src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRFCut_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_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_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     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  95     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  96     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  97     real             rswitch_scalar,d_scalar;
  98     __m256d          dummy_mask,cutoff_mask;
  99     __m128           tmpmask0,tmpmask1;
 100     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 101     __m256d          one     = _mm256_set1_pd(1.0);
 102     __m256d          two     = _mm256_set1_pd(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm256_set1_pd(fr->ic->epsfac);
 115     charge           = mdatoms->chargeA;
 116     krf              = _mm256_set1_pd(fr->ic->k_rf);
 117     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 118     crf              = _mm256_set1_pd(fr->ic->c_rf);
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 124     rcutoff_scalar   = fr->ic->rcoulomb;
 125     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 126     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 127
 128     rswitch_scalar   = fr->ic->rvdw_switch;
 129     rswitch          = _mm256_set1_pd(rswitch_scalar);
 130     /* Setup switch parameters */
 131     d_scalar         = rcutoff_scalar-rswitch_scalar;
 132     d                = _mm256_set1_pd(d_scalar);
 133     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 134     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 135     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 136     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 137     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 138     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 139
 140     /* Avoid stupid compiler warnings */
 141     jnrA = jnrB = jnrC = jnrD = 0;
 142     j_coord_offsetA = 0;
 143     j_coord_offsetB = 0;
 144     j_coord_offsetC = 0;
 145     j_coord_offsetD = 0;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     for(iidx=0;iidx<4*DIM;iidx++)
 151     {
 152         scratch[iidx] = 0.0;
 153     }
 154
 155     /* Start outer loop over neighborlists */
 156     for(iidx=0; iidx<nri; iidx++)
 157     {
 158         /* Load shift vector for this list */
 159         i_shift_offset   = DIM*shiftidx[iidx];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 171
 172         fix0             = _mm256_setzero_pd();
 173         fiy0             = _mm256_setzero_pd();
 174         fiz0             = _mm256_setzero_pd();
 175
 176         /* Load parameters for i particles */
 177         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 178         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 179
 180         /* Reset potential sums */
 181         velecsum         = _mm256_setzero_pd();
 182         vvdwsum          = _mm256_setzero_pd();
 183
 184         /* Start inner kernel loop */
 185         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 186         {
 187
 188             /* Get j neighbor index, and coordinate index */
 189             jnrA             = jjnr[jidx];
 190             jnrB             = jjnr[jidx+1];
 191             jnrC             = jjnr[jidx+2];
 192             jnrD             = jjnr[jidx+3];
 193             j_coord_offsetA  = DIM*jnrA;
 194             j_coord_offsetB  = DIM*jnrB;
 195             j_coord_offsetC  = DIM*jnrC;
 196             j_coord_offsetD  = DIM*jnrD;
 197
 198             /* load j atom coordinates */
 199             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 200                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 201                                                  &jx0,&jy0,&jz0);
 202
 203             /* Calculate displacement vector */
 204             dx00             = _mm256_sub_pd(ix0,jx0);
 205             dy00             = _mm256_sub_pd(iy0,jy0);
 206             dz00             = _mm256_sub_pd(iz0,jz0);
 207
 208             /* Calculate squared distance and things based on it */
 209             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 210
 211             rinv00           = avx256_invsqrt_d(rsq00);
 212
 213             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 214
 215             /* Load parameters for j particles */
 216             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 217                                                                  charge+jnrC+0,charge+jnrD+0);
 218             vdwjidx0A        = 2*vdwtype[jnrA+0];
 219             vdwjidx0B        = 2*vdwtype[jnrB+0];
 220             vdwjidx0C        = 2*vdwtype[jnrC+0];
 221             vdwjidx0D        = 2*vdwtype[jnrD+0];
 222
 223             /**************************
 224              * CALCULATE INTERACTIONS *
 225              **************************/
 226
 227             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 228             {
 229
 230             r00              = _mm256_mul_pd(rsq00,rinv00);
 231
 232             /* Compute parameters for interactions between i and j atoms */
 233             qq00             = _mm256_mul_pd(iq0,jq0);
 234             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 235                                             vdwioffsetptr0+vdwjidx0B,
 236                                             vdwioffsetptr0+vdwjidx0C,
 237                                             vdwioffsetptr0+vdwjidx0D,
 238                                             &c6_00,&c12_00);
 239
 240             /* REACTION-FIELD ELECTROSTATICS */
 241             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 242             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 243
 244             /* LENNARD-JONES DISPERSION/REPULSION */
 245
 246             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 247             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 248             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 249             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 250             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 251
 252             d                = _mm256_sub_pd(r00,rswitch);
 253             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 254             d2               = _mm256_mul_pd(d,d);
 255             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)))))));
 256
 257             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 258
 259             /* Evaluate switch function */
 260             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 261             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 262             vvdw             = _mm256_mul_pd(vvdw,sw);
 263             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 264
 265             /* Update potential sum for this i atom from the interaction with this j atom. */
 266             velec            = _mm256_and_pd(velec,cutoff_mask);
 267             velecsum         = _mm256_add_pd(velecsum,velec);
 268             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 269             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 270
 271             fscal            = _mm256_add_pd(felec,fvdw);
 272
 273             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 274
 275             /* Calculate temporary vectorial force */
 276             tx               = _mm256_mul_pd(fscal,dx00);
 277             ty               = _mm256_mul_pd(fscal,dy00);
 278             tz               = _mm256_mul_pd(fscal,dz00);
 279
 280             /* Update vectorial force */
 281             fix0             = _mm256_add_pd(fix0,tx);
 282             fiy0             = _mm256_add_pd(fiy0,ty);
 283             fiz0             = _mm256_add_pd(fiz0,tz);
 284
 285             fjptrA             = f+j_coord_offsetA;
 286             fjptrB             = f+j_coord_offsetB;
 287             fjptrC             = f+j_coord_offsetC;
 288             fjptrD             = f+j_coord_offsetD;
 289             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 290
 291             }
 292
 293             /* Inner loop uses 70 flops */
 294         }
 295
 296         if(jidx<j_index_end)
 297         {
 298
 299             /* Get j neighbor index, and coordinate index */
 300             jnrlistA         = jjnr[jidx];
 301             jnrlistB         = jjnr[jidx+1];
 302             jnrlistC         = jjnr[jidx+2];
 303             jnrlistD         = jjnr[jidx+3];
 304             /* Sign of each element will be negative for non-real atoms.
 305              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 306              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 307              */
 308             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 309
 310             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 311             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 312             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 313
 314             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 315             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 316             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 317             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 318             j_coord_offsetA  = DIM*jnrA;
 319             j_coord_offsetB  = DIM*jnrB;
 320             j_coord_offsetC  = DIM*jnrC;
 321             j_coord_offsetD  = DIM*jnrD;
 322
 323             /* load j atom coordinates */
 324             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 325                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 326                                                  &jx0,&jy0,&jz0);
 327
 328             /* Calculate displacement vector */
 329             dx00             = _mm256_sub_pd(ix0,jx0);
 330             dy00             = _mm256_sub_pd(iy0,jy0);
 331             dz00             = _mm256_sub_pd(iz0,jz0);
 332
 333             /* Calculate squared distance and things based on it */
 334             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 335
 336             rinv00           = avx256_invsqrt_d(rsq00);
 337
 338             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 339
 340             /* Load parameters for j particles */
 341             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 342                                                                  charge+jnrC+0,charge+jnrD+0);
 343             vdwjidx0A        = 2*vdwtype[jnrA+0];
 344             vdwjidx0B        = 2*vdwtype[jnrB+0];
 345             vdwjidx0C        = 2*vdwtype[jnrC+0];
 346             vdwjidx0D        = 2*vdwtype[jnrD+0];
 347
 348             /**************************
 349              * CALCULATE INTERACTIONS *
 350              **************************/
 351
 352             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 353             {
 354
 355             r00              = _mm256_mul_pd(rsq00,rinv00);
 356             r00              = _mm256_andnot_pd(dummy_mask,r00);
 357
 358             /* Compute parameters for interactions between i and j atoms */
 359             qq00             = _mm256_mul_pd(iq0,jq0);
 360             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 361                                             vdwioffsetptr0+vdwjidx0B,
 362                                             vdwioffsetptr0+vdwjidx0C,
 363                                             vdwioffsetptr0+vdwjidx0D,
 364                                             &c6_00,&c12_00);
 365
 366             /* REACTION-FIELD ELECTROSTATICS */
 367             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 368             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 369
 370             /* LENNARD-JONES DISPERSION/REPULSION */
 371
 372             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 373             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 374             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 375             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 376             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 377
 378             d                = _mm256_sub_pd(r00,rswitch);
 379             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 380             d2               = _mm256_mul_pd(d,d);
 381             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)))))));
 382
 383             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 384
 385             /* Evaluate switch function */
 386             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 387             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 388             vvdw             = _mm256_mul_pd(vvdw,sw);
 389             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 390
 391             /* Update potential sum for this i atom from the interaction with this j atom. */
 392             velec            = _mm256_and_pd(velec,cutoff_mask);
 393             velec            = _mm256_andnot_pd(dummy_mask,velec);
 394             velecsum         = _mm256_add_pd(velecsum,velec);
 395             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 396             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 397             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 398
 399             fscal            = _mm256_add_pd(felec,fvdw);
 400
 401             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 402
 403             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 404
 405             /* Calculate temporary vectorial force */
 406             tx               = _mm256_mul_pd(fscal,dx00);
 407             ty               = _mm256_mul_pd(fscal,dy00);
 408             tz               = _mm256_mul_pd(fscal,dz00);
 409
 410             /* Update vectorial force */
 411             fix0             = _mm256_add_pd(fix0,tx);
 412             fiy0             = _mm256_add_pd(fiy0,ty);
 413             fiz0             = _mm256_add_pd(fiz0,tz);
 414
 415             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 416             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 417             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 418             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 419             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 420
 421             }
 422
 423             /* Inner loop uses 71 flops */
 424         }
 425
 426         /* End of innermost loop */
 427
 428         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 429                                                  f+i_coord_offset,fshift+i_shift_offset);
 430
 431         ggid                        = gid[iidx];
 432         /* Update potential energies */
 433         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 434         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 435
 436         /* Increment number of inner iterations */
 437         inneriter                  += j_index_end - j_index_start;
 438
 439         /* Outer loop uses 9 flops */
 440     }
 441
 442     /* Increment number of outer iterations */
 443     outeriter        += nri;
 444
 445     /* Update outer/inner flops */
 446
 447     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 448 }
 449 /*
 450  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 451  * Electrostatics interaction: ReactionField
 452  * VdW interaction:            LennardJones
 453  * Geometry:                   Particle-Particle
 454  * Calculate force/pot:        Force
 455  */
 456 void
 457 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 458                     (t_nblist                    * gmx_restrict       nlist,
 459                      rvec                        * gmx_restrict          xx,
 460                      rvec                        * gmx_restrict          ff,
 461                      struct t_forcerec           * gmx_restrict          fr,
 462                      t_mdatoms                   * gmx_restrict     mdatoms,
 463                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 464                      t_nrnb                      * gmx_restrict        nrnb)
 465 {
 466     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 467      * just 0 for non-waters.
 468      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 469      * jnr indices corresponding to data put in the four positions in the SIMD register.
 470      */
 471     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 472     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 473     int              jnrA,jnrB,jnrC,jnrD;
 474     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 475     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 476     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 477     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 478     real             rcutoff_scalar;
 479     real             *shiftvec,*fshift,*x,*f;
 480     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 481     real             scratch[4*DIM];
 482     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 483     real *           vdwioffsetptr0;
 484     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 485     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 486     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 487     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 488     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 489     real             *charge;
 490     int              nvdwtype;
 491     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 492     int              *vdwtype;
 493     real             *vdwparam;
 494     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 495     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 496     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 497     real             rswitch_scalar,d_scalar;
 498     __m256d          dummy_mask,cutoff_mask;
 499     __m128           tmpmask0,tmpmask1;
 500     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 501     __m256d          one     = _mm256_set1_pd(1.0);
 502     __m256d          two     = _mm256_set1_pd(2.0);
 503     x                = xx[0];
 504     f                = ff[0];
 505
 506     nri              = nlist->nri;
 507     iinr             = nlist->iinr;
 508     jindex           = nlist->jindex;
 509     jjnr             = nlist->jjnr;
 510     shiftidx         = nlist->shift;
 511     gid              = nlist->gid;
 512     shiftvec         = fr->shift_vec[0];
 513     fshift           = fr->fshift[0];
 514     facel            = _mm256_set1_pd(fr->ic->epsfac);
 515     charge           = mdatoms->chargeA;
 516     krf              = _mm256_set1_pd(fr->ic->k_rf);
 517     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 518     crf              = _mm256_set1_pd(fr->ic->c_rf);
 519     nvdwtype         = fr->ntype;
 520     vdwparam         = fr->nbfp;
 521     vdwtype          = mdatoms->typeA;
 522
 523     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 524     rcutoff_scalar   = fr->ic->rcoulomb;
 525     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 526     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 527
 528     rswitch_scalar   = fr->ic->rvdw_switch;
 529     rswitch          = _mm256_set1_pd(rswitch_scalar);
 530     /* Setup switch parameters */
 531     d_scalar         = rcutoff_scalar-rswitch_scalar;
 532     d                = _mm256_set1_pd(d_scalar);
 533     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 534     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 535     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 536     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 537     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 538     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 539
 540     /* Avoid stupid compiler warnings */
 541     jnrA = jnrB = jnrC = jnrD = 0;
 542     j_coord_offsetA = 0;
 543     j_coord_offsetB = 0;
 544     j_coord_offsetC = 0;
 545     j_coord_offsetD = 0;
 546
 547     outeriter        = 0;
 548     inneriter        = 0;
 549
 550     for(iidx=0;iidx<4*DIM;iidx++)
 551     {
 552         scratch[iidx] = 0.0;
 553     }
 554
 555     /* Start outer loop over neighborlists */
 556     for(iidx=0; iidx<nri; iidx++)
 557     {
 558         /* Load shift vector for this list */
 559         i_shift_offset   = DIM*shiftidx[iidx];
 560
 561         /* Load limits for loop over neighbors */
 562         j_index_start    = jindex[iidx];
 563         j_index_end      = jindex[iidx+1];
 564
 565         /* Get outer coordinate index */
 566         inr              = iinr[iidx];
 567         i_coord_offset   = DIM*inr;
 568
 569         /* Load i particle coords and add shift vector */
 570         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 571
 572         fix0             = _mm256_setzero_pd();
 573         fiy0             = _mm256_setzero_pd();
 574         fiz0             = _mm256_setzero_pd();
 575
 576         /* Load parameters for i particles */
 577         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 578         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 579
 580         /* Start inner kernel loop */
 581         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 582         {
 583
 584             /* Get j neighbor index, and coordinate index */
 585             jnrA             = jjnr[jidx];
 586             jnrB             = jjnr[jidx+1];
 587             jnrC             = jjnr[jidx+2];
 588             jnrD             = jjnr[jidx+3];
 589             j_coord_offsetA  = DIM*jnrA;
 590             j_coord_offsetB  = DIM*jnrB;
 591             j_coord_offsetC  = DIM*jnrC;
 592             j_coord_offsetD  = DIM*jnrD;
 593
 594             /* load j atom coordinates */
 595             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 596                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 597                                                  &jx0,&jy0,&jz0);
 598
 599             /* Calculate displacement vector */
 600             dx00             = _mm256_sub_pd(ix0,jx0);
 601             dy00             = _mm256_sub_pd(iy0,jy0);
 602             dz00             = _mm256_sub_pd(iz0,jz0);
 603
 604             /* Calculate squared distance and things based on it */
 605             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 606
 607             rinv00           = avx256_invsqrt_d(rsq00);
 608
 609             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 610
 611             /* Load parameters for j particles */
 612             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 613                                                                  charge+jnrC+0,charge+jnrD+0);
 614             vdwjidx0A        = 2*vdwtype[jnrA+0];
 615             vdwjidx0B        = 2*vdwtype[jnrB+0];
 616             vdwjidx0C        = 2*vdwtype[jnrC+0];
 617             vdwjidx0D        = 2*vdwtype[jnrD+0];
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 624             {
 625
 626             r00              = _mm256_mul_pd(rsq00,rinv00);
 627
 628             /* Compute parameters for interactions between i and j atoms */
 629             qq00             = _mm256_mul_pd(iq0,jq0);
 630             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 631                                             vdwioffsetptr0+vdwjidx0B,
 632                                             vdwioffsetptr0+vdwjidx0C,
 633                                             vdwioffsetptr0+vdwjidx0D,
 634                                             &c6_00,&c12_00);
 635
 636             /* REACTION-FIELD ELECTROSTATICS */
 637             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 638
 639             /* LENNARD-JONES DISPERSION/REPULSION */
 640
 641             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 642             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 643             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 644             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 645             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 646
 647             d                = _mm256_sub_pd(r00,rswitch);
 648             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 649             d2               = _mm256_mul_pd(d,d);
 650             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)))))));
 651
 652             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 653
 654             /* Evaluate switch function */
 655             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 656             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 657             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 658
 659             fscal            = _mm256_add_pd(felec,fvdw);
 660
 661             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 662
 663             /* Calculate temporary vectorial force */
 664             tx               = _mm256_mul_pd(fscal,dx00);
 665             ty               = _mm256_mul_pd(fscal,dy00);
 666             tz               = _mm256_mul_pd(fscal,dz00);
 667
 668             /* Update vectorial force */
 669             fix0             = _mm256_add_pd(fix0,tx);
 670             fiy0             = _mm256_add_pd(fiy0,ty);
 671             fiz0             = _mm256_add_pd(fiz0,tz);
 672
 673             fjptrA             = f+j_coord_offsetA;
 674             fjptrB             = f+j_coord_offsetB;
 675             fjptrC             = f+j_coord_offsetC;
 676             fjptrD             = f+j_coord_offsetD;
 677             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 678
 679             }
 680
 681             /* Inner loop uses 61 flops */
 682         }
 683
 684         if(jidx<j_index_end)
 685         {
 686
 687             /* Get j neighbor index, and coordinate index */
 688             jnrlistA         = jjnr[jidx];
 689             jnrlistB         = jjnr[jidx+1];
 690             jnrlistC         = jjnr[jidx+2];
 691             jnrlistD         = jjnr[jidx+3];
 692             /* Sign of each element will be negative for non-real atoms.
 693              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 694              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 695              */
 696             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 697
 698             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 699             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 700             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 701
 702             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 703             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 704             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 705             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 706             j_coord_offsetA  = DIM*jnrA;
 707             j_coord_offsetB  = DIM*jnrB;
 708             j_coord_offsetC  = DIM*jnrC;
 709             j_coord_offsetD  = DIM*jnrD;
 710
 711             /* load j atom coordinates */
 712             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 713                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 714                                                  &jx0,&jy0,&jz0);
 715
 716             /* Calculate displacement vector */
 717             dx00             = _mm256_sub_pd(ix0,jx0);
 718             dy00             = _mm256_sub_pd(iy0,jy0);
 719             dz00             = _mm256_sub_pd(iz0,jz0);
 720
 721             /* Calculate squared distance and things based on it */
 722             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 723
 724             rinv00           = avx256_invsqrt_d(rsq00);
 725
 726             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 727
 728             /* Load parameters for j particles */
 729             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 730                                                                  charge+jnrC+0,charge+jnrD+0);
 731             vdwjidx0A        = 2*vdwtype[jnrA+0];
 732             vdwjidx0B        = 2*vdwtype[jnrB+0];
 733             vdwjidx0C        = 2*vdwtype[jnrC+0];
 734             vdwjidx0D        = 2*vdwtype[jnrD+0];
 735
 736             /**************************
 737              * CALCULATE INTERACTIONS *
 738              **************************/
 739
 740             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 741             {
 742
 743             r00              = _mm256_mul_pd(rsq00,rinv00);
 744             r00              = _mm256_andnot_pd(dummy_mask,r00);
 745
 746             /* Compute parameters for interactions between i and j atoms */
 747             qq00             = _mm256_mul_pd(iq0,jq0);
 748             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 749                                             vdwioffsetptr0+vdwjidx0B,
 750                                             vdwioffsetptr0+vdwjidx0C,
 751                                             vdwioffsetptr0+vdwjidx0D,
 752                                             &c6_00,&c12_00);
 753
 754             /* REACTION-FIELD ELECTROSTATICS */
 755             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 756
 757             /* LENNARD-JONES DISPERSION/REPULSION */
 758
 759             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 760             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 761             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 762             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 763             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 764
 765             d                = _mm256_sub_pd(r00,rswitch);
 766             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 767             d2               = _mm256_mul_pd(d,d);
 768             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)))))));
 769
 770             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 771
 772             /* Evaluate switch function */
 773             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 774             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 775             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 776
 777             fscal            = _mm256_add_pd(felec,fvdw);
 778
 779             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 780
 781             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 782
 783             /* Calculate temporary vectorial force */
 784             tx               = _mm256_mul_pd(fscal,dx00);
 785             ty               = _mm256_mul_pd(fscal,dy00);
 786             tz               = _mm256_mul_pd(fscal,dz00);
 787
 788             /* Update vectorial force */
 789             fix0             = _mm256_add_pd(fix0,tx);
 790             fiy0             = _mm256_add_pd(fiy0,ty);
 791             fiz0             = _mm256_add_pd(fiz0,tz);
 792
 793             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 794             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 795             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 796             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 797             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 798
 799             }
 800
 801             /* Inner loop uses 62 flops */
 802         }
 803
 804         /* End of innermost loop */
 805
 806         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 807                                                  f+i_coord_offset,fshift+i_shift_offset);
 808
 809         /* Increment number of inner iterations */
 810         inneriter                  += j_index_end - j_index_start;
 811
 812         /* Outer loop uses 7 flops */
 813     }
 814
 815     /* Increment number of outer iterations */
 816     outeriter        += nri;
 817
 818     /* Update outer/inner flops */
 819
 820     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);
 821 }