src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwLJSw_GeomP1P1_avx_256_double.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_double 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_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: None
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_double
  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 refer to j loop unrolling done with AVX, e.g. for the four 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              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     real *           vdwioffsetptr0;
  84     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     int              nvdwtype;
  89     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  90     int              *vdwtype;
  91     real             *vdwparam;
  92     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  93     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  94     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  95     real             rswitch_scalar,d_scalar;
  96     __m256d          dummy_mask,cutoff_mask;
  97     __m128           tmpmask0,tmpmask1;
  98     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  99     __m256d          one     = _mm256_set1_pd(1.0);
 100     __m256d          two     = _mm256_set1_pd(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     nvdwtype         = fr->ntype;
 113     vdwparam         = fr->nbfp;
 114     vdwtype          = mdatoms->typeA;
 115
 116     rcutoff_scalar   = fr->ic->rvdw;
 117     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 118     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 119
 120     rswitch_scalar   = fr->ic->rvdw_switch;
 121     rswitch          = _mm256_set1_pd(rswitch_scalar);
 122     /* Setup switch parameters */
 123     d_scalar         = rcutoff_scalar-rswitch_scalar;
 124     d                = _mm256_set1_pd(d_scalar);
 125     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 126     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 127     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 128     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 129     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 130     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = jnrC = jnrD = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136     j_coord_offsetC = 0;
 137     j_coord_offsetD = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 163
 164         fix0             = _mm256_setzero_pd();
 165         fiy0             = _mm256_setzero_pd();
 166         fiz0             = _mm256_setzero_pd();
 167
 168         /* Load parameters for i particles */
 169         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 170
 171         /* Reset potential sums */
 172         vvdwsum          = _mm256_setzero_pd();
 173
 174         /* Start inner kernel loop */
 175         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 176         {
 177
 178             /* Get j neighbor index, and coordinate index */
 179             jnrA             = jjnr[jidx];
 180             jnrB             = jjnr[jidx+1];
 181             jnrC             = jjnr[jidx+2];
 182             jnrD             = jjnr[jidx+3];
 183             j_coord_offsetA  = DIM*jnrA;
 184             j_coord_offsetB  = DIM*jnrB;
 185             j_coord_offsetC  = DIM*jnrC;
 186             j_coord_offsetD  = DIM*jnrD;
 187
 188             /* load j atom coordinates */
 189             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 190                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 191                                                  &jx0,&jy0,&jz0);
 192
 193             /* Calculate displacement vector */
 194             dx00             = _mm256_sub_pd(ix0,jx0);
 195             dy00             = _mm256_sub_pd(iy0,jy0);
 196             dz00             = _mm256_sub_pd(iz0,jz0);
 197
 198             /* Calculate squared distance and things based on it */
 199             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 200
 201             rinv00           = avx256_invsqrt_d(rsq00);
 202
 203             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 204
 205             /* Load parameters for j particles */
 206             vdwjidx0A        = 2*vdwtype[jnrA+0];
 207             vdwjidx0B        = 2*vdwtype[jnrB+0];
 208             vdwjidx0C        = 2*vdwtype[jnrC+0];
 209             vdwjidx0D        = 2*vdwtype[jnrD+0];
 210
 211             /**************************
 212              * CALCULATE INTERACTIONS *
 213              **************************/
 214
 215             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 216             {
 217
 218             r00              = _mm256_mul_pd(rsq00,rinv00);
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 222                                             vdwioffsetptr0+vdwjidx0B,
 223                                             vdwioffsetptr0+vdwjidx0C,
 224                                             vdwioffsetptr0+vdwjidx0D,
 225                                             &c6_00,&c12_00);
 226
 227             /* LENNARD-JONES DISPERSION/REPULSION */
 228
 229             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 230             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 231             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 232             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 233             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 234
 235             d                = _mm256_sub_pd(r00,rswitch);
 236             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 237             d2               = _mm256_mul_pd(d,d);
 238             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)))))));
 239
 240             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 241
 242             /* Evaluate switch function */
 243             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 244             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 245             vvdw             = _mm256_mul_pd(vvdw,sw);
 246             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 250             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 251
 252             fscal            = fvdw;
 253
 254             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 255
 256             /* Calculate temporary vectorial force */
 257             tx               = _mm256_mul_pd(fscal,dx00);
 258             ty               = _mm256_mul_pd(fscal,dy00);
 259             tz               = _mm256_mul_pd(fscal,dz00);
 260
 261             /* Update vectorial force */
 262             fix0             = _mm256_add_pd(fix0,tx);
 263             fiy0             = _mm256_add_pd(fiy0,ty);
 264             fiz0             = _mm256_add_pd(fiz0,tz);
 265
 266             fjptrA             = f+j_coord_offsetA;
 267             fjptrB             = f+j_coord_offsetB;
 268             fjptrC             = f+j_coord_offsetC;
 269             fjptrD             = f+j_coord_offsetD;
 270             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 271
 272             }
 273
 274             /* Inner loop uses 59 flops */
 275         }
 276
 277         if(jidx<j_index_end)
 278         {
 279
 280             /* Get j neighbor index, and coordinate index */
 281             jnrlistA         = jjnr[jidx];
 282             jnrlistB         = jjnr[jidx+1];
 283             jnrlistC         = jjnr[jidx+2];
 284             jnrlistD         = jjnr[jidx+3];
 285             /* Sign of each element will be negative for non-real atoms.
 286              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 287              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 288              */
 289             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 290
 291             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 292             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 293             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 294
 295             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 296             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 297             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 298             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 299             j_coord_offsetA  = DIM*jnrA;
 300             j_coord_offsetB  = DIM*jnrB;
 301             j_coord_offsetC  = DIM*jnrC;
 302             j_coord_offsetD  = DIM*jnrD;
 303
 304             /* load j atom coordinates */
 305             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 306                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 307                                                  &jx0,&jy0,&jz0);
 308
 309             /* Calculate displacement vector */
 310             dx00             = _mm256_sub_pd(ix0,jx0);
 311             dy00             = _mm256_sub_pd(iy0,jy0);
 312             dz00             = _mm256_sub_pd(iz0,jz0);
 313
 314             /* Calculate squared distance and things based on it */
 315             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 316
 317             rinv00           = avx256_invsqrt_d(rsq00);
 318
 319             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 320
 321             /* Load parameters for j particles */
 322             vdwjidx0A        = 2*vdwtype[jnrA+0];
 323             vdwjidx0B        = 2*vdwtype[jnrB+0];
 324             vdwjidx0C        = 2*vdwtype[jnrC+0];
 325             vdwjidx0D        = 2*vdwtype[jnrD+0];
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 332             {
 333
 334             r00              = _mm256_mul_pd(rsq00,rinv00);
 335             r00              = _mm256_andnot_pd(dummy_mask,r00);
 336
 337             /* Compute parameters for interactions between i and j atoms */
 338             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 339                                             vdwioffsetptr0+vdwjidx0B,
 340                                             vdwioffsetptr0+vdwjidx0C,
 341                                             vdwioffsetptr0+vdwjidx0D,
 342                                             &c6_00,&c12_00);
 343
 344             /* LENNARD-JONES DISPERSION/REPULSION */
 345
 346             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 347             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 348             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 349             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 350             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 351
 352             d                = _mm256_sub_pd(r00,rswitch);
 353             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 354             d2               = _mm256_mul_pd(d,d);
 355             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)))))));
 356
 357             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 358
 359             /* Evaluate switch function */
 360             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 361             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 362             vvdw             = _mm256_mul_pd(vvdw,sw);
 363             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 367             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 368             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 369
 370             fscal            = fvdw;
 371
 372             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 373
 374             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 375
 376             /* Calculate temporary vectorial force */
 377             tx               = _mm256_mul_pd(fscal,dx00);
 378             ty               = _mm256_mul_pd(fscal,dy00);
 379             tz               = _mm256_mul_pd(fscal,dz00);
 380
 381             /* Update vectorial force */
 382             fix0             = _mm256_add_pd(fix0,tx);
 383             fiy0             = _mm256_add_pd(fiy0,ty);
 384             fiz0             = _mm256_add_pd(fiz0,tz);
 385
 386             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 387             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 388             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 389             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 390             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 391
 392             }
 393
 394             /* Inner loop uses 60 flops */
 395         }
 396
 397         /* End of innermost loop */
 398
 399         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 400                                                  f+i_coord_offset,fshift+i_shift_offset);
 401
 402         ggid                        = gid[iidx];
 403         /* Update potential energies */
 404         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 405
 406         /* Increment number of inner iterations */
 407         inneriter                  += j_index_end - j_index_start;
 408
 409         /* Outer loop uses 7 flops */
 410     }
 411
 412     /* Increment number of outer iterations */
 413     outeriter        += nri;
 414
 415     /* Update outer/inner flops */
 416
 417     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 418 }
 419 /*
 420  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_double
 421  * Electrostatics interaction: None
 422  * VdW interaction:            LennardJones
 423  * Geometry:                   Particle-Particle
 424  * Calculate force/pot:        Force
 425  */
 426 void
 427 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_double
 428                     (t_nblist                    * gmx_restrict       nlist,
 429                      rvec                        * gmx_restrict          xx,
 430                      rvec                        * gmx_restrict          ff,
 431                      struct t_forcerec           * gmx_restrict          fr,
 432                      t_mdatoms                   * gmx_restrict     mdatoms,
 433                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 434                      t_nrnb                      * gmx_restrict        nrnb)
 435 {
 436     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 437      * just 0 for non-waters.
 438      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 439      * jnr indices corresponding to data put in the four positions in the SIMD register.
 440      */
 441     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 442     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 443     int              jnrA,jnrB,jnrC,jnrD;
 444     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 445     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 446     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 447     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 448     real             rcutoff_scalar;
 449     real             *shiftvec,*fshift,*x,*f;
 450     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 451     real             scratch[4*DIM];
 452     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 453     real *           vdwioffsetptr0;
 454     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 455     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 456     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 457     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 458     int              nvdwtype;
 459     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 460     int              *vdwtype;
 461     real             *vdwparam;
 462     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 463     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 464     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 465     real             rswitch_scalar,d_scalar;
 466     __m256d          dummy_mask,cutoff_mask;
 467     __m128           tmpmask0,tmpmask1;
 468     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 469     __m256d          one     = _mm256_set1_pd(1.0);
 470     __m256d          two     = _mm256_set1_pd(2.0);
 471     x                = xx[0];
 472     f                = ff[0];
 473
 474     nri              = nlist->nri;
 475     iinr             = nlist->iinr;
 476     jindex           = nlist->jindex;
 477     jjnr             = nlist->jjnr;
 478     shiftidx         = nlist->shift;
 479     gid              = nlist->gid;
 480     shiftvec         = fr->shift_vec[0];
 481     fshift           = fr->fshift[0];
 482     nvdwtype         = fr->ntype;
 483     vdwparam         = fr->nbfp;
 484     vdwtype          = mdatoms->typeA;
 485
 486     rcutoff_scalar   = fr->ic->rvdw;
 487     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 488     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 489
 490     rswitch_scalar   = fr->ic->rvdw_switch;
 491     rswitch          = _mm256_set1_pd(rswitch_scalar);
 492     /* Setup switch parameters */
 493     d_scalar         = rcutoff_scalar-rswitch_scalar;
 494     d                = _mm256_set1_pd(d_scalar);
 495     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 496     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 497     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 498     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 499     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 500     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 501
 502     /* Avoid stupid compiler warnings */
 503     jnrA = jnrB = jnrC = jnrD = 0;
 504     j_coord_offsetA = 0;
 505     j_coord_offsetB = 0;
 506     j_coord_offsetC = 0;
 507     j_coord_offsetD = 0;
 508
 509     outeriter        = 0;
 510     inneriter        = 0;
 511
 512     for(iidx=0;iidx<4*DIM;iidx++)
 513     {
 514         scratch[iidx] = 0.0;
 515     }
 516
 517     /* Start outer loop over neighborlists */
 518     for(iidx=0; iidx<nri; iidx++)
 519     {
 520         /* Load shift vector for this list */
 521         i_shift_offset   = DIM*shiftidx[iidx];
 522
 523         /* Load limits for loop over neighbors */
 524         j_index_start    = jindex[iidx];
 525         j_index_end      = jindex[iidx+1];
 526
 527         /* Get outer coordinate index */
 528         inr              = iinr[iidx];
 529         i_coord_offset   = DIM*inr;
 530
 531         /* Load i particle coords and add shift vector */
 532         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 533
 534         fix0             = _mm256_setzero_pd();
 535         fiy0             = _mm256_setzero_pd();
 536         fiz0             = _mm256_setzero_pd();
 537
 538         /* Load parameters for i particles */
 539         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 540
 541         /* Start inner kernel loop */
 542         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 543         {
 544
 545             /* Get j neighbor index, and coordinate index */
 546             jnrA             = jjnr[jidx];
 547             jnrB             = jjnr[jidx+1];
 548             jnrC             = jjnr[jidx+2];
 549             jnrD             = jjnr[jidx+3];
 550             j_coord_offsetA  = DIM*jnrA;
 551             j_coord_offsetB  = DIM*jnrB;
 552             j_coord_offsetC  = DIM*jnrC;
 553             j_coord_offsetD  = DIM*jnrD;
 554
 555             /* load j atom coordinates */
 556             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 557                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 558                                                  &jx0,&jy0,&jz0);
 559
 560             /* Calculate displacement vector */
 561             dx00             = _mm256_sub_pd(ix0,jx0);
 562             dy00             = _mm256_sub_pd(iy0,jy0);
 563             dz00             = _mm256_sub_pd(iz0,jz0);
 564
 565             /* Calculate squared distance and things based on it */
 566             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 567
 568             rinv00           = avx256_invsqrt_d(rsq00);
 569
 570             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 571
 572             /* Load parameters for j particles */
 573             vdwjidx0A        = 2*vdwtype[jnrA+0];
 574             vdwjidx0B        = 2*vdwtype[jnrB+0];
 575             vdwjidx0C        = 2*vdwtype[jnrC+0];
 576             vdwjidx0D        = 2*vdwtype[jnrD+0];
 577
 578             /**************************
 579              * CALCULATE INTERACTIONS *
 580              **************************/
 581
 582             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 583             {
 584
 585             r00              = _mm256_mul_pd(rsq00,rinv00);
 586
 587             /* Compute parameters for interactions between i and j atoms */
 588             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 589                                             vdwioffsetptr0+vdwjidx0B,
 590                                             vdwioffsetptr0+vdwjidx0C,
 591                                             vdwioffsetptr0+vdwjidx0D,
 592                                             &c6_00,&c12_00);
 593
 594             /* LENNARD-JONES DISPERSION/REPULSION */
 595
 596             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 597             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 598             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 599             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 600             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 601
 602             d                = _mm256_sub_pd(r00,rswitch);
 603             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 604             d2               = _mm256_mul_pd(d,d);
 605             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)))))));
 606
 607             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 608
 609             /* Evaluate switch function */
 610             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 611             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 612             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 613
 614             fscal            = fvdw;
 615
 616             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 617
 618             /* Calculate temporary vectorial force */
 619             tx               = _mm256_mul_pd(fscal,dx00);
 620             ty               = _mm256_mul_pd(fscal,dy00);
 621             tz               = _mm256_mul_pd(fscal,dz00);
 622
 623             /* Update vectorial force */
 624             fix0             = _mm256_add_pd(fix0,tx);
 625             fiy0             = _mm256_add_pd(fiy0,ty);
 626             fiz0             = _mm256_add_pd(fiz0,tz);
 627
 628             fjptrA             = f+j_coord_offsetA;
 629             fjptrB             = f+j_coord_offsetB;
 630             fjptrC             = f+j_coord_offsetC;
 631             fjptrD             = f+j_coord_offsetD;
 632             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 633
 634             }
 635
 636             /* Inner loop uses 56 flops */
 637         }
 638
 639         if(jidx<j_index_end)
 640         {
 641
 642             /* Get j neighbor index, and coordinate index */
 643             jnrlistA         = jjnr[jidx];
 644             jnrlistB         = jjnr[jidx+1];
 645             jnrlistC         = jjnr[jidx+2];
 646             jnrlistD         = jjnr[jidx+3];
 647             /* Sign of each element will be negative for non-real atoms.
 648              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 649              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 650              */
 651             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 652
 653             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 654             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 655             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 656
 657             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 658             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 659             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 660             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 661             j_coord_offsetA  = DIM*jnrA;
 662             j_coord_offsetB  = DIM*jnrB;
 663             j_coord_offsetC  = DIM*jnrC;
 664             j_coord_offsetD  = DIM*jnrD;
 665
 666             /* load j atom coordinates */
 667             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 668                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 669                                                  &jx0,&jy0,&jz0);
 670
 671             /* Calculate displacement vector */
 672             dx00             = _mm256_sub_pd(ix0,jx0);
 673             dy00             = _mm256_sub_pd(iy0,jy0);
 674             dz00             = _mm256_sub_pd(iz0,jz0);
 675
 676             /* Calculate squared distance and things based on it */
 677             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 678
 679             rinv00           = avx256_invsqrt_d(rsq00);
 680
 681             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 682
 683             /* Load parameters for j particles */
 684             vdwjidx0A        = 2*vdwtype[jnrA+0];
 685             vdwjidx0B        = 2*vdwtype[jnrB+0];
 686             vdwjidx0C        = 2*vdwtype[jnrC+0];
 687             vdwjidx0D        = 2*vdwtype[jnrD+0];
 688
 689             /**************************
 690              * CALCULATE INTERACTIONS *
 691              **************************/
 692
 693             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 694             {
 695
 696             r00              = _mm256_mul_pd(rsq00,rinv00);
 697             r00              = _mm256_andnot_pd(dummy_mask,r00);
 698
 699             /* Compute parameters for interactions between i and j atoms */
 700             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 701                                             vdwioffsetptr0+vdwjidx0B,
 702                                             vdwioffsetptr0+vdwjidx0C,
 703                                             vdwioffsetptr0+vdwjidx0D,
 704                                             &c6_00,&c12_00);
 705
 706             /* LENNARD-JONES DISPERSION/REPULSION */
 707
 708             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 709             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 710             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 711             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 712             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 713
 714             d                = _mm256_sub_pd(r00,rswitch);
 715             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 716             d2               = _mm256_mul_pd(d,d);
 717             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)))))));
 718
 719             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 720
 721             /* Evaluate switch function */
 722             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 723             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 724             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 725
 726             fscal            = fvdw;
 727
 728             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 729
 730             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 731
 732             /* Calculate temporary vectorial force */
 733             tx               = _mm256_mul_pd(fscal,dx00);
 734             ty               = _mm256_mul_pd(fscal,dy00);
 735             tz               = _mm256_mul_pd(fscal,dz00);
 736
 737             /* Update vectorial force */
 738             fix0             = _mm256_add_pd(fix0,tx);
 739             fiy0             = _mm256_add_pd(fiy0,ty);
 740             fiz0             = _mm256_add_pd(fiz0,tz);
 741
 742             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 743             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 744             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 745             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 746             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 747
 748             }
 749
 750             /* Inner loop uses 57 flops */
 751         }
 752
 753         /* End of innermost loop */
 754
 755         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 756                                                  f+i_coord_offset,fshift+i_shift_offset);
 757
 758         /* Increment number of inner iterations */
 759         inneriter                  += j_index_end - j_index_start;
 760
 761         /* Outer loop uses 6 flops */
 762     }
 763
 764     /* Increment number of outer iterations */
 765     outeriter        += nri;
 766
 767     /* Update outer/inner flops */
 768
 769     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*57);
 770 }