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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  98     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  99     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 100     real             rswitch_scalar,d_scalar;
 101     __m256d          dummy_mask,cutoff_mask;
 102     __m128           tmpmask0,tmpmask1;
 103     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 104     __m256d          one     = _mm256_set1_pd(1.0);
 105     __m256d          two     = _mm256_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm256_set1_pd(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = _mm256_set1_pd(fr->ic->k_rf);
 120     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 121     crf              = _mm256_set1_pd(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 127     rcutoff_scalar   = fr->rcoulomb;
 128     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 129     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 130
 131     rswitch_scalar   = fr->rvdw_switch;
 132     rswitch          = _mm256_set1_pd(rswitch_scalar);
 133     /* Setup switch parameters */
 134     d_scalar         = rcutoff_scalar-rswitch_scalar;
 135     d                = _mm256_set1_pd(d_scalar);
 136     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 137     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 138     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 140     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 141     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 142
 143     /* Avoid stupid compiler warnings */
 144     jnrA = jnrB = jnrC = jnrD = 0;
 145     j_coord_offsetA = 0;
 146     j_coord_offsetB = 0;
 147     j_coord_offsetC = 0;
 148     j_coord_offsetD = 0;
 149
 150     outeriter        = 0;
 151     inneriter        = 0;
 152
 153     for(iidx=0;iidx<4*DIM;iidx++)
 154     {
 155         scratch[iidx] = 0.0;
 156     }
 157
 158     /* Start outer loop over neighborlists */
 159     for(iidx=0; iidx<nri; iidx++)
 160     {
 161         /* Load shift vector for this list */
 162         i_shift_offset   = DIM*shiftidx[iidx];
 163
 164         /* Load limits for loop over neighbors */
 165         j_index_start    = jindex[iidx];
 166         j_index_end      = jindex[iidx+1];
 167
 168         /* Get outer coordinate index */
 169         inr              = iinr[iidx];
 170         i_coord_offset   = DIM*inr;
 171
 172         /* Load i particle coords and add shift vector */
 173         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 174
 175         fix0             = _mm256_setzero_pd();
 176         fiy0             = _mm256_setzero_pd();
 177         fiz0             = _mm256_setzero_pd();
 178
 179         /* Load parameters for i particles */
 180         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 181         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 182
 183         /* Reset potential sums */
 184         velecsum         = _mm256_setzero_pd();
 185         vvdwsum          = _mm256_setzero_pd();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             jnrC             = jjnr[jidx+2];
 195             jnrD             = jjnr[jidx+3];
 196             j_coord_offsetA  = DIM*jnrA;
 197             j_coord_offsetB  = DIM*jnrB;
 198             j_coord_offsetC  = DIM*jnrC;
 199             j_coord_offsetD  = DIM*jnrD;
 200
 201             /* load j atom coordinates */
 202             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 203                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 204                                                  &jx0,&jy0,&jz0);
 205
 206             /* Calculate displacement vector */
 207             dx00             = _mm256_sub_pd(ix0,jx0);
 208             dy00             = _mm256_sub_pd(iy0,jy0);
 209             dz00             = _mm256_sub_pd(iz0,jz0);
 210
 211             /* Calculate squared distance and things based on it */
 212             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 213
 214             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 215
 216             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 217
 218             /* Load parameters for j particles */
 219             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 220                                                                  charge+jnrC+0,charge+jnrD+0);
 221             vdwjidx0A        = 2*vdwtype[jnrA+0];
 222             vdwjidx0B        = 2*vdwtype[jnrB+0];
 223             vdwjidx0C        = 2*vdwtype[jnrC+0];
 224             vdwjidx0D        = 2*vdwtype[jnrD+0];
 225
 226             /**************************
 227              * CALCULATE INTERACTIONS *
 228              **************************/
 229
 230             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 231             {
 232
 233             r00              = _mm256_mul_pd(rsq00,rinv00);
 234
 235             /* Compute parameters for interactions between i and j atoms */
 236             qq00             = _mm256_mul_pd(iq0,jq0);
 237             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 238                                             vdwioffsetptr0+vdwjidx0B,
 239                                             vdwioffsetptr0+vdwjidx0C,
 240                                             vdwioffsetptr0+vdwjidx0D,
 241                                             &c6_00,&c12_00);
 242
 243             /* REACTION-FIELD ELECTROSTATICS */
 244             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 245             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 246
 247             /* LENNARD-JONES DISPERSION/REPULSION */
 248
 249             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 250             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 251             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 252             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 253             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 254
 255             d                = _mm256_sub_pd(r00,rswitch);
 256             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 257             d2               = _mm256_mul_pd(d,d);
 258             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)))))));
 259
 260             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 261
 262             /* Evaluate switch function */
 263             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 264             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 265             vvdw             = _mm256_mul_pd(vvdw,sw);
 266             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 267
 268             /* Update potential sum for this i atom from the interaction with this j atom. */
 269             velec            = _mm256_and_pd(velec,cutoff_mask);
 270             velecsum         = _mm256_add_pd(velecsum,velec);
 271             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 272             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 273
 274             fscal            = _mm256_add_pd(felec,fvdw);
 275
 276             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 277
 278             /* Calculate temporary vectorial force */
 279             tx               = _mm256_mul_pd(fscal,dx00);
 280             ty               = _mm256_mul_pd(fscal,dy00);
 281             tz               = _mm256_mul_pd(fscal,dz00);
 282
 283             /* Update vectorial force */
 284             fix0             = _mm256_add_pd(fix0,tx);
 285             fiy0             = _mm256_add_pd(fiy0,ty);
 286             fiz0             = _mm256_add_pd(fiz0,tz);
 287
 288             fjptrA             = f+j_coord_offsetA;
 289             fjptrB             = f+j_coord_offsetB;
 290             fjptrC             = f+j_coord_offsetC;
 291             fjptrD             = f+j_coord_offsetD;
 292             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 293
 294             }
 295
 296             /* Inner loop uses 70 flops */
 297         }
 298
 299         if(jidx<j_index_end)
 300         {
 301
 302             /* Get j neighbor index, and coordinate index */
 303             jnrlistA         = jjnr[jidx];
 304             jnrlistB         = jjnr[jidx+1];
 305             jnrlistC         = jjnr[jidx+2];
 306             jnrlistD         = jjnr[jidx+3];
 307             /* Sign of each element will be negative for non-real atoms.
 308              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 309              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 310              */
 311             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 312
 313             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 314             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 315             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 316
 317             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 318             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 319             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 320             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 321             j_coord_offsetA  = DIM*jnrA;
 322             j_coord_offsetB  = DIM*jnrB;
 323             j_coord_offsetC  = DIM*jnrC;
 324             j_coord_offsetD  = DIM*jnrD;
 325
 326             /* load j atom coordinates */
 327             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 328                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 329                                                  &jx0,&jy0,&jz0);
 330
 331             /* Calculate displacement vector */
 332             dx00             = _mm256_sub_pd(ix0,jx0);
 333             dy00             = _mm256_sub_pd(iy0,jy0);
 334             dz00             = _mm256_sub_pd(iz0,jz0);
 335
 336             /* Calculate squared distance and things based on it */
 337             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 338
 339             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 340
 341             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 342
 343             /* Load parameters for j particles */
 344             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 345                                                                  charge+jnrC+0,charge+jnrD+0);
 346             vdwjidx0A        = 2*vdwtype[jnrA+0];
 347             vdwjidx0B        = 2*vdwtype[jnrB+0];
 348             vdwjidx0C        = 2*vdwtype[jnrC+0];
 349             vdwjidx0D        = 2*vdwtype[jnrD+0];
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 356             {
 357
 358             r00              = _mm256_mul_pd(rsq00,rinv00);
 359             r00              = _mm256_andnot_pd(dummy_mask,r00);
 360
 361             /* Compute parameters for interactions between i and j atoms */
 362             qq00             = _mm256_mul_pd(iq0,jq0);
 363             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 364                                             vdwioffsetptr0+vdwjidx0B,
 365                                             vdwioffsetptr0+vdwjidx0C,
 366                                             vdwioffsetptr0+vdwjidx0D,
 367                                             &c6_00,&c12_00);
 368
 369             /* REACTION-FIELD ELECTROSTATICS */
 370             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
 371             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 372
 373             /* LENNARD-JONES DISPERSION/REPULSION */
 374
 375             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 376             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 377             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 378             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 379             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 380
 381             d                = _mm256_sub_pd(r00,rswitch);
 382             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 383             d2               = _mm256_mul_pd(d,d);
 384             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)))))));
 385
 386             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 387
 388             /* Evaluate switch function */
 389             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 390             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 391             vvdw             = _mm256_mul_pd(vvdw,sw);
 392             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 393
 394             /* Update potential sum for this i atom from the interaction with this j atom. */
 395             velec            = _mm256_and_pd(velec,cutoff_mask);
 396             velec            = _mm256_andnot_pd(dummy_mask,velec);
 397             velecsum         = _mm256_add_pd(velecsum,velec);
 398             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 399             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 400             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 401
 402             fscal            = _mm256_add_pd(felec,fvdw);
 403
 404             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 405
 406             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 407
 408             /* Calculate temporary vectorial force */
 409             tx               = _mm256_mul_pd(fscal,dx00);
 410             ty               = _mm256_mul_pd(fscal,dy00);
 411             tz               = _mm256_mul_pd(fscal,dz00);
 412
 413             /* Update vectorial force */
 414             fix0             = _mm256_add_pd(fix0,tx);
 415             fiy0             = _mm256_add_pd(fiy0,ty);
 416             fiz0             = _mm256_add_pd(fiz0,tz);
 417
 418             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 419             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 420             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 421             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 422             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 423
 424             }
 425
 426             /* Inner loop uses 71 flops */
 427         }
 428
 429         /* End of innermost loop */
 430
 431         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 432                                                  f+i_coord_offset,fshift+i_shift_offset);
 433
 434         ggid                        = gid[iidx];
 435         /* Update potential energies */
 436         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 437         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 438
 439         /* Increment number of inner iterations */
 440         inneriter                  += j_index_end - j_index_start;
 441
 442         /* Outer loop uses 9 flops */
 443     }
 444
 445     /* Increment number of outer iterations */
 446     outeriter        += nri;
 447
 448     /* Update outer/inner flops */
 449
 450     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*71);
 451 }
 452 /*
 453  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 454  * Electrostatics interaction: ReactionField
 455  * VdW interaction:            LennardJones
 456  * Geometry:                   Particle-Particle
 457  * Calculate force/pot:        Force
 458  */
 459 void
 460 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_256_double
 461                     (t_nblist                    * gmx_restrict       nlist,
 462                      rvec                        * gmx_restrict          xx,
 463                      rvec                        * gmx_restrict          ff,
 464                      t_forcerec                  * gmx_restrict          fr,
 465                      t_mdatoms                   * gmx_restrict     mdatoms,
 466                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 467                      t_nrnb                      * gmx_restrict        nrnb)
 468 {
 469     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 470      * just 0 for non-waters.
 471      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 472      * jnr indices corresponding to data put in the four positions in the SIMD register.
 473      */
 474     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 475     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 476     int              jnrA,jnrB,jnrC,jnrD;
 477     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 478     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 479     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 480     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 481     real             rcutoff_scalar;
 482     real             *shiftvec,*fshift,*x,*f;
 483     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 484     real             scratch[4*DIM];
 485     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 486     real *           vdwioffsetptr0;
 487     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 488     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 489     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 490     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 491     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 492     real             *charge;
 493     int              nvdwtype;
 494     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 495     int              *vdwtype;
 496     real             *vdwparam;
 497     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 498     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 499     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 500     real             rswitch_scalar,d_scalar;
 501     __m256d          dummy_mask,cutoff_mask;
 502     __m128           tmpmask0,tmpmask1;
 503     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 504     __m256d          one     = _mm256_set1_pd(1.0);
 505     __m256d          two     = _mm256_set1_pd(2.0);
 506     x                = xx[0];
 507     f                = ff[0];
 508
 509     nri              = nlist->nri;
 510     iinr             = nlist->iinr;
 511     jindex           = nlist->jindex;
 512     jjnr             = nlist->jjnr;
 513     shiftidx         = nlist->shift;
 514     gid              = nlist->gid;
 515     shiftvec         = fr->shift_vec[0];
 516     fshift           = fr->fshift[0];
 517     facel            = _mm256_set1_pd(fr->epsfac);
 518     charge           = mdatoms->chargeA;
 519     krf              = _mm256_set1_pd(fr->ic->k_rf);
 520     krf2             = _mm256_set1_pd(fr->ic->k_rf*2.0);
 521     crf              = _mm256_set1_pd(fr->ic->c_rf);
 522     nvdwtype         = fr->ntype;
 523     vdwparam         = fr->nbfp;
 524     vdwtype          = mdatoms->typeA;
 525
 526     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 527     rcutoff_scalar   = fr->rcoulomb;
 528     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 529     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 530
 531     rswitch_scalar   = fr->rvdw_switch;
 532     rswitch          = _mm256_set1_pd(rswitch_scalar);
 533     /* Setup switch parameters */
 534     d_scalar         = rcutoff_scalar-rswitch_scalar;
 535     d                = _mm256_set1_pd(d_scalar);
 536     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 537     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 538     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 539     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 540     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 541     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 542
 543     /* Avoid stupid compiler warnings */
 544     jnrA = jnrB = jnrC = jnrD = 0;
 545     j_coord_offsetA = 0;
 546     j_coord_offsetB = 0;
 547     j_coord_offsetC = 0;
 548     j_coord_offsetD = 0;
 549
 550     outeriter        = 0;
 551     inneriter        = 0;
 552
 553     for(iidx=0;iidx<4*DIM;iidx++)
 554     {
 555         scratch[iidx] = 0.0;
 556     }
 557
 558     /* Start outer loop over neighborlists */
 559     for(iidx=0; iidx<nri; iidx++)
 560     {
 561         /* Load shift vector for this list */
 562         i_shift_offset   = DIM*shiftidx[iidx];
 563
 564         /* Load limits for loop over neighbors */
 565         j_index_start    = jindex[iidx];
 566         j_index_end      = jindex[iidx+1];
 567
 568         /* Get outer coordinate index */
 569         inr              = iinr[iidx];
 570         i_coord_offset   = DIM*inr;
 571
 572         /* Load i particle coords and add shift vector */
 573         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 574
 575         fix0             = _mm256_setzero_pd();
 576         fiy0             = _mm256_setzero_pd();
 577         fiz0             = _mm256_setzero_pd();
 578
 579         /* Load parameters for i particles */
 580         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 581         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 582
 583         /* Start inner kernel loop */
 584         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 585         {
 586
 587             /* Get j neighbor index, and coordinate index */
 588             jnrA             = jjnr[jidx];
 589             jnrB             = jjnr[jidx+1];
 590             jnrC             = jjnr[jidx+2];
 591             jnrD             = jjnr[jidx+3];
 592             j_coord_offsetA  = DIM*jnrA;
 593             j_coord_offsetB  = DIM*jnrB;
 594             j_coord_offsetC  = DIM*jnrC;
 595             j_coord_offsetD  = DIM*jnrD;
 596
 597             /* load j atom coordinates */
 598             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 599                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 600                                                  &jx0,&jy0,&jz0);
 601
 602             /* Calculate displacement vector */
 603             dx00             = _mm256_sub_pd(ix0,jx0);
 604             dy00             = _mm256_sub_pd(iy0,jy0);
 605             dz00             = _mm256_sub_pd(iz0,jz0);
 606
 607             /* Calculate squared distance and things based on it */
 608             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 609
 610             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 611
 612             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 613
 614             /* Load parameters for j particles */
 615             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 616                                                                  charge+jnrC+0,charge+jnrD+0);
 617             vdwjidx0A        = 2*vdwtype[jnrA+0];
 618             vdwjidx0B        = 2*vdwtype[jnrB+0];
 619             vdwjidx0C        = 2*vdwtype[jnrC+0];
 620             vdwjidx0D        = 2*vdwtype[jnrD+0];
 621
 622             /**************************
 623              * CALCULATE INTERACTIONS *
 624              **************************/
 625
 626             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 627             {
 628
 629             r00              = _mm256_mul_pd(rsq00,rinv00);
 630
 631             /* Compute parameters for interactions between i and j atoms */
 632             qq00             = _mm256_mul_pd(iq0,jq0);
 633             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 634                                             vdwioffsetptr0+vdwjidx0B,
 635                                             vdwioffsetptr0+vdwjidx0C,
 636                                             vdwioffsetptr0+vdwjidx0D,
 637                                             &c6_00,&c12_00);
 638
 639             /* REACTION-FIELD ELECTROSTATICS */
 640             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 641
 642             /* LENNARD-JONES DISPERSION/REPULSION */
 643
 644             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 645             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 646             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 647             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 648             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 649
 650             d                = _mm256_sub_pd(r00,rswitch);
 651             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 652             d2               = _mm256_mul_pd(d,d);
 653             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)))))));
 654
 655             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 656
 657             /* Evaluate switch function */
 658             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 659             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 660             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 661
 662             fscal            = _mm256_add_pd(felec,fvdw);
 663
 664             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 665
 666             /* Calculate temporary vectorial force */
 667             tx               = _mm256_mul_pd(fscal,dx00);
 668             ty               = _mm256_mul_pd(fscal,dy00);
 669             tz               = _mm256_mul_pd(fscal,dz00);
 670
 671             /* Update vectorial force */
 672             fix0             = _mm256_add_pd(fix0,tx);
 673             fiy0             = _mm256_add_pd(fiy0,ty);
 674             fiz0             = _mm256_add_pd(fiz0,tz);
 675
 676             fjptrA             = f+j_coord_offsetA;
 677             fjptrB             = f+j_coord_offsetB;
 678             fjptrC             = f+j_coord_offsetC;
 679             fjptrD             = f+j_coord_offsetD;
 680             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 681
 682             }
 683
 684             /* Inner loop uses 61 flops */
 685         }
 686
 687         if(jidx<j_index_end)
 688         {
 689
 690             /* Get j neighbor index, and coordinate index */
 691             jnrlistA         = jjnr[jidx];
 692             jnrlistB         = jjnr[jidx+1];
 693             jnrlistC         = jjnr[jidx+2];
 694             jnrlistD         = jjnr[jidx+3];
 695             /* Sign of each element will be negative for non-real atoms.
 696              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 697              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 698              */
 699             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 700
 701             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 702             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 703             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 704
 705             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 706             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 707             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 708             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 709             j_coord_offsetA  = DIM*jnrA;
 710             j_coord_offsetB  = DIM*jnrB;
 711             j_coord_offsetC  = DIM*jnrC;
 712             j_coord_offsetD  = DIM*jnrD;
 713
 714             /* load j atom coordinates */
 715             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 716                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 717                                                  &jx0,&jy0,&jz0);
 718
 719             /* Calculate displacement vector */
 720             dx00             = _mm256_sub_pd(ix0,jx0);
 721             dy00             = _mm256_sub_pd(iy0,jy0);
 722             dz00             = _mm256_sub_pd(iz0,jz0);
 723
 724             /* Calculate squared distance and things based on it */
 725             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 726
 727             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 728
 729             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 730
 731             /* Load parameters for j particles */
 732             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 733                                                                  charge+jnrC+0,charge+jnrD+0);
 734             vdwjidx0A        = 2*vdwtype[jnrA+0];
 735             vdwjidx0B        = 2*vdwtype[jnrB+0];
 736             vdwjidx0C        = 2*vdwtype[jnrC+0];
 737             vdwjidx0D        = 2*vdwtype[jnrD+0];
 738
 739             /**************************
 740              * CALCULATE INTERACTIONS *
 741              **************************/
 742
 743             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 744             {
 745
 746             r00              = _mm256_mul_pd(rsq00,rinv00);
 747             r00              = _mm256_andnot_pd(dummy_mask,r00);
 748
 749             /* Compute parameters for interactions between i and j atoms */
 750             qq00             = _mm256_mul_pd(iq0,jq0);
 751             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 752                                             vdwioffsetptr0+vdwjidx0B,
 753                                             vdwioffsetptr0+vdwjidx0C,
 754                                             vdwioffsetptr0+vdwjidx0D,
 755                                             &c6_00,&c12_00);
 756
 757             /* REACTION-FIELD ELECTROSTATICS */
 758             felec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
 759
 760             /* LENNARD-JONES DISPERSION/REPULSION */
 761
 762             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 763             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 764             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 765             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 766             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 767
 768             d                = _mm256_sub_pd(r00,rswitch);
 769             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 770             d2               = _mm256_mul_pd(d,d);
 771             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)))))));
 772
 773             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 774
 775             /* Evaluate switch function */
 776             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 777             fvdw             = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
 778             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 779
 780             fscal            = _mm256_add_pd(felec,fvdw);
 781
 782             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 783
 784             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 785
 786             /* Calculate temporary vectorial force */
 787             tx               = _mm256_mul_pd(fscal,dx00);
 788             ty               = _mm256_mul_pd(fscal,dy00);
 789             tz               = _mm256_mul_pd(fscal,dz00);
 790
 791             /* Update vectorial force */
 792             fix0             = _mm256_add_pd(fix0,tx);
 793             fiy0             = _mm256_add_pd(fiy0,ty);
 794             fiz0             = _mm256_add_pd(fiz0,tz);
 795
 796             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 797             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 798             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 799             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 800             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 801
 802             }
 803
 804             /* Inner loop uses 62 flops */
 805         }
 806
 807         /* End of innermost loop */
 808
 809         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 810                                                  f+i_coord_offset,fshift+i_shift_offset);
 811
 812         /* Increment number of inner iterations */
 813         inneriter                  += j_index_end - j_index_start;
 814
 815         /* Outer loop uses 7 flops */
 816     }
 817
 818     /* Increment number of outer iterations */
 819     outeriter        += nri;
 820
 821     /* Update outer/inner flops */
 822
 823     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);
 824 }