src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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_128_fma_single 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_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  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_128, 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              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  85     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  95     __m128           dummy_mask,cutoff_mask;
  96     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  97     __m128           one     = _mm_set1_ps(1.0);
  98     __m128           two     = _mm_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_ps(fr->ic->epsfac);
 111     charge           = mdatoms->chargeA;
 112     krf              = _mm_set1_ps(fr->ic->k_rf);
 113     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 114     crf              = _mm_set1_ps(fr->ic->c_rf);
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 120     rcutoff_scalar   = fr->ic->rcoulomb;
 121     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 122     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 123
 124     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 125     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 126
 127     /* Avoid stupid compiler warnings */
 128     jnrA = jnrB = jnrC = jnrD = 0;
 129     j_coord_offsetA = 0;
 130     j_coord_offsetB = 0;
 131     j_coord_offsetC = 0;
 132     j_coord_offsetD = 0;
 133
 134     outeriter        = 0;
 135     inneriter        = 0;
 136
 137     for(iidx=0;iidx<4*DIM;iidx++)
 138     {
 139         scratch[iidx] = 0.0;
 140     }
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 158
 159         fix0             = _mm_setzero_ps();
 160         fiy0             = _mm_setzero_ps();
 161         fiz0             = _mm_setzero_ps();
 162
 163         /* Load parameters for i particles */
 164         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 165         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 166
 167         /* Reset potential sums */
 168         velecsum         = _mm_setzero_ps();
 169         vvdwsum          = _mm_setzero_ps();
 170
 171         /* Start inner kernel loop */
 172         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 173         {
 174
 175             /* Get j neighbor index, and coordinate index */
 176             jnrA             = jjnr[jidx];
 177             jnrB             = jjnr[jidx+1];
 178             jnrC             = jjnr[jidx+2];
 179             jnrD             = jjnr[jidx+3];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182             j_coord_offsetC  = DIM*jnrC;
 183             j_coord_offsetD  = DIM*jnrD;
 184
 185             /* load j atom coordinates */
 186             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 187                                               x+j_coord_offsetC,x+j_coord_offsetD,
 188                                               &jx0,&jy0,&jz0);
 189
 190             /* Calculate displacement vector */
 191             dx00             = _mm_sub_ps(ix0,jx0);
 192             dy00             = _mm_sub_ps(iy0,jy0);
 193             dz00             = _mm_sub_ps(iz0,jz0);
 194
 195             /* Calculate squared distance and things based on it */
 196             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 197
 198             rinv00           = avx128fma_invsqrt_f(rsq00);
 199
 200             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 201
 202             /* Load parameters for j particles */
 203             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 204                                                               charge+jnrC+0,charge+jnrD+0);
 205             vdwjidx0A        = 2*vdwtype[jnrA+0];
 206             vdwjidx0B        = 2*vdwtype[jnrB+0];
 207             vdwjidx0C        = 2*vdwtype[jnrC+0];
 208             vdwjidx0D        = 2*vdwtype[jnrD+0];
 209
 210             /**************************
 211              * CALCULATE INTERACTIONS *
 212              **************************/
 213
 214             if (gmx_mm_any_lt(rsq00,rcutoff2))
 215             {
 216
 217             /* Compute parameters for interactions between i and j atoms */
 218             qq00             = _mm_mul_ps(iq0,jq0);
 219             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 220                                          vdwparam+vdwioffset0+vdwjidx0B,
 221                                          vdwparam+vdwioffset0+vdwjidx0C,
 222                                          vdwparam+vdwioffset0+vdwjidx0D,
 223                                          &c6_00,&c12_00);
 224
 225             /* REACTION-FIELD ELECTROSTATICS */
 226             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 227             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 228
 229             /* LENNARD-JONES DISPERSION/REPULSION */
 230
 231             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 232             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 233             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 234             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 235                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 236             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 237
 238             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 239
 240             /* Update potential sum for this i atom from the interaction with this j atom. */
 241             velec            = _mm_and_ps(velec,cutoff_mask);
 242             velecsum         = _mm_add_ps(velecsum,velec);
 243             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 244             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 245
 246             fscal            = _mm_add_ps(felec,fvdw);
 247
 248             fscal            = _mm_and_ps(fscal,cutoff_mask);
 249
 250              /* Update vectorial force */
 251             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 252             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 253             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 254
 255             fjptrA             = f+j_coord_offsetA;
 256             fjptrB             = f+j_coord_offsetB;
 257             fjptrC             = f+j_coord_offsetC;
 258             fjptrD             = f+j_coord_offsetD;
 259             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 260                                                    _mm_mul_ps(dx00,fscal),
 261                                                    _mm_mul_ps(dy00,fscal),
 262                                                    _mm_mul_ps(dz00,fscal));
 263
 264             }
 265
 266             /* Inner loop uses 57 flops */
 267         }
 268
 269         if(jidx<j_index_end)
 270         {
 271
 272             /* Get j neighbor index, and coordinate index */
 273             jnrlistA         = jjnr[jidx];
 274             jnrlistB         = jjnr[jidx+1];
 275             jnrlistC         = jjnr[jidx+2];
 276             jnrlistD         = jjnr[jidx+3];
 277             /* Sign of each element will be negative for non-real atoms.
 278              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 279              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 280              */
 281             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 282             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 283             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 284             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 285             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 286             j_coord_offsetA  = DIM*jnrA;
 287             j_coord_offsetB  = DIM*jnrB;
 288             j_coord_offsetC  = DIM*jnrC;
 289             j_coord_offsetD  = DIM*jnrD;
 290
 291             /* load j atom coordinates */
 292             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 293                                               x+j_coord_offsetC,x+j_coord_offsetD,
 294                                               &jx0,&jy0,&jz0);
 295
 296             /* Calculate displacement vector */
 297             dx00             = _mm_sub_ps(ix0,jx0);
 298             dy00             = _mm_sub_ps(iy0,jy0);
 299             dz00             = _mm_sub_ps(iz0,jz0);
 300
 301             /* Calculate squared distance and things based on it */
 302             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 303
 304             rinv00           = avx128fma_invsqrt_f(rsq00);
 305
 306             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 307
 308             /* Load parameters for j particles */
 309             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 310                                                               charge+jnrC+0,charge+jnrD+0);
 311             vdwjidx0A        = 2*vdwtype[jnrA+0];
 312             vdwjidx0B        = 2*vdwtype[jnrB+0];
 313             vdwjidx0C        = 2*vdwtype[jnrC+0];
 314             vdwjidx0D        = 2*vdwtype[jnrD+0];
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             if (gmx_mm_any_lt(rsq00,rcutoff2))
 321             {
 322
 323             /* Compute parameters for interactions between i and j atoms */
 324             qq00             = _mm_mul_ps(iq0,jq0);
 325             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 326                                          vdwparam+vdwioffset0+vdwjidx0B,
 327                                          vdwparam+vdwioffset0+vdwjidx0C,
 328                                          vdwparam+vdwioffset0+vdwjidx0D,
 329                                          &c6_00,&c12_00);
 330
 331             /* REACTION-FIELD ELECTROSTATICS */
 332             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 333             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 334
 335             /* LENNARD-JONES DISPERSION/REPULSION */
 336
 337             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 338             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 339             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 340             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 341                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 342             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 343
 344             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velec            = _mm_and_ps(velec,cutoff_mask);
 348             velec            = _mm_andnot_ps(dummy_mask,velec);
 349             velecsum         = _mm_add_ps(velecsum,velec);
 350             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 351             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 352             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 353
 354             fscal            = _mm_add_ps(felec,fvdw);
 355
 356             fscal            = _mm_and_ps(fscal,cutoff_mask);
 357
 358             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 359
 360              /* Update vectorial force */
 361             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 362             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 363             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 364
 365             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 366             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 367             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 368             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 369             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 370                                                    _mm_mul_ps(dx00,fscal),
 371                                                    _mm_mul_ps(dy00,fscal),
 372                                                    _mm_mul_ps(dz00,fscal));
 373
 374             }
 375
 376             /* Inner loop uses 57 flops */
 377         }
 378
 379         /* End of innermost loop */
 380
 381         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 382                                               f+i_coord_offset,fshift+i_shift_offset);
 383
 384         ggid                        = gid[iidx];
 385         /* Update potential energies */
 386         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 387         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 388
 389         /* Increment number of inner iterations */
 390         inneriter                  += j_index_end - j_index_start;
 391
 392         /* Outer loop uses 9 flops */
 393     }
 394
 395     /* Increment number of outer iterations */
 396     outeriter        += nri;
 397
 398     /* Update outer/inner flops */
 399
 400     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*57);
 401 }
 402 /*
 403  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 404  * Electrostatics interaction: ReactionField
 405  * VdW interaction:            LennardJones
 406  * Geometry:                   Particle-Particle
 407  * Calculate force/pot:        Force
 408  */
 409 void
 410 nb_kernel_ElecRFCut_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 411                     (t_nblist                    * gmx_restrict       nlist,
 412                      rvec                        * gmx_restrict          xx,
 413                      rvec                        * gmx_restrict          ff,
 414                      struct t_forcerec           * gmx_restrict          fr,
 415                      t_mdatoms                   * gmx_restrict     mdatoms,
 416                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 417                      t_nrnb                      * gmx_restrict        nrnb)
 418 {
 419     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 420      * just 0 for non-waters.
 421      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 422      * jnr indices corresponding to data put in the four positions in the SIMD register.
 423      */
 424     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 425     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 426     int              jnrA,jnrB,jnrC,jnrD;
 427     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 428     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 429     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 430     real             rcutoff_scalar;
 431     real             *shiftvec,*fshift,*x,*f;
 432     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 433     real             scratch[4*DIM];
 434     __m128           fscal,rcutoff,rcutoff2,jidxall;
 435     int              vdwioffset0;
 436     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 437     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 438     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 439     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 440     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 441     real             *charge;
 442     int              nvdwtype;
 443     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 444     int              *vdwtype;
 445     real             *vdwparam;
 446     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 447     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 448     __m128           dummy_mask,cutoff_mask;
 449     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 450     __m128           one     = _mm_set1_ps(1.0);
 451     __m128           two     = _mm_set1_ps(2.0);
 452     x                = xx[0];
 453     f                = ff[0];
 454
 455     nri              = nlist->nri;
 456     iinr             = nlist->iinr;
 457     jindex           = nlist->jindex;
 458     jjnr             = nlist->jjnr;
 459     shiftidx         = nlist->shift;
 460     gid              = nlist->gid;
 461     shiftvec         = fr->shift_vec[0];
 462     fshift           = fr->fshift[0];
 463     facel            = _mm_set1_ps(fr->ic->epsfac);
 464     charge           = mdatoms->chargeA;
 465     krf              = _mm_set1_ps(fr->ic->k_rf);
 466     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 467     crf              = _mm_set1_ps(fr->ic->c_rf);
 468     nvdwtype         = fr->ntype;
 469     vdwparam         = fr->nbfp;
 470     vdwtype          = mdatoms->typeA;
 471
 472     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 473     rcutoff_scalar   = fr->ic->rcoulomb;
 474     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 475     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 476
 477     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 478     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 479
 480     /* Avoid stupid compiler warnings */
 481     jnrA = jnrB = jnrC = jnrD = 0;
 482     j_coord_offsetA = 0;
 483     j_coord_offsetB = 0;
 484     j_coord_offsetC = 0;
 485     j_coord_offsetD = 0;
 486
 487     outeriter        = 0;
 488     inneriter        = 0;
 489
 490     for(iidx=0;iidx<4*DIM;iidx++)
 491     {
 492         scratch[iidx] = 0.0;
 493     }
 494
 495     /* Start outer loop over neighborlists */
 496     for(iidx=0; iidx<nri; iidx++)
 497     {
 498         /* Load shift vector for this list */
 499         i_shift_offset   = DIM*shiftidx[iidx];
 500
 501         /* Load limits for loop over neighbors */
 502         j_index_start    = jindex[iidx];
 503         j_index_end      = jindex[iidx+1];
 504
 505         /* Get outer coordinate index */
 506         inr              = iinr[iidx];
 507         i_coord_offset   = DIM*inr;
 508
 509         /* Load i particle coords and add shift vector */
 510         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 511
 512         fix0             = _mm_setzero_ps();
 513         fiy0             = _mm_setzero_ps();
 514         fiz0             = _mm_setzero_ps();
 515
 516         /* Load parameters for i particles */
 517         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 518         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 519
 520         /* Start inner kernel loop */
 521         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 522         {
 523
 524             /* Get j neighbor index, and coordinate index */
 525             jnrA             = jjnr[jidx];
 526             jnrB             = jjnr[jidx+1];
 527             jnrC             = jjnr[jidx+2];
 528             jnrD             = jjnr[jidx+3];
 529             j_coord_offsetA  = DIM*jnrA;
 530             j_coord_offsetB  = DIM*jnrB;
 531             j_coord_offsetC  = DIM*jnrC;
 532             j_coord_offsetD  = DIM*jnrD;
 533
 534             /* load j atom coordinates */
 535             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 536                                               x+j_coord_offsetC,x+j_coord_offsetD,
 537                                               &jx0,&jy0,&jz0);
 538
 539             /* Calculate displacement vector */
 540             dx00             = _mm_sub_ps(ix0,jx0);
 541             dy00             = _mm_sub_ps(iy0,jy0);
 542             dz00             = _mm_sub_ps(iz0,jz0);
 543
 544             /* Calculate squared distance and things based on it */
 545             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 546
 547             rinv00           = avx128fma_invsqrt_f(rsq00);
 548
 549             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 550
 551             /* Load parameters for j particles */
 552             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 553                                                               charge+jnrC+0,charge+jnrD+0);
 554             vdwjidx0A        = 2*vdwtype[jnrA+0];
 555             vdwjidx0B        = 2*vdwtype[jnrB+0];
 556             vdwjidx0C        = 2*vdwtype[jnrC+0];
 557             vdwjidx0D        = 2*vdwtype[jnrD+0];
 558
 559             /**************************
 560              * CALCULATE INTERACTIONS *
 561              **************************/
 562
 563             if (gmx_mm_any_lt(rsq00,rcutoff2))
 564             {
 565
 566             /* Compute parameters for interactions between i and j atoms */
 567             qq00             = _mm_mul_ps(iq0,jq0);
 568             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 569                                          vdwparam+vdwioffset0+vdwjidx0B,
 570                                          vdwparam+vdwioffset0+vdwjidx0C,
 571                                          vdwparam+vdwioffset0+vdwjidx0D,
 572                                          &c6_00,&c12_00);
 573
 574             /* REACTION-FIELD ELECTROSTATICS */
 575             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 576
 577             /* LENNARD-JONES DISPERSION/REPULSION */
 578
 579             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 580             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 581
 582             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 583
 584             fscal            = _mm_add_ps(felec,fvdw);
 585
 586             fscal            = _mm_and_ps(fscal,cutoff_mask);
 587
 588              /* Update vectorial force */
 589             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 590             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 591             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 592
 593             fjptrA             = f+j_coord_offsetA;
 594             fjptrB             = f+j_coord_offsetB;
 595             fjptrC             = f+j_coord_offsetC;
 596             fjptrD             = f+j_coord_offsetD;
 597             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 598                                                    _mm_mul_ps(dx00,fscal),
 599                                                    _mm_mul_ps(dy00,fscal),
 600                                                    _mm_mul_ps(dz00,fscal));
 601
 602             }
 603
 604             /* Inner loop uses 40 flops */
 605         }
 606
 607         if(jidx<j_index_end)
 608         {
 609
 610             /* Get j neighbor index, and coordinate index */
 611             jnrlistA         = jjnr[jidx];
 612             jnrlistB         = jjnr[jidx+1];
 613             jnrlistC         = jjnr[jidx+2];
 614             jnrlistD         = jjnr[jidx+3];
 615             /* Sign of each element will be negative for non-real atoms.
 616              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 617              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 618              */
 619             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 620             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 621             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 622             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 623             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 624             j_coord_offsetA  = DIM*jnrA;
 625             j_coord_offsetB  = DIM*jnrB;
 626             j_coord_offsetC  = DIM*jnrC;
 627             j_coord_offsetD  = DIM*jnrD;
 628
 629             /* load j atom coordinates */
 630             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 631                                               x+j_coord_offsetC,x+j_coord_offsetD,
 632                                               &jx0,&jy0,&jz0);
 633
 634             /* Calculate displacement vector */
 635             dx00             = _mm_sub_ps(ix0,jx0);
 636             dy00             = _mm_sub_ps(iy0,jy0);
 637             dz00             = _mm_sub_ps(iz0,jz0);
 638
 639             /* Calculate squared distance and things based on it */
 640             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 641
 642             rinv00           = avx128fma_invsqrt_f(rsq00);
 643
 644             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 645
 646             /* Load parameters for j particles */
 647             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 648                                                               charge+jnrC+0,charge+jnrD+0);
 649             vdwjidx0A        = 2*vdwtype[jnrA+0];
 650             vdwjidx0B        = 2*vdwtype[jnrB+0];
 651             vdwjidx0C        = 2*vdwtype[jnrC+0];
 652             vdwjidx0D        = 2*vdwtype[jnrD+0];
 653
 654             /**************************
 655              * CALCULATE INTERACTIONS *
 656              **************************/
 657
 658             if (gmx_mm_any_lt(rsq00,rcutoff2))
 659             {
 660
 661             /* Compute parameters for interactions between i and j atoms */
 662             qq00             = _mm_mul_ps(iq0,jq0);
 663             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 664                                          vdwparam+vdwioffset0+vdwjidx0B,
 665                                          vdwparam+vdwioffset0+vdwjidx0C,
 666                                          vdwparam+vdwioffset0+vdwjidx0D,
 667                                          &c6_00,&c12_00);
 668
 669             /* REACTION-FIELD ELECTROSTATICS */
 670             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 671
 672             /* LENNARD-JONES DISPERSION/REPULSION */
 673
 674             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 675             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 676
 677             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 678
 679             fscal            = _mm_add_ps(felec,fvdw);
 680
 681             fscal            = _mm_and_ps(fscal,cutoff_mask);
 682
 683             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 684
 685              /* Update vectorial force */
 686             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 687             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 688             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 689
 690             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 691             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 692             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 693             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 694             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 695                                                    _mm_mul_ps(dx00,fscal),
 696                                                    _mm_mul_ps(dy00,fscal),
 697                                                    _mm_mul_ps(dz00,fscal));
 698
 699             }
 700
 701             /* Inner loop uses 40 flops */
 702         }
 703
 704         /* End of innermost loop */
 705
 706         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 707                                               f+i_coord_offset,fshift+i_shift_offset);
 708
 709         /* Increment number of inner iterations */
 710         inneriter                  += j_index_end - j_index_start;
 711
 712         /* Outer loop uses 7 flops */
 713     }
 714
 715     /* Increment number of outer iterations */
 716     outeriter        += nri;
 717
 718     /* Update outer/inner flops */
 719
 720     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*40);
 721 }