src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  91     int              nvdwtype;
  92     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  96     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  97     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  98     real             rswitch_scalar,d_scalar;
  99     __m256d          dummy_mask,cutoff_mask;
 100     __m128           tmpmask0,tmpmask1;
 101     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 102     __m256d          one     = _mm256_set1_pd(1.0);
 103     __m256d          two     = _mm256_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm256_set1_pd(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     krf              = _mm256_set1_pd(fr->ic->k_rf);
 118     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 119     crf              = _mm256_set1_pd(fr->ic->c_rf);
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 125     rcutoff_scalar   = fr->rcoulomb;
 126     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 127     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 128
 129     rswitch_scalar   = fr->rvdw_switch;
 130     rswitch          = _mm256_set1_pd(rswitch_scalar);
 131     /* Setup switch parameters */
 132     d_scalar         = rcutoff_scalar-rswitch_scalar;
 133     d                = _mm256_set1_pd(d_scalar);
 134     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 135     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 136     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 137     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 138     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = jnrC = jnrD = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145     j_coord_offsetC = 0;
 146     j_coord_offsetD = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     for(iidx=0;iidx<4*DIM;iidx++)
 152     {
 153         scratch[iidx] = 0.0;
 154     }
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 172
 173         fix0             = _mm256_setzero_pd();
 174         fiy0             = _mm256_setzero_pd();
 175         fiz0             = _mm256_setzero_pd();
 176
 177         /* Load parameters for i particles */
 178         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 179         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 180
 181         /* Reset potential sums */
 182         velecsum         = _mm256_setzero_pd();
 183         vvdwsum          = _mm256_setzero_pd();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             jnrC             = jjnr[jidx+2];
 193             jnrD             = jjnr[jidx+3];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196             j_coord_offsetC  = DIM*jnrC;
 197             j_coord_offsetD  = DIM*jnrD;
 198
 199             /* load j atom coordinates */
 200             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 201                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 202                                                  &jx0,&jy0,&jz0);
 203
 204             /* Calculate displacement vector */
 205             dx00             = _mm256_sub_pd(ix0,jx0);
 206             dy00             = _mm256_sub_pd(iy0,jy0);
 207             dz00             = _mm256_sub_pd(iz0,jz0);
 208
 209             /* Calculate squared distance and things based on it */
 210             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 211
 212             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 213
 214             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 215
 216             /* Load parameters for j particles */
 217             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 218                                                                  charge+jnrC+0,charge+jnrD+0);
 219             vdwjidx0A        = 2*vdwtype[jnrA+0];
 220             vdwjidx0B        = 2*vdwtype[jnrB+0];
 221             vdwjidx0C        = 2*vdwtype[jnrC+0];
 222             vdwjidx0D        = 2*vdwtype[jnrD+0];
 223
 224             /**************************
 225              * CALCULATE INTERACTIONS *
 226              **************************/
 227
 228             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 229             {
 230
 231             r00              = _mm256_mul_pd(rsq00,rinv00);
 232
 233             /* Compute parameters for interactions between i and j atoms */
 234             qq00             = _mm256_mul_pd(iq0,jq0);
 235             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 236                                             vdwioffsetptr0+vdwjidx0B,
 237                                             vdwioffsetptr0+vdwjidx0C,
 238                                             vdwioffsetptr0+vdwjidx0D,
 239                                             &c6_00,&c12_00);
 240
 241             /* REACTION-FIELD ELECTROSTATICS */
 242             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 243             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 244
 245             /* LENNARD-JONES DISPERSION/REPULSION */
 246
 247             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 248             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 249             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 250             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 251             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 252
 253             d                = _mm256_sub_pd(r00,rswitch);
 254             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 255             d2               = _mm256_mul_pd(d,d);
 256             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)))))));
 257
 258             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 259
 260             /* Evaluate switch function */
 261             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 262             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 263             vvdw             = _mm256_mul_pd(vvdw,sw);
 264             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 265
 266             /* Update potential sum for this i atom from the interaction with this j atom. */
 267             velec            = _mm256_and_pd(velec,cutoff_mask);
 268             velecsum         = _mm256_add_pd(velecsum,velec);
 269             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 270             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 271
 272             fscal            = _mm256_add_pd(felec,fvdw);
 273
 274             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 275
 276             /* Calculate temporary vectorial force */
 277             tx               = _mm256_mul_pd(fscal,dx00);
 278             ty               = _mm256_mul_pd(fscal,dy00);
 279             tz               = _mm256_mul_pd(fscal,dz00);
 280
 281             /* Update vectorial force */
 282             fix0             = _mm256_add_pd(fix0,tx);
 283             fiy0             = _mm256_add_pd(fiy0,ty);
 284             fiz0             = _mm256_add_pd(fiz0,tz);
 285
 286             fjptrA             = f+j_coord_offsetA;
 287             fjptrB             = f+j_coord_offsetB;
 288             fjptrC             = f+j_coord_offsetC;
 289             fjptrD             = f+j_coord_offsetD;
 290             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 291
 292             }
 293
 294             /* Inner loop uses 70 flops */
 295         }
 296
 297         if(jidx<j_index_end)
 298         {
 299
 300             /* Get j neighbor index, and coordinate index */
 301             jnrlistA         = jjnr[jidx];
 302             jnrlistB         = jjnr[jidx+1];
 303             jnrlistC         = jjnr[jidx+2];
 304             jnrlistD         = jjnr[jidx+3];
 305             /* Sign of each element will be negative for non-real atoms.
 306              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 307              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 308              */
 309             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 310
 311             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 312             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 313             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 314
 315             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 316             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 317             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 318             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 319             j_coord_offsetA  = DIM*jnrA;
 320             j_coord_offsetB  = DIM*jnrB;
 321             j_coord_offsetC  = DIM*jnrC;
 322             j_coord_offsetD  = DIM*jnrD;
 323
 324             /* load j atom coordinates */
 325             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 326                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 327                                                  &jx0,&jy0,&jz0);
 328
 329             /* Calculate displacement vector */
 330             dx00             = _mm256_sub_pd(ix0,jx0);
 331             dy00             = _mm256_sub_pd(iy0,jy0);
 332             dz00             = _mm256_sub_pd(iz0,jz0);
 333
 334             /* Calculate squared distance and things based on it */
 335             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 336
 337             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 338
 339             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 340
 341             /* Load parameters for j particles */
 342             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 343                                                                  charge+jnrC+0,charge+jnrD+0);
 344             vdwjidx0A        = 2*vdwtype[jnrA+0];
 345             vdwjidx0B        = 2*vdwtype[jnrB+0];
 346             vdwjidx0C        = 2*vdwtype[jnrC+0];
 347             vdwjidx0D        = 2*vdwtype[jnrD+0];
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 354             {
 355
 356             r00              = _mm256_mul_pd(rsq00,rinv00);
 357             r00              = _mm256_andnot_pd(dummy_mask,r00);
 358
 359             /* Compute parameters for interactions between i and j atoms */
 360             qq00             = _mm256_mul_pd(iq0,jq0);
 361             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 362                                             vdwioffsetptr0+vdwjidx0B,
 363                                             vdwioffsetptr0+vdwjidx0C,
 364                                             vdwioffsetptr0+vdwjidx0D,
 365                                             &c6_00,&c12_00);
 366
 367             /* REACTION-FIELD ELECTROSTATICS */
 368             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 369             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 370
 371             /* LENNARD-JONES DISPERSION/REPULSION */
 372
 373             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 374             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 375             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 376             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 377             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 378
 379             d                = _mm256_sub_pd(r00,rswitch);
 380             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 381             d2               = _mm256_mul_pd(d,d);
 382             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)))))));
 383
 384             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 385
 386             /* Evaluate switch function */
 387             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 388             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 389             vvdw             = _mm256_mul_pd(vvdw,sw);
 390             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 391
 392             /* Update potential sum for this i atom from the interaction with this j atom. */
 393             velec            = _mm256_and_pd(velec,cutoff_mask);
 394             velec            = _mm256_andnot_pd(dummy_mask,velec);
 395             velecsum         = _mm256_add_pd(velecsum,velec);
 396             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 397             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 398             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 399
 400             fscal            = _mm256_add_pd(felec,fvdw);
 401
 402             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 403
 404             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = _mm256_mul_pd(fscal,dx00);
 408             ty               = _mm256_mul_pd(fscal,dy00);
 409             tz               = _mm256_mul_pd(fscal,dz00);
 410
 411             /* Update vectorial force */
 412             fix0             = _mm256_add_pd(fix0,tx);
 413             fiy0             = _mm256_add_pd(fiy0,ty);
 414             fiz0             = _mm256_add_pd(fiz0,tz);
 415
 416             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 417             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 418             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 419             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 420             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 421
 422             }
 423
 424             /* Inner loop uses 71 flops */
 425         }
 426
 427         /* End of innermost loop */
 428
 429         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 430                                                  f+i_coord_offset,fshift+i_shift_offset);
 431
 432         ggid                        = gid[iidx];
 433         /* Update potential energies */
 434         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 435         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 436
 437         /* Increment number of inner iterations */
 438         inneriter                  += j_index_end - j_index_start;
 439
 440         /* Outer loop uses 9 flops */
 441     }
 442
 443     /* Increment number of outer iterations */
 444     outeriter        += nri;
 445
 446     /* Update outer/inner flops */
 447
 448     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 449 }
 450 /*
 451  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 452  * Electrostatics interaction: ReactionField
 453  * VdW interaction:            LennardJones
 454  * Geometry:                   Particle-Particle
 455  * Calculate force/pot:        Force
 456  */
 457 void
 458 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 459                     (t_nblist                    * gmx_restrict       nlist,
 460                      rvec                        * gmx_restrict          xx,
 461                      rvec                        * gmx_restrict          ff,
 462                      t_forcerec                  * gmx_restrict          fr,
 463                      t_mdatoms                   * gmx_restrict     mdatoms,
 464                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 465                      t_nrnb                      * gmx_restrict        nrnb)
 466 {
 467     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 468      * just 0 for non-waters.
 469      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 470      * jnr indices corresponding to data put in the four positions in the SIMD register.
 471      */
 472     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 473     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 474     int              jnrA,jnrB,jnrC,jnrD;
 475     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 476     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 477     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 478     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 479     real             rcutoff_scalar;
 480     real             *shiftvec,*fshift,*x,*f;
 481     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 482     real             scratch[4*DIM];
 483     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 484     real *           vdwioffsetptr0;
 485     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 486     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 487     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 488     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 489     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 490     real             *charge;
 491     int              nvdwtype;
 492     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 493     int              *vdwtype;
 494     real             *vdwparam;
 495     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 496     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 497     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 498     real             rswitch_scalar,d_scalar;
 499     __m256d          dummy_mask,cutoff_mask;
 500     __m128           tmpmask0,tmpmask1;
 501     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 502     __m256d          one     = _mm256_set1_pd(1.0);
 503     __m256d          two     = _mm256_set1_pd(2.0);
 504     x                = xx[0];
 505     f                = ff[0];
 506
 507     nri              = nlist->nri;
 508     iinr             = nlist->iinr;
 509     jindex           = nlist->jindex;
 510     jjnr             = nlist->jjnr;
 511     shiftidx         = nlist->shift;
 512     gid              = nlist->gid;
 513     shiftvec         = fr->shift_vec[0];
 514     fshift           = fr->fshift[0];
 515     facel            = _mm256_set1_pd(fr->epsfac);
 516     charge           = mdatoms->chargeA;
 517     krf              = _mm256_set1_pd(fr->ic->k_rf);
 518     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 519     crf              = _mm256_set1_pd(fr->ic->c_rf);
 520     nvdwtype         = fr->ntype;
 521     vdwparam         = fr->nbfp;
 522     vdwtype          = mdatoms->typeA;
 523
 524     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 525     rcutoff_scalar   = fr->rcoulomb;
 526     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 527     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 528
 529     rswitch_scalar   = fr->rvdw_switch;
 530     rswitch          = _mm256_set1_pd(rswitch_scalar);
 531     /* Setup switch parameters */
 532     d_scalar         = rcutoff_scalar-rswitch_scalar;
 533     d                = _mm256_set1_pd(d_scalar);
 534     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 535     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 536     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 537     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 538     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 539     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 540
 541     /* Avoid stupid compiler warnings */
 542     jnrA = jnrB = jnrC = jnrD = 0;
 543     j_coord_offsetA = 0;
 544     j_coord_offsetB = 0;
 545     j_coord_offsetC = 0;
 546     j_coord_offsetD = 0;
 547
 548     outeriter        = 0;
 549     inneriter        = 0;
 550
 551     for(iidx=0;iidx<4*DIM;iidx++)
 552     {
 553         scratch[iidx] = 0.0;
 554     }
 555
 556     /* Start outer loop over neighborlists */
 557     for(iidx=0; iidx<nri; iidx++)
 558     {
 559         /* Load shift vector for this list */
 560         i_shift_offset   = DIM*shiftidx[iidx];
 561
 562         /* Load limits for loop over neighbors */
 563         j_index_start    = jindex[iidx];
 564         j_index_end      = jindex[iidx+1];
 565
 566         /* Get outer coordinate index */
 567         inr              = iinr[iidx];
 568         i_coord_offset   = DIM*inr;
 569
 570         /* Load i particle coords and add shift vector */
 571         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 572
 573         fix0             = _mm256_setzero_pd();
 574         fiy0             = _mm256_setzero_pd();
 575         fiz0             = _mm256_setzero_pd();
 576
 577         /* Load parameters for i particles */
 578         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 579         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 580
 581         /* Start inner kernel loop */
 582         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 583         {
 584
 585             /* Get j neighbor index, and coordinate index */
 586             jnrA             = jjnr[jidx];
 587             jnrB             = jjnr[jidx+1];
 588             jnrC             = jjnr[jidx+2];
 589             jnrD             = jjnr[jidx+3];
 590             j_coord_offsetA  = DIM*jnrA;
 591             j_coord_offsetB  = DIM*jnrB;
 592             j_coord_offsetC  = DIM*jnrC;
 593             j_coord_offsetD  = DIM*jnrD;
 594
 595             /* load j atom coordinates */
 596             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 597                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 598                                                  &jx0,&jy0,&jz0);
 599
 600             /* Calculate displacement vector */
 601             dx00             = _mm256_sub_pd(ix0,jx0);
 602             dy00             = _mm256_sub_pd(iy0,jy0);
 603             dz00             = _mm256_sub_pd(iz0,jz0);
 604
 605             /* Calculate squared distance and things based on it */
 606             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 607
 608             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 609
 610             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 611
 612             /* Load parameters for j particles */
 613             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 614                                                                  charge+jnrC+0,charge+jnrD+0);
 615             vdwjidx0A        = 2*vdwtype[jnrA+0];
 616             vdwjidx0B        = 2*vdwtype[jnrB+0];
 617             vdwjidx0C        = 2*vdwtype[jnrC+0];
 618             vdwjidx0D        = 2*vdwtype[jnrD+0];
 619
 620             /**************************
 621              * CALCULATE INTERACTIONS *
 622              **************************/
 623
 624             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 625             {
 626
 627             r00              = _mm256_mul_pd(rsq00,rinv00);
 628
 629             /* Compute parameters for interactions between i and j atoms */
 630             qq00             = _mm256_mul_pd(iq0,jq0);
 631             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 632                                             vdwioffsetptr0+vdwjidx0B,
 633                                             vdwioffsetptr0+vdwjidx0C,
 634                                             vdwioffsetptr0+vdwjidx0D,
 635                                             &c6_00,&c12_00);
 636
 637             /* REACTION-FIELD ELECTROSTATICS */
 638             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 639
 640             /* LENNARD-JONES DISPERSION/REPULSION */
 641
 642             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 643             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 644             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 645             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 646             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 647
 648             d                = _mm256_sub_pd(r00,rswitch);
 649             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 650             d2               = _mm256_mul_pd(d,d);
 651             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)))))));
 652
 653             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 654
 655             /* Evaluate switch function */
 656             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 657             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 658             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 659
 660             fscal            = _mm256_add_pd(felec,fvdw);
 661
 662             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 663
 664             /* Calculate temporary vectorial force */
 665             tx               = _mm256_mul_pd(fscal,dx00);
 666             ty               = _mm256_mul_pd(fscal,dy00);
 667             tz               = _mm256_mul_pd(fscal,dz00);
 668
 669             /* Update vectorial force */
 670             fix0             = _mm256_add_pd(fix0,tx);
 671             fiy0             = _mm256_add_pd(fiy0,ty);
 672             fiz0             = _mm256_add_pd(fiz0,tz);
 673
 674             fjptrA             = f+j_coord_offsetA;
 675             fjptrB             = f+j_coord_offsetB;
 676             fjptrC             = f+j_coord_offsetC;
 677             fjptrD             = f+j_coord_offsetD;
 678             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 679
 680             }
 681
 682             /* Inner loop uses 61 flops */
 683         }
 684
 685         if(jidx<j_index_end)
 686         {
 687
 688             /* Get j neighbor index, and coordinate index */
 689             jnrlistA         = jjnr[jidx];
 690             jnrlistB         = jjnr[jidx+1];
 691             jnrlistC         = jjnr[jidx+2];
 692             jnrlistD         = jjnr[jidx+3];
 693             /* Sign of each element will be negative for non-real atoms.
 694              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 695              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 696              */
 697             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 698
 699             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 700             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 701             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 702
 703             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 704             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 705             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 706             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 707             j_coord_offsetA  = DIM*jnrA;
 708             j_coord_offsetB  = DIM*jnrB;
 709             j_coord_offsetC  = DIM*jnrC;
 710             j_coord_offsetD  = DIM*jnrD;
 711
 712             /* load j atom coordinates */
 713             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 714                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 715                                                  &jx0,&jy0,&jz0);
 716
 717             /* Calculate displacement vector */
 718             dx00             = _mm256_sub_pd(ix0,jx0);
 719             dy00             = _mm256_sub_pd(iy0,jy0);
 720             dz00             = _mm256_sub_pd(iz0,jz0);
 721
 722             /* Calculate squared distance and things based on it */
 723             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 724
 725             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 726
 727             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 728
 729             /* Load parameters for j particles */
 730             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 731                                                                  charge+jnrC+0,charge+jnrD+0);
 732             vdwjidx0A        = 2*vdwtype[jnrA+0];
 733             vdwjidx0B        = 2*vdwtype[jnrB+0];
 734             vdwjidx0C        = 2*vdwtype[jnrC+0];
 735             vdwjidx0D        = 2*vdwtype[jnrD+0];
 736
 737             /**************************
 738              * CALCULATE INTERACTIONS *
 739              **************************/
 740
 741             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 742             {
 743
 744             r00              = _mm256_mul_pd(rsq00,rinv00);
 745             r00              = _mm256_andnot_pd(dummy_mask,r00);
 746
 747             /* Compute parameters for interactions between i and j atoms */
 748             qq00             = _mm256_mul_pd(iq0,jq0);
 749             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 750                                             vdwioffsetptr0+vdwjidx0B,
 751                                             vdwioffsetptr0+vdwjidx0C,
 752                                             vdwioffsetptr0+vdwjidx0D,
 753                                             &c6_00,&c12_00);
 754
 755             /* REACTION-FIELD ELECTROSTATICS */
 756             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 757
 758             /* LENNARD-JONES DISPERSION/REPULSION */
 759
 760             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 761             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 762             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 763             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 764             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 765
 766             d                = _mm256_sub_pd(r00,rswitch);
 767             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 768             d2               = _mm256_mul_pd(d,d);
 769             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)))))));
 770
 771             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 772
 773             /* Evaluate switch function */
 774             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 775             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 776             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 777
 778             fscal            = _mm256_add_pd(felec,fvdw);
 779
 780             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 781
 782             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 783
 784             /* Calculate temporary vectorial force */
 785             tx               = _mm256_mul_pd(fscal,dx00);
 786             ty               = _mm256_mul_pd(fscal,dy00);
 787             tz               = _mm256_mul_pd(fscal,dz00);
 788
 789             /* Update vectorial force */
 790             fix0             = _mm256_add_pd(fix0,tx);
 791             fiy0             = _mm256_add_pd(fiy0,ty);
 792             fiz0             = _mm256_add_pd(fiz0,tz);
 793
 794             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 795             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 796             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 797             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 798             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 799
 800             }
 801
 802             /* Inner loop uses 62 flops */
 803         }
 804
 805         /* End of innermost loop */
 806
 807         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 808                                                  f+i_coord_offset,fshift+i_shift_offset);
 809
 810         /* Increment number of inner iterations */
 811         inneriter                  += j_index_end - j_index_start;
 812
 813         /* Outer loop uses 7 flops */
 814     }
 815
 816     /* Increment number of outer iterations */
 817     outeriter        += nri;
 818
 819     /* Update outer/inner flops */
 820
 821     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);
 822 }