src/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwLJSw_GeomP1P1_avx_256_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_256_double.h"
  34 #include "kernelutil_x86_avx_256_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: None
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  63     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  64     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  65     real             rcutoff_scalar;
  66     real             *shiftvec,*fshift,*x,*f;
  67     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  68     real             scratch[4*DIM];
  69     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  70     real *           vdwioffsetptr0;
  71     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     int              nvdwtype;
  76     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  77     int              *vdwtype;
  78     real             *vdwparam;
  79     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  80     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  81     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  82     real             rswitch_scalar,d_scalar;
  83     __m256d          dummy_mask,cutoff_mask;
  84     __m128           tmpmask0,tmpmask1;
  85     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  86     __m256d          one     = _mm256_set1_pd(1.0);
  87     __m256d          two     = _mm256_set1_pd(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     nvdwtype         = fr->ntype;
 100     vdwparam         = fr->nbfp;
 101     vdwtype          = mdatoms->typeA;
 102
 103     rcutoff_scalar   = fr->rvdw;
 104     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 105     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 106
 107     rswitch_scalar   = fr->rvdw_switch;
 108     rswitch          = _mm256_set1_pd(rswitch_scalar);
 109     /* Setup switch parameters */
 110     d_scalar         = rcutoff_scalar-rswitch_scalar;
 111     d                = _mm256_set1_pd(d_scalar);
 112     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 113     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 114     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 115     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 116     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 117     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 118
 119     /* Avoid stupid compiler warnings */
 120     jnrA = jnrB = jnrC = jnrD = 0;
 121     j_coord_offsetA = 0;
 122     j_coord_offsetB = 0;
 123     j_coord_offsetC = 0;
 124     j_coord_offsetD = 0;
 125
 126     outeriter        = 0;
 127     inneriter        = 0;
 128
 129     for(iidx=0;iidx<4*DIM;iidx++)
 130     {
 131         scratch[iidx] = 0.0;
 132     }
 133
 134     /* Start outer loop over neighborlists */
 135     for(iidx=0; iidx<nri; iidx++)
 136     {
 137         /* Load shift vector for this list */
 138         i_shift_offset   = DIM*shiftidx[iidx];
 139
 140         /* Load limits for loop over neighbors */
 141         j_index_start    = jindex[iidx];
 142         j_index_end      = jindex[iidx+1];
 143
 144         /* Get outer coordinate index */
 145         inr              = iinr[iidx];
 146         i_coord_offset   = DIM*inr;
 147
 148         /* Load i particle coords and add shift vector */
 149         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 150
 151         fix0             = _mm256_setzero_pd();
 152         fiy0             = _mm256_setzero_pd();
 153         fiz0             = _mm256_setzero_pd();
 154
 155         /* Load parameters for i particles */
 156         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 157
 158         /* Reset potential sums */
 159         vvdwsum          = _mm256_setzero_pd();
 160
 161         /* Start inner kernel loop */
 162         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 163         {
 164
 165             /* Get j neighbor index, and coordinate index */
 166             jnrA             = jjnr[jidx];
 167             jnrB             = jjnr[jidx+1];
 168             jnrC             = jjnr[jidx+2];
 169             jnrD             = jjnr[jidx+3];
 170             j_coord_offsetA  = DIM*jnrA;
 171             j_coord_offsetB  = DIM*jnrB;
 172             j_coord_offsetC  = DIM*jnrC;
 173             j_coord_offsetD  = DIM*jnrD;
 174
 175             /* load j atom coordinates */
 176             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 177                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 178                                                  &jx0,&jy0,&jz0);
 179
 180             /* Calculate displacement vector */
 181             dx00             = _mm256_sub_pd(ix0,jx0);
 182             dy00             = _mm256_sub_pd(iy0,jy0);
 183             dz00             = _mm256_sub_pd(iz0,jz0);
 184
 185             /* Calculate squared distance and things based on it */
 186             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 187
 188             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 189
 190             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 191
 192             /* Load parameters for j particles */
 193             vdwjidx0A        = 2*vdwtype[jnrA+0];
 194             vdwjidx0B        = 2*vdwtype[jnrB+0];
 195             vdwjidx0C        = 2*vdwtype[jnrC+0];
 196             vdwjidx0D        = 2*vdwtype[jnrD+0];
 197
 198             /**************************
 199              * CALCULATE INTERACTIONS *
 200              **************************/
 201
 202             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 203             {
 204
 205             r00              = _mm256_mul_pd(rsq00,rinv00);
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 209                                             vdwioffsetptr0+vdwjidx0B,
 210                                             vdwioffsetptr0+vdwjidx0C,
 211                                             vdwioffsetptr0+vdwjidx0D,
 212                                             &c6_00,&c12_00);
 213
 214             /* LENNARD-JONES DISPERSION/REPULSION */
 215
 216             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 217             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 218             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 219             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 220             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 221
 222             d                = _mm256_sub_pd(r00,rswitch);
 223             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 224             d2               = _mm256_mul_pd(d,d);
 225             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)))))));
 226
 227             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 228
 229             /* Evaluate switch function */
 230             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 231             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 232             vvdw             = _mm256_mul_pd(vvdw,sw);
 233             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 234
 235             /* Update potential sum for this i atom from the interaction with this j atom. */
 236             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 237             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 238
 239             fscal            = fvdw;
 240
 241             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 242
 243             /* Calculate temporary vectorial force */
 244             tx               = _mm256_mul_pd(fscal,dx00);
 245             ty               = _mm256_mul_pd(fscal,dy00);
 246             tz               = _mm256_mul_pd(fscal,dz00);
 247
 248             /* Update vectorial force */
 249             fix0             = _mm256_add_pd(fix0,tx);
 250             fiy0             = _mm256_add_pd(fiy0,ty);
 251             fiz0             = _mm256_add_pd(fiz0,tz);
 252
 253             fjptrA             = f+j_coord_offsetA;
 254             fjptrB             = f+j_coord_offsetB;
 255             fjptrC             = f+j_coord_offsetC;
 256             fjptrD             = f+j_coord_offsetD;
 257             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 258
 259             }
 260
 261             /* Inner loop uses 59 flops */
 262         }
 263
 264         if(jidx<j_index_end)
 265         {
 266
 267             /* Get j neighbor index, and coordinate index */
 268             jnrlistA         = jjnr[jidx];
 269             jnrlistB         = jjnr[jidx+1];
 270             jnrlistC         = jjnr[jidx+2];
 271             jnrlistD         = jjnr[jidx+3];
 272             /* Sign of each element will be negative for non-real atoms.
 273              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 274              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 275              */
 276             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 277
 278             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 279             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 280             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 281
 282             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 283             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 284             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 285             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 286             j_coord_offsetA  = DIM*jnrA;
 287             j_coord_offsetB  = DIM*jnrB;
 288             j_coord_offsetC  = DIM*jnrC;
 289             j_coord_offsetD  = DIM*jnrD;
 290
 291             /* load j atom coordinates */
 292             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 293                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 294                                                  &jx0,&jy0,&jz0);
 295
 296             /* Calculate displacement vector */
 297             dx00             = _mm256_sub_pd(ix0,jx0);
 298             dy00             = _mm256_sub_pd(iy0,jy0);
 299             dz00             = _mm256_sub_pd(iz0,jz0);
 300
 301             /* Calculate squared distance and things based on it */
 302             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 303
 304             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 305
 306             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 307
 308             /* Load parameters for j particles */
 309             vdwjidx0A        = 2*vdwtype[jnrA+0];
 310             vdwjidx0B        = 2*vdwtype[jnrB+0];
 311             vdwjidx0C        = 2*vdwtype[jnrC+0];
 312             vdwjidx0D        = 2*vdwtype[jnrD+0];
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 319             {
 320
 321             r00              = _mm256_mul_pd(rsq00,rinv00);
 322             r00              = _mm256_andnot_pd(dummy_mask,r00);
 323
 324             /* Compute parameters for interactions between i and j atoms */
 325             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 326                                             vdwioffsetptr0+vdwjidx0B,
 327                                             vdwioffsetptr0+vdwjidx0C,
 328                                             vdwioffsetptr0+vdwjidx0D,
 329                                             &c6_00,&c12_00);
 330
 331             /* LENNARD-JONES DISPERSION/REPULSION */
 332
 333             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 334             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 335             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 336             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 337             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 338
 339             d                = _mm256_sub_pd(r00,rswitch);
 340             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 341             d2               = _mm256_mul_pd(d,d);
 342             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)))))));
 343
 344             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 345
 346             /* Evaluate switch function */
 347             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 348             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 349             vvdw             = _mm256_mul_pd(vvdw,sw);
 350             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 354             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 355             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 356
 357             fscal            = fvdw;
 358
 359             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 360
 361             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 362
 363             /* Calculate temporary vectorial force */
 364             tx               = _mm256_mul_pd(fscal,dx00);
 365             ty               = _mm256_mul_pd(fscal,dy00);
 366             tz               = _mm256_mul_pd(fscal,dz00);
 367
 368             /* Update vectorial force */
 369             fix0             = _mm256_add_pd(fix0,tx);
 370             fiy0             = _mm256_add_pd(fiy0,ty);
 371             fiz0             = _mm256_add_pd(fiz0,tz);
 372
 373             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 374             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 375             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 376             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 377             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 378
 379             }
 380
 381             /* Inner loop uses 60 flops */
 382         }
 383
 384         /* End of innermost loop */
 385
 386         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 387                                                  f+i_coord_offset,fshift+i_shift_offset);
 388
 389         ggid                        = gid[iidx];
 390         /* Update potential energies */
 391         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 392
 393         /* Increment number of inner iterations */
 394         inneriter                  += j_index_end - j_index_start;
 395
 396         /* Outer loop uses 7 flops */
 397     }
 398
 399     /* Increment number of outer iterations */
 400     outeriter        += nri;
 401
 402     /* Update outer/inner flops */
 403
 404     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 405 }
 406 /*
 407  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_double
 408  * Electrostatics interaction: None
 409  * VdW interaction:            LennardJones
 410  * Geometry:                   Particle-Particle
 411  * Calculate force/pot:        Force
 412  */
 413 void
 414 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_double
 415                     (t_nblist * gmx_restrict                nlist,
 416                      rvec * gmx_restrict                    xx,
 417                      rvec * gmx_restrict                    ff,
 418                      t_forcerec * gmx_restrict              fr,
 419                      t_mdatoms * gmx_restrict               mdatoms,
 420                      nb_kernel_data_t * gmx_restrict        kernel_data,
 421                      t_nrnb * gmx_restrict                  nrnb)
 422 {
 423     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 424      * just 0 for non-waters.
 425      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 426      * jnr indices corresponding to data put in the four positions in the SIMD register.
 427      */
 428     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 429     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 430     int              jnrA,jnrB,jnrC,jnrD;
 431     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 432     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 433     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 434     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 435     real             rcutoff_scalar;
 436     real             *shiftvec,*fshift,*x,*f;
 437     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 438     real             scratch[4*DIM];
 439     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 440     real *           vdwioffsetptr0;
 441     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 442     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 443     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 444     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 445     int              nvdwtype;
 446     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 447     int              *vdwtype;
 448     real             *vdwparam;
 449     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 450     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 451     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 452     real             rswitch_scalar,d_scalar;
 453     __m256d          dummy_mask,cutoff_mask;
 454     __m128           tmpmask0,tmpmask1;
 455     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 456     __m256d          one     = _mm256_set1_pd(1.0);
 457     __m256d          two     = _mm256_set1_pd(2.0);
 458     x                = xx[0];
 459     f                = ff[0];
 460
 461     nri              = nlist->nri;
 462     iinr             = nlist->iinr;
 463     jindex           = nlist->jindex;
 464     jjnr             = nlist->jjnr;
 465     shiftidx         = nlist->shift;
 466     gid              = nlist->gid;
 467     shiftvec         = fr->shift_vec[0];
 468     fshift           = fr->fshift[0];
 469     nvdwtype         = fr->ntype;
 470     vdwparam         = fr->nbfp;
 471     vdwtype          = mdatoms->typeA;
 472
 473     rcutoff_scalar   = fr->rvdw;
 474     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 475     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 476
 477     rswitch_scalar   = fr->rvdw_switch;
 478     rswitch          = _mm256_set1_pd(rswitch_scalar);
 479     /* Setup switch parameters */
 480     d_scalar         = rcutoff_scalar-rswitch_scalar;
 481     d                = _mm256_set1_pd(d_scalar);
 482     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 483     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 484     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 485     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 486     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 487     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 488
 489     /* Avoid stupid compiler warnings */
 490     jnrA = jnrB = jnrC = jnrD = 0;
 491     j_coord_offsetA = 0;
 492     j_coord_offsetB = 0;
 493     j_coord_offsetC = 0;
 494     j_coord_offsetD = 0;
 495
 496     outeriter        = 0;
 497     inneriter        = 0;
 498
 499     for(iidx=0;iidx<4*DIM;iidx++)
 500     {
 501         scratch[iidx] = 0.0;
 502     }
 503
 504     /* Start outer loop over neighborlists */
 505     for(iidx=0; iidx<nri; iidx++)
 506     {
 507         /* Load shift vector for this list */
 508         i_shift_offset   = DIM*shiftidx[iidx];
 509
 510         /* Load limits for loop over neighbors */
 511         j_index_start    = jindex[iidx];
 512         j_index_end      = jindex[iidx+1];
 513
 514         /* Get outer coordinate index */
 515         inr              = iinr[iidx];
 516         i_coord_offset   = DIM*inr;
 517
 518         /* Load i particle coords and add shift vector */
 519         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 520
 521         fix0             = _mm256_setzero_pd();
 522         fiy0             = _mm256_setzero_pd();
 523         fiz0             = _mm256_setzero_pd();
 524
 525         /* Load parameters for i particles */
 526         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 527
 528         /* Start inner kernel loop */
 529         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 530         {
 531
 532             /* Get j neighbor index, and coordinate index */
 533             jnrA             = jjnr[jidx];
 534             jnrB             = jjnr[jidx+1];
 535             jnrC             = jjnr[jidx+2];
 536             jnrD             = jjnr[jidx+3];
 537             j_coord_offsetA  = DIM*jnrA;
 538             j_coord_offsetB  = DIM*jnrB;
 539             j_coord_offsetC  = DIM*jnrC;
 540             j_coord_offsetD  = DIM*jnrD;
 541
 542             /* load j atom coordinates */
 543             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 544                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 545                                                  &jx0,&jy0,&jz0);
 546
 547             /* Calculate displacement vector */
 548             dx00             = _mm256_sub_pd(ix0,jx0);
 549             dy00             = _mm256_sub_pd(iy0,jy0);
 550             dz00             = _mm256_sub_pd(iz0,jz0);
 551
 552             /* Calculate squared distance and things based on it */
 553             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 554
 555             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 556
 557             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 558
 559             /* Load parameters for j particles */
 560             vdwjidx0A        = 2*vdwtype[jnrA+0];
 561             vdwjidx0B        = 2*vdwtype[jnrB+0];
 562             vdwjidx0C        = 2*vdwtype[jnrC+0];
 563             vdwjidx0D        = 2*vdwtype[jnrD+0];
 564
 565             /**************************
 566              * CALCULATE INTERACTIONS *
 567              **************************/
 568
 569             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 570             {
 571
 572             r00              = _mm256_mul_pd(rsq00,rinv00);
 573
 574             /* Compute parameters for interactions between i and j atoms */
 575             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 576                                             vdwioffsetptr0+vdwjidx0B,
 577                                             vdwioffsetptr0+vdwjidx0C,
 578                                             vdwioffsetptr0+vdwjidx0D,
 579                                             &c6_00,&c12_00);
 580
 581             /* LENNARD-JONES DISPERSION/REPULSION */
 582
 583             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 584             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 585             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 586             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 587             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 588
 589             d                = _mm256_sub_pd(r00,rswitch);
 590             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 591             d2               = _mm256_mul_pd(d,d);
 592             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)))))));
 593
 594             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 595
 596             /* Evaluate switch function */
 597             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 598             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 599             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 600
 601             fscal            = fvdw;
 602
 603             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 604
 605             /* Calculate temporary vectorial force */
 606             tx               = _mm256_mul_pd(fscal,dx00);
 607             ty               = _mm256_mul_pd(fscal,dy00);
 608             tz               = _mm256_mul_pd(fscal,dz00);
 609
 610             /* Update vectorial force */
 611             fix0             = _mm256_add_pd(fix0,tx);
 612             fiy0             = _mm256_add_pd(fiy0,ty);
 613             fiz0             = _mm256_add_pd(fiz0,tz);
 614
 615             fjptrA             = f+j_coord_offsetA;
 616             fjptrB             = f+j_coord_offsetB;
 617             fjptrC             = f+j_coord_offsetC;
 618             fjptrD             = f+j_coord_offsetD;
 619             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 620
 621             }
 622
 623             /* Inner loop uses 56 flops */
 624         }
 625
 626         if(jidx<j_index_end)
 627         {
 628
 629             /* Get j neighbor index, and coordinate index */
 630             jnrlistA         = jjnr[jidx];
 631             jnrlistB         = jjnr[jidx+1];
 632             jnrlistC         = jjnr[jidx+2];
 633             jnrlistD         = jjnr[jidx+3];
 634             /* Sign of each element will be negative for non-real atoms.
 635              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 636              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 637              */
 638             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 639
 640             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 641             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 642             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 643
 644             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 645             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 646             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 647             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 648             j_coord_offsetA  = DIM*jnrA;
 649             j_coord_offsetB  = DIM*jnrB;
 650             j_coord_offsetC  = DIM*jnrC;
 651             j_coord_offsetD  = DIM*jnrD;
 652
 653             /* load j atom coordinates */
 654             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 655                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 656                                                  &jx0,&jy0,&jz0);
 657
 658             /* Calculate displacement vector */
 659             dx00             = _mm256_sub_pd(ix0,jx0);
 660             dy00             = _mm256_sub_pd(iy0,jy0);
 661             dz00             = _mm256_sub_pd(iz0,jz0);
 662
 663             /* Calculate squared distance and things based on it */
 664             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 665
 666             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 667
 668             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 669
 670             /* Load parameters for j particles */
 671             vdwjidx0A        = 2*vdwtype[jnrA+0];
 672             vdwjidx0B        = 2*vdwtype[jnrB+0];
 673             vdwjidx0C        = 2*vdwtype[jnrC+0];
 674             vdwjidx0D        = 2*vdwtype[jnrD+0];
 675
 676             /**************************
 677              * CALCULATE INTERACTIONS *
 678              **************************/
 679
 680             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 681             {
 682
 683             r00              = _mm256_mul_pd(rsq00,rinv00);
 684             r00              = _mm256_andnot_pd(dummy_mask,r00);
 685
 686             /* Compute parameters for interactions between i and j atoms */
 687             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 688                                             vdwioffsetptr0+vdwjidx0B,
 689                                             vdwioffsetptr0+vdwjidx0C,
 690                                             vdwioffsetptr0+vdwjidx0D,
 691                                             &c6_00,&c12_00);
 692
 693             /* LENNARD-JONES DISPERSION/REPULSION */
 694
 695             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 696             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 697             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 698             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 699             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 700
 701             d                = _mm256_sub_pd(r00,rswitch);
 702             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 703             d2               = _mm256_mul_pd(d,d);
 704             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)))))));
 705
 706             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 707
 708             /* Evaluate switch function */
 709             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 710             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 711             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 712
 713             fscal            = fvdw;
 714
 715             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 716
 717             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 718
 719             /* Calculate temporary vectorial force */
 720             tx               = _mm256_mul_pd(fscal,dx00);
 721             ty               = _mm256_mul_pd(fscal,dy00);
 722             tz               = _mm256_mul_pd(fscal,dz00);
 723
 724             /* Update vectorial force */
 725             fix0             = _mm256_add_pd(fix0,tx);
 726             fiy0             = _mm256_add_pd(fiy0,ty);
 727             fiz0             = _mm256_add_pd(fiz0,tz);
 728
 729             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 730             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 731             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 732             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 733             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 734
 735             }
 736
 737             /* Inner loop uses 57 flops */
 738         }
 739
 740         /* End of innermost loop */
 741
 742         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 743                                                  f+i_coord_offset,fshift+i_shift_offset);
 744
 745         /* Increment number of inner iterations */
 746         inneriter                  += j_index_end - j_index_start;
 747
 748         /* Outer loop uses 6 flops */
 749     }
 750
 751     /* Increment number of outer iterations */
 752     outeriter        += nri;
 753
 754     /* Update outer/inner flops */
 755
 756     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*57);
 757 }