545         /* Start inner kernel loop */
 546         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 547         {
 548
 549             /* Get j neighbor index, and coordinate index */
 550             jnrA             = jjnr[jidx];
 551             jnrB             = jjnr[jidx+1];
 552             jnrC             = jjnr[jidx+2];
 553             jnrD             = jjnr[jidx+3];
 554             jnrE             = jjnr[jidx+4];
 555             jnrF             = jjnr[jidx+5];
 556             jnrG             = jjnr[jidx+6];
 557             jnrH             = jjnr[jidx+7];
 558             j_coord_offsetA  = DIM*jnrA;
 559             j_coord_offsetB  = DIM*jnrB;
 560             j_coord_offsetC  = DIM*jnrC;
 561             j_coord_offsetD  = DIM*jnrD;
 562             j_coord_offsetE  = DIM*jnrE;
 563             j_coord_offsetF  = DIM*jnrF;
 564             j_coord_offsetG  = DIM*jnrG;
 565             j_coord_offsetH  = DIM*jnrH;
 566
 567             /* load j atom coordinates */
 568             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 569                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 570                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 571                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 572                                                  &jx0,&jy0,&jz0);
 573
 574             /* Calculate displacement vector */
 575             dx00             = _mm256_sub_ps(ix0,jx0);
 576             dy00             = _mm256_sub_ps(iy0,jy0);
 577             dz00             = _mm256_sub_ps(iz0,jz0);
 578
 579             /* Calculate squared distance and things based on it */
 580             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 581
 582             rinvsq00         = avx256_inv_f(rsq00);
 583
 584             /* Load parameters for j particles */
 585             vdwjidx0A        = 2*vdwtype[jnrA+0];
 586             vdwjidx0B        = 2*vdwtype[jnrB+0];
 587             vdwjidx0C        = 2*vdwtype[jnrC+0];
 588             vdwjidx0D        = 2*vdwtype[jnrD+0];
 589             vdwjidx0E        = 2*vdwtype[jnrE+0];
 590             vdwjidx0F        = 2*vdwtype[jnrF+0];
 591             vdwjidx0G        = 2*vdwtype[jnrG+0];
 592             vdwjidx0H        = 2*vdwtype[jnrH+0];
 593
 594             /**************************
 595              * CALCULATE INTERACTIONS *
 596              **************************/
 597
 598             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 599             {
 600
 601             /* Compute parameters for interactions between i and j atoms */
 602             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 603                                             vdwioffsetptr0+vdwjidx0B,
 604                                             vdwioffsetptr0+vdwjidx0C,
 605                                             vdwioffsetptr0+vdwjidx0D,
 606                                             vdwioffsetptr0+vdwjidx0E,
 607                                             vdwioffsetptr0+vdwjidx0F,
 608                                             vdwioffsetptr0+vdwjidx0G,
 609                                             vdwioffsetptr0+vdwjidx0H,
 610                                             &c6_00,&c12_00);
 611
 612             /* LENNARD-JONES DISPERSION/REPULSION */
 613
 614             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 615             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 616
 617             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 618
 619             fscal            = fvdw;
 620
 621             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 622
 623             /* Calculate temporary vectorial force */
 624             tx               = _mm256_mul_ps(fscal,dx00);
 625             ty               = _mm256_mul_ps(fscal,dy00);
 626             tz               = _mm256_mul_ps(fscal,dz00);
 627
 628             /* Update vectorial force */
 629             fix0             = _mm256_add_ps(fix0,tx);
 630             fiy0             = _mm256_add_ps(fiy0,ty);
 631             fiz0             = _mm256_add_ps(fiz0,tz);
 632
 633             fjptrA             = f+j_coord_offsetA;
 634             fjptrB             = f+j_coord_offsetB;
 635             fjptrC             = f+j_coord_offsetC;
 636             fjptrD             = f+j_coord_offsetD;
 637             fjptrE             = f+j_coord_offsetE;
 638             fjptrF             = f+j_coord_offsetF;
 639             fjptrG             = f+j_coord_offsetG;
 640             fjptrH             = f+j_coord_offsetH;
 641             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 642
 643             }
 644
 645             /* Inner loop uses 30 flops */
 646         }
 647
 648         if(jidx<j_index_end)
 649         {
 650
 651             /* Get j neighbor index, and coordinate index */
 652             jnrlistA         = jjnr[jidx];
 653             jnrlistB         = jjnr[jidx+1];
 654             jnrlistC         = jjnr[jidx+2];
 655             jnrlistD         = jjnr[jidx+3];
 656             jnrlistE         = jjnr[jidx+4];
 657             jnrlistF         = jjnr[jidx+5];
 658             jnrlistG         = jjnr[jidx+6];
 659             jnrlistH         = jjnr[jidx+7];
 660             /* Sign of each element will be negative for non-real atoms.
 661              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 662              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 663              */
 664             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 665                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 666
 667             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 668             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 669             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 670             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 671             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 672             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 673             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 674             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 675             j_coord_offsetA  = DIM*jnrA;
 676             j_coord_offsetB  = DIM*jnrB;
 677             j_coord_offsetC  = DIM*jnrC;
 678             j_coord_offsetD  = DIM*jnrD;
 679             j_coord_offsetE  = DIM*jnrE;
 680             j_coord_offsetF  = DIM*jnrF;
 681             j_coord_offsetG  = DIM*jnrG;
 682             j_coord_offsetH  = DIM*jnrH;
 683
 684             /* load j atom coordinates */
 685             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 686                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 687                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 688                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 689                                                  &jx0,&jy0,&jz0);
 690
 691             /* Calculate displacement vector */
 692             dx00             = _mm256_sub_ps(ix0,jx0);
 693             dy00             = _mm256_sub_ps(iy0,jy0);
 694             dz00             = _mm256_sub_ps(iz0,jz0);
 695
 696             /* Calculate squared distance and things based on it */
 697             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 698
 699             rinvsq00         = avx256_inv_f(rsq00);
 700
 701             /* Load parameters for j particles */
 702             vdwjidx0A        = 2*vdwtype[jnrA+0];
 703             vdwjidx0B        = 2*vdwtype[jnrB+0];
 704             vdwjidx0C        = 2*vdwtype[jnrC+0];
 705             vdwjidx0D        = 2*vdwtype[jnrD+0];
 706             vdwjidx0E        = 2*vdwtype[jnrE+0];
 707             vdwjidx0F        = 2*vdwtype[jnrF+0];
 708             vdwjidx0G        = 2*vdwtype[jnrG+0];
 709             vdwjidx0H        = 2*vdwtype[jnrH+0];
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 716             {
 717
 718             /* Compute parameters for interactions between i and j atoms */
 719             gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
 720                                             vdwioffsetptr0+vdwjidx0B,
 721                                             vdwioffsetptr0+vdwjidx0C,
 722                                             vdwioffsetptr0+vdwjidx0D,
 723                                             vdwioffsetptr0+vdwjidx0E,
 724                                             vdwioffsetptr0+vdwjidx0F,
 725                                             vdwioffsetptr0+vdwjidx0G,
 726                                             vdwioffsetptr0+vdwjidx0H,
 727                                             &c6_00,&c12_00);
 728
 729             /* LENNARD-JONES DISPERSION/REPULSION */
 730
 731             rinvsix          = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 732             fvdw             = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(c12_00,rinvsix),c6_00),_mm256_mul_ps(rinvsix,rinvsq00));
 733
 734             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 735
 736             fscal            = fvdw;
 737
 738             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 739
 740             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 741
 742             /* Calculate temporary vectorial force */
 743             tx               = _mm256_mul_ps(fscal,dx00);
 744             ty               = _mm256_mul_ps(fscal,dy00);
 745             tz               = _mm256_mul_ps(fscal,dz00);
 746
 747             /* Update vectorial force */
 748             fix0             = _mm256_add_ps(fix0,tx);
 749             fiy0             = _mm256_add_ps(fiy0,ty);
 750             fiz0             = _mm256_add_ps(fiz0,tz);
 751
 752             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 753             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 754             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 755             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 756             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 757             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 758             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 759             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 760             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 761
 762             }
 763
 764             /* Inner loop uses 30 flops */
 765         }
 766
 767         /* End of innermost loop */
 768
 769         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 770                                                  f+i_coord_offset,fshift+i_shift_offset);
 771
 772         /* Increment number of inner iterations */
 773         inneriter                  += j_index_end - j_index_start;
 774
 775         /* Outer loop uses 6 flops */
 776     }
 777
 778     /* Increment number of outer iterations */
 779     outeriter        += nri;
 780
 781     /* Update outer/inner flops */
 782
 783     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*30);
 784 }