src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwLJSw_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: None
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     int              nvdwtype;
  90     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  94     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  95     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  96     real             rswitch_scalar,d_scalar;
  97     __m256d          dummy_mask,cutoff_mask;
  98     __m128           tmpmask0,tmpmask1;
  99     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 100     __m256d          one     = _mm256_set1_pd(1.0);
 101     __m256d          two     = _mm256_set1_pd(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     nvdwtype         = fr->ntype;
 114     vdwparam         = fr->nbfp;
 115     vdwtype          = mdatoms->typeA;
 116
 117     rcutoff_scalar   = fr->rvdw;
 118     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 119     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 120
 121     rswitch_scalar   = fr->rvdw_switch;
 122     rswitch          = _mm256_set1_pd(rswitch_scalar);
 123     /* Setup switch parameters */
 124     d_scalar         = rcutoff_scalar-rswitch_scalar;
 125     d                = _mm256_set1_pd(d_scalar);
 126     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 127     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 128     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 129     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 130     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 131     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 132
 133     /* Avoid stupid compiler warnings */
 134     jnrA = jnrB = jnrC = jnrD = 0;
 135     j_coord_offsetA = 0;
 136     j_coord_offsetB = 0;
 137     j_coord_offsetC = 0;
 138     j_coord_offsetD = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     for(iidx=0;iidx<4*DIM;iidx++)
 144     {
 145         scratch[iidx] = 0.0;
 146     }
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 164
 165         fix0             = _mm256_setzero_pd();
 166         fiy0             = _mm256_setzero_pd();
 167         fiz0             = _mm256_setzero_pd();
 168
 169         /* Load parameters for i particles */
 170         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         vvdwsum          = _mm256_setzero_pd();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 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
 189             /* load j atom coordinates */
 190             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 192                                                  &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm256_sub_pd(ix0,jx0);
 196             dy00             = _mm256_sub_pd(iy0,jy0);
 197             dz00             = _mm256_sub_pd(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 203
 204             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             vdwjidx0A        = 2*vdwtype[jnrA+0];
 208             vdwjidx0B        = 2*vdwtype[jnrB+0];
 209             vdwjidx0C        = 2*vdwtype[jnrC+0];
 210             vdwjidx0D        = 2*vdwtype[jnrD+0];
 211
 212             /**************************
 213              * CALCULATE INTERACTIONS *
 214              **************************/
 215
 216             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 217             {
 218
 219             r00              = _mm256_mul_pd(rsq00,rinv00);
 220
 221             /* Compute parameters for interactions between i and j atoms */
 222             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 223                                             vdwioffsetptr0+vdwjidx0B,
 224                                             vdwioffsetptr0+vdwjidx0C,
 225                                             vdwioffsetptr0+vdwjidx0D,
 226                                             &c6_00,&c12_00);
 227
 228             /* LENNARD-JONES DISPERSION/REPULSION */
 229
 230             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 231             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 232             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 233             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 234             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 235
 236             d                = _mm256_sub_pd(r00,rswitch);
 237             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 238             d2               = _mm256_mul_pd(d,d);
 239             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 240
 241             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 242
 243             /* Evaluate switch function */
 244             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 245             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 246             vvdw             = _mm256_mul_pd(vvdw,sw);
 247             cutoff_mask      = _mm256_cmp_pd(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_pd(vvdw,cutoff_mask);
 251             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 252
 253             fscal            = fvdw;
 254
 255             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 256
 257             /* Calculate temporary vectorial force */
 258             tx               = _mm256_mul_pd(fscal,dx00);
 259             ty               = _mm256_mul_pd(fscal,dy00);
 260             tz               = _mm256_mul_pd(fscal,dz00);
 261
 262             /* Update vectorial force */
 263             fix0             = _mm256_add_pd(fix0,tx);
 264             fiy0             = _mm256_add_pd(fiy0,ty);
 265             fiz0             = _mm256_add_pd(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             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 272
 273             }
 274
 275             /* Inner loop uses 59 flops */
 276         }
 277
 278         if(jidx<j_index_end)
 279         {
 280
 281             /* Get j neighbor index, and coordinate index */
 282             jnrlistA         = jjnr[jidx];
 283             jnrlistB         = jjnr[jidx+1];
 284             jnrlistC         = jjnr[jidx+2];
 285             jnrlistD         = jjnr[jidx+3];
 286             /* Sign of each element will be negative for non-real atoms.
 287              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 288              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 289              */
 290             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 291
 292             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 293             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 294             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 295
 296             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 297             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 298             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 299             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 300             j_coord_offsetA  = DIM*jnrA;
 301             j_coord_offsetB  = DIM*jnrB;
 302             j_coord_offsetC  = DIM*jnrC;
 303             j_coord_offsetD  = DIM*jnrD;
 304
 305             /* load j atom coordinates */
 306             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 307                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 308                                                  &jx0,&jy0,&jz0);
 309
 310             /* Calculate displacement vector */
 311             dx00             = _mm256_sub_pd(ix0,jx0);
 312             dy00             = _mm256_sub_pd(iy0,jy0);
 313             dz00             = _mm256_sub_pd(iz0,jz0);
 314
 315             /* Calculate squared distance and things based on it */
 316             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 317
 318             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 319
 320             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 321
 322             /* Load parameters for j particles */
 323             vdwjidx0A        = 2*vdwtype[jnrA+0];
 324             vdwjidx0B        = 2*vdwtype[jnrB+0];
 325             vdwjidx0C        = 2*vdwtype[jnrC+0];
 326             vdwjidx0D        = 2*vdwtype[jnrD+0];
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 333             {
 334
 335             r00              = _mm256_mul_pd(rsq00,rinv00);
 336             r00              = _mm256_andnot_pd(dummy_mask,r00);
 337
 338             /* Compute parameters for interactions between i and j atoms */
 339             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 340                                             vdwioffsetptr0+vdwjidx0B,
 341                                             vdwioffsetptr0+vdwjidx0C,
 342                                             vdwioffsetptr0+vdwjidx0D,
 343                                             &c6_00,&c12_00);
 344
 345             /* LENNARD-JONES DISPERSION/REPULSION */
 346
 347             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 348             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 349             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 350             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 351             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 352
 353             d                = _mm256_sub_pd(r00,rswitch);
 354             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 355             d2               = _mm256_mul_pd(d,d);
 356             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 357
 358             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 359
 360             /* Evaluate switch function */
 361             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 362             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 363             vvdw             = _mm256_mul_pd(vvdw,sw);
 364             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 368             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 369             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 370
 371             fscal            = fvdw;
 372
 373             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 374
 375             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = _mm256_mul_pd(fscal,dx00);
 379             ty               = _mm256_mul_pd(fscal,dy00);
 380             tz               = _mm256_mul_pd(fscal,dz00);
 381
 382             /* Update vectorial force */
 383             fix0             = _mm256_add_pd(fix0,tx);
 384             fiy0             = _mm256_add_pd(fiy0,ty);
 385             fiz0             = _mm256_add_pd(fiz0,tz);
 386
 387             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 388             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 389             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 390             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 391             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 392
 393             }
 394
 395             /* Inner loop uses 60 flops */
 396         }
 397
 398         /* End of innermost loop */
 399
 400         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 401                                                  f+i_coord_offset,fshift+i_shift_offset);
 402
 403         ggid                        = gid[iidx];
 404         /* Update potential energies */
 405         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 406
 407         /* Increment number of inner iterations */
 408         inneriter                  += j_index_end - j_index_start;
 409
 410         /* Outer loop uses 7 flops */
 411     }
 412
 413     /* Increment number of outer iterations */
 414     outeriter        += nri;
 415
 416     /* Update outer/inner flops */
 417
 418     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 419 }
 420 /*
 421  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_double
 422  * Electrostatics interaction: None
 423  * VdW interaction:            LennardJones
 424  * Geometry:                   Particle-Particle
 425  * Calculate force/pot:        Force
 426  */
 427 void
 428 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_double
 429                     (t_nblist                    * gmx_restrict       nlist,
 430                      rvec                        * gmx_restrict          xx,
 431                      rvec                        * gmx_restrict          ff,
 432                      t_forcerec                  * gmx_restrict          fr,
 433                      t_mdatoms                   * gmx_restrict     mdatoms,
 434                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 435                      t_nrnb                      * gmx_restrict        nrnb)
 436 {
 437     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 438      * just 0 for non-waters.
 439      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 440      * jnr indices corresponding to data put in the four positions in the SIMD register.
 441      */
 442     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 443     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 444     int              jnrA,jnrB,jnrC,jnrD;
 445     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 446     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 447     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 448     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 449     real             rcutoff_scalar;
 450     real             *shiftvec,*fshift,*x,*f;
 451     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 452     real             scratch[4*DIM];
 453     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 454     real *           vdwioffsetptr0;
 455     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 456     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 457     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 458     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 459     int              nvdwtype;
 460     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 461     int              *vdwtype;
 462     real             *vdwparam;
 463     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 464     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 465     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 466     real             rswitch_scalar,d_scalar;
 467     __m256d          dummy_mask,cutoff_mask;
 468     __m128           tmpmask0,tmpmask1;
 469     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 470     __m256d          one     = _mm256_set1_pd(1.0);
 471     __m256d          two     = _mm256_set1_pd(2.0);
 472     x                = xx[0];
 473     f                = ff[0];
 474
 475     nri              = nlist->nri;
 476     iinr             = nlist->iinr;
 477     jindex           = nlist->jindex;
 478     jjnr             = nlist->jjnr;
 479     shiftidx         = nlist->shift;
 480     gid              = nlist->gid;
 481     shiftvec         = fr->shift_vec[0];
 482     fshift           = fr->fshift[0];
 483     nvdwtype         = fr->ntype;
 484     vdwparam         = fr->nbfp;
 485     vdwtype          = mdatoms->typeA;
 486
 487     rcutoff_scalar   = fr->rvdw;
 488     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 489     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 490
 491     rswitch_scalar   = fr->rvdw_switch;
 492     rswitch          = _mm256_set1_pd(rswitch_scalar);
 493     /* Setup switch parameters */
 494     d_scalar         = rcutoff_scalar-rswitch_scalar;
 495     d                = _mm256_set1_pd(d_scalar);
 496     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 497     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 498     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 499     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 500     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 501     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 502
 503     /* Avoid stupid compiler warnings */
 504     jnrA = jnrB = jnrC = jnrD = 0;
 505     j_coord_offsetA = 0;
 506     j_coord_offsetB = 0;
 507     j_coord_offsetC = 0;
 508     j_coord_offsetD = 0;
 509
 510     outeriter        = 0;
 511     inneriter        = 0;
 512
 513     for(iidx=0;iidx<4*DIM;iidx++)
 514     {
 515         scratch[iidx] = 0.0;
 516     }
 517
 518     /* Start outer loop over neighborlists */
 519     for(iidx=0; iidx<nri; iidx++)
 520     {
 521         /* Load shift vector for this list */
 522         i_shift_offset   = DIM*shiftidx[iidx];
 523
 524         /* Load limits for loop over neighbors */
 525         j_index_start    = jindex[iidx];
 526         j_index_end      = jindex[iidx+1];
 527
 528         /* Get outer coordinate index */
 529         inr              = iinr[iidx];
 530         i_coord_offset   = DIM*inr;
 531
 532         /* Load i particle coords and add shift vector */
 533         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 534
 535         fix0             = _mm256_setzero_pd();
 536         fiy0             = _mm256_setzero_pd();
 537         fiz0             = _mm256_setzero_pd();
 538
 539         /* Load parameters for i particles */
 540         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 541
 542         /* Start inner kernel loop */
 543         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 544         {
 545
 546             /* Get j neighbor index, and coordinate index */
 547             jnrA             = jjnr[jidx];
 548             jnrB             = jjnr[jidx+1];
 549             jnrC             = jjnr[jidx+2];
 550             jnrD             = jjnr[jidx+3];
 551             j_coord_offsetA  = DIM*jnrA;
 552             j_coord_offsetB  = DIM*jnrB;
 553             j_coord_offsetC  = DIM*jnrC;
 554             j_coord_offsetD  = DIM*jnrD;
 555
 556             /* load j atom coordinates */
 557             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 558                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 559                                                  &jx0,&jy0,&jz0);
 560
 561             /* Calculate displacement vector */
 562             dx00             = _mm256_sub_pd(ix0,jx0);
 563             dy00             = _mm256_sub_pd(iy0,jy0);
 564             dz00             = _mm256_sub_pd(iz0,jz0);
 565
 566             /* Calculate squared distance and things based on it */
 567             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 568
 569             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 570
 571             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 572
 573             /* Load parameters for j particles */
 574             vdwjidx0A        = 2*vdwtype[jnrA+0];
 575             vdwjidx0B        = 2*vdwtype[jnrB+0];
 576             vdwjidx0C        = 2*vdwtype[jnrC+0];
 577             vdwjidx0D        = 2*vdwtype[jnrD+0];
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 584             {
 585
 586             r00              = _mm256_mul_pd(rsq00,rinv00);
 587
 588             /* Compute parameters for interactions between i and j atoms */
 589             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 590                                             vdwioffsetptr0+vdwjidx0B,
 591                                             vdwioffsetptr0+vdwjidx0C,
 592                                             vdwioffsetptr0+vdwjidx0D,
 593                                             &c6_00,&c12_00);
 594
 595             /* LENNARD-JONES DISPERSION/REPULSION */
 596
 597             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 598             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 599             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 600             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 601             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 602
 603             d                = _mm256_sub_pd(r00,rswitch);
 604             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 605             d2               = _mm256_mul_pd(d,d);
 606             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 607
 608             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 609
 610             /* Evaluate switch function */
 611             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 612             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 613             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 614
 615             fscal            = fvdw;
 616
 617             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 618
 619             /* Calculate temporary vectorial force */
 620             tx               = _mm256_mul_pd(fscal,dx00);
 621             ty               = _mm256_mul_pd(fscal,dy00);
 622             tz               = _mm256_mul_pd(fscal,dz00);
 623
 624             /* Update vectorial force */
 625             fix0             = _mm256_add_pd(fix0,tx);
 626             fiy0             = _mm256_add_pd(fiy0,ty);
 627             fiz0             = _mm256_add_pd(fiz0,tz);
 628
 629             fjptrA             = f+j_coord_offsetA;
 630             fjptrB             = f+j_coord_offsetB;
 631             fjptrC             = f+j_coord_offsetC;
 632             fjptrD             = f+j_coord_offsetD;
 633             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 634
 635             }
 636
 637             /* Inner loop uses 56 flops */
 638         }
 639
 640         if(jidx<j_index_end)
 641         {
 642
 643             /* Get j neighbor index, and coordinate index */
 644             jnrlistA         = jjnr[jidx];
 645             jnrlistB         = jjnr[jidx+1];
 646             jnrlistC         = jjnr[jidx+2];
 647             jnrlistD         = jjnr[jidx+3];
 648             /* Sign of each element will be negative for non-real atoms.
 649              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 650              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 651              */
 652             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 653
 654             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 655             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 656             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 657
 658             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 659             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 660             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 661             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 662             j_coord_offsetA  = DIM*jnrA;
 663             j_coord_offsetB  = DIM*jnrB;
 664             j_coord_offsetC  = DIM*jnrC;
 665             j_coord_offsetD  = DIM*jnrD;
 666
 667             /* load j atom coordinates */
 668             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 669                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 670                                                  &jx0,&jy0,&jz0);
 671
 672             /* Calculate displacement vector */
 673             dx00             = _mm256_sub_pd(ix0,jx0);
 674             dy00             = _mm256_sub_pd(iy0,jy0);
 675             dz00             = _mm256_sub_pd(iz0,jz0);
 676
 677             /* Calculate squared distance and things based on it */
 678             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 679
 680             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 681
 682             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 683
 684             /* Load parameters for j particles */
 685             vdwjidx0A        = 2*vdwtype[jnrA+0];
 686             vdwjidx0B        = 2*vdwtype[jnrB+0];
 687             vdwjidx0C        = 2*vdwtype[jnrC+0];
 688             vdwjidx0D        = 2*vdwtype[jnrD+0];
 689
 690             /**************************
 691              * CALCULATE INTERACTIONS *
 692              **************************/
 693
 694             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 695             {
 696
 697             r00              = _mm256_mul_pd(rsq00,rinv00);
 698             r00              = _mm256_andnot_pd(dummy_mask,r00);
 699
 700             /* Compute parameters for interactions between i and j atoms */
 701             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 702                                             vdwioffsetptr0+vdwjidx0B,
 703                                             vdwioffsetptr0+vdwjidx0C,
 704                                             vdwioffsetptr0+vdwjidx0D,
 705                                             &c6_00,&c12_00);
 706
 707             /* LENNARD-JONES DISPERSION/REPULSION */
 708
 709             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 710             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 711             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 712             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 713             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 714
 715             d                = _mm256_sub_pd(r00,rswitch);
 716             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 717             d2               = _mm256_mul_pd(d,d);
 718             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 719
 720             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 721
 722             /* Evaluate switch function */
 723             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 724             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 725             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 726
 727             fscal            = fvdw;
 728
 729             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 730
 731             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 732
 733             /* Calculate temporary vectorial force */
 734             tx               = _mm256_mul_pd(fscal,dx00);
 735             ty               = _mm256_mul_pd(fscal,dy00);
 736             tz               = _mm256_mul_pd(fscal,dz00);
 737
 738             /* Update vectorial force */
 739             fix0             = _mm256_add_pd(fix0,tx);
 740             fiy0             = _mm256_add_pd(fiy0,ty);
 741             fiz0             = _mm256_add_pd(fiz0,tz);
 742
 743             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 744             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 745             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 746             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 747             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 748
 749             }
 750
 751             /* Inner loop uses 57 flops */
 752         }
 753
 754         /* End of innermost loop */
 755
 756         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 757                                                  f+i_coord_offset,fshift+i_shift_offset);
 758
 759         /* Increment number of inner iterations */
 760         inneriter                  += j_index_end - j_index_start;
 761
 762         /* Outer loop uses 6 flops */
 763     }
 764
 765     /* Increment number of outer iterations */
 766     outeriter        += nri;
 767
 768     /* Update outer/inner flops */
 769
 770     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*57);
 771 }