src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_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_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_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     __m128i          ewitab;
  96     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  97     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  98     real             *ewtab;
  99     __m128           dummy_mask,cutoff_mask;
 100     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 101     __m128           one     = _mm_set1_ps(1.0);
 102     __m128           two     = _mm_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_ps(fr->ic->epsfac);
 115     charge           = mdatoms->chargeA;
 116     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119
 120     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 121     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 122     beta2            = _mm_mul_ps(beta,beta);
 123     beta3            = _mm_mul_ps(beta,beta2);
 124     ewtab            = fr->ic->tabq_coul_FDV0;
 125     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 126     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 127
 128     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 129     rcutoff_scalar   = fr->ic->rcoulomb;
 130     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 131     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 132
 133     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 134     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 135
 136     /* Avoid stupid compiler warnings */
 137     jnrA = jnrB = jnrC = jnrD = 0;
 138     j_coord_offsetA = 0;
 139     j_coord_offsetB = 0;
 140     j_coord_offsetC = 0;
 141     j_coord_offsetD = 0;
 142
 143     outeriter        = 0;
 144     inneriter        = 0;
 145
 146     for(iidx=0;iidx<4*DIM;iidx++)
 147     {
 148         scratch[iidx] = 0.0;
 149     }
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 167
 168         fix0             = _mm_setzero_ps();
 169         fiy0             = _mm_setzero_ps();
 170         fiz0             = _mm_setzero_ps();
 171
 172         /* Load parameters for i particles */
 173         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 174         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 175
 176         /* Reset potential sums */
 177         velecsum         = _mm_setzero_ps();
 178         vvdwsum          = _mm_setzero_ps();
 179
 180         /* Start inner kernel loop */
 181         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 182         {
 183
 184             /* Get j neighbor index, and coordinate index */
 185             jnrA             = jjnr[jidx];
 186             jnrB             = jjnr[jidx+1];
 187             jnrC             = jjnr[jidx+2];
 188             jnrD             = jjnr[jidx+3];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193
 194             /* load j atom coordinates */
 195             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 196                                               x+j_coord_offsetC,x+j_coord_offsetD,
 197                                               &jx0,&jy0,&jz0);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _mm_sub_ps(ix0,jx0);
 201             dy00             = _mm_sub_ps(iy0,jy0);
 202             dz00             = _mm_sub_ps(iz0,jz0);
 203
 204             /* Calculate squared distance and things based on it */
 205             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 206
 207             rinv00           = avx128fma_invsqrt_f(rsq00);
 208
 209             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 213                                                               charge+jnrC+0,charge+jnrD+0);
 214             vdwjidx0A        = 2*vdwtype[jnrA+0];
 215             vdwjidx0B        = 2*vdwtype[jnrB+0];
 216             vdwjidx0C        = 2*vdwtype[jnrC+0];
 217             vdwjidx0D        = 2*vdwtype[jnrD+0];
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             if (gmx_mm_any_lt(rsq00,rcutoff2))
 224             {
 225
 226             r00              = _mm_mul_ps(rsq00,rinv00);
 227
 228             /* Compute parameters for interactions between i and j atoms */
 229             qq00             = _mm_mul_ps(iq0,jq0);
 230             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 231                                          vdwparam+vdwioffset0+vdwjidx0B,
 232                                          vdwparam+vdwioffset0+vdwjidx0C,
 233                                          vdwparam+vdwioffset0+vdwjidx0D,
 234                                          &c6_00,&c12_00);
 235
 236             /* EWALD ELECTROSTATICS */
 237
 238             /* Analytical PME correction */
 239             zeta2            = _mm_mul_ps(beta2,rsq00);
 240             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 241             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 242             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 243             felec            = _mm_mul_ps(qq00,felec);
 244             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 245             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 246             velec            = _mm_mul_ps(qq00,velec);
 247
 248             /* LENNARD-JONES DISPERSION/REPULSION */
 249
 250             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 251             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 252             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 253             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 254                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 255             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 256
 257             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 258
 259             /* Update potential sum for this i atom from the interaction with this j atom. */
 260             velec            = _mm_and_ps(velec,cutoff_mask);
 261             velecsum         = _mm_add_ps(velecsum,velec);
 262             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 263             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 264
 265             fscal            = _mm_add_ps(felec,fvdw);
 266
 267             fscal            = _mm_and_ps(fscal,cutoff_mask);
 268
 269              /* Update vectorial force */
 270             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 271             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 272             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 273
 274             fjptrA             = f+j_coord_offsetA;
 275             fjptrB             = f+j_coord_offsetB;
 276             fjptrC             = f+j_coord_offsetC;
 277             fjptrD             = f+j_coord_offsetD;
 278             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 279                                                    _mm_mul_ps(dx00,fscal),
 280                                                    _mm_mul_ps(dy00,fscal),
 281                                                    _mm_mul_ps(dz00,fscal));
 282
 283             }
 284
 285             /* Inner loop uses 51 flops */
 286         }
 287
 288         if(jidx<j_index_end)
 289         {
 290
 291             /* Get j neighbor index, and coordinate index */
 292             jnrlistA         = jjnr[jidx];
 293             jnrlistB         = jjnr[jidx+1];
 294             jnrlistC         = jjnr[jidx+2];
 295             jnrlistD         = jjnr[jidx+3];
 296             /* Sign of each element will be negative for non-real atoms.
 297              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 298              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 299              */
 300             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 301             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 302             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 303             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 304             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 305             j_coord_offsetA  = DIM*jnrA;
 306             j_coord_offsetB  = DIM*jnrB;
 307             j_coord_offsetC  = DIM*jnrC;
 308             j_coord_offsetD  = DIM*jnrD;
 309
 310             /* load j atom coordinates */
 311             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 312                                               x+j_coord_offsetC,x+j_coord_offsetD,
 313                                               &jx0,&jy0,&jz0);
 314
 315             /* Calculate displacement vector */
 316             dx00             = _mm_sub_ps(ix0,jx0);
 317             dy00             = _mm_sub_ps(iy0,jy0);
 318             dz00             = _mm_sub_ps(iz0,jz0);
 319
 320             /* Calculate squared distance and things based on it */
 321             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 322
 323             rinv00           = avx128fma_invsqrt_f(rsq00);
 324
 325             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 326
 327             /* Load parameters for j particles */
 328             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 329                                                               charge+jnrC+0,charge+jnrD+0);
 330             vdwjidx0A        = 2*vdwtype[jnrA+0];
 331             vdwjidx0B        = 2*vdwtype[jnrB+0];
 332             vdwjidx0C        = 2*vdwtype[jnrC+0];
 333             vdwjidx0D        = 2*vdwtype[jnrD+0];
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             if (gmx_mm_any_lt(rsq00,rcutoff2))
 340             {
 341
 342             r00              = _mm_mul_ps(rsq00,rinv00);
 343             r00              = _mm_andnot_ps(dummy_mask,r00);
 344
 345             /* Compute parameters for interactions between i and j atoms */
 346             qq00             = _mm_mul_ps(iq0,jq0);
 347             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 348                                          vdwparam+vdwioffset0+vdwjidx0B,
 349                                          vdwparam+vdwioffset0+vdwjidx0C,
 350                                          vdwparam+vdwioffset0+vdwjidx0D,
 351                                          &c6_00,&c12_00);
 352
 353             /* EWALD ELECTROSTATICS */
 354
 355             /* Analytical PME correction */
 356             zeta2            = _mm_mul_ps(beta2,rsq00);
 357             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 358             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 359             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 360             felec            = _mm_mul_ps(qq00,felec);
 361             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 362             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 363             velec            = _mm_mul_ps(qq00,velec);
 364
 365             /* LENNARD-JONES DISPERSION/REPULSION */
 366
 367             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 368             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 369             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 370             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 371                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 372             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 373
 374             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velec            = _mm_and_ps(velec,cutoff_mask);
 378             velec            = _mm_andnot_ps(dummy_mask,velec);
 379             velecsum         = _mm_add_ps(velecsum,velec);
 380             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 381             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 382             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 383
 384             fscal            = _mm_add_ps(felec,fvdw);
 385
 386             fscal            = _mm_and_ps(fscal,cutoff_mask);
 387
 388             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 389
 390              /* Update vectorial force */
 391             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 392             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 393             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 394
 395             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 396             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 397             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 398             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 399             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 400                                                    _mm_mul_ps(dx00,fscal),
 401                                                    _mm_mul_ps(dy00,fscal),
 402                                                    _mm_mul_ps(dz00,fscal));
 403
 404             }
 405
 406             /* Inner loop uses 52 flops */
 407         }
 408
 409         /* End of innermost loop */
 410
 411         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 412                                               f+i_coord_offset,fshift+i_shift_offset);
 413
 414         ggid                        = gid[iidx];
 415         /* Update potential energies */
 416         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 417         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 418
 419         /* Increment number of inner iterations */
 420         inneriter                  += j_index_end - j_index_start;
 421
 422         /* Outer loop uses 9 flops */
 423     }
 424
 425     /* Increment number of outer iterations */
 426     outeriter        += nri;
 427
 428     /* Update outer/inner flops */
 429
 430     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*52);
 431 }
 432 /*
 433  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 434  * Electrostatics interaction: Ewald
 435  * VdW interaction:            LennardJones
 436  * Geometry:                   Particle-Particle
 437  * Calculate force/pot:        Force
 438  */
 439 void
 440 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 441                     (t_nblist                    * gmx_restrict       nlist,
 442                      rvec                        * gmx_restrict          xx,
 443                      rvec                        * gmx_restrict          ff,
 444                      struct t_forcerec           * gmx_restrict          fr,
 445                      t_mdatoms                   * gmx_restrict     mdatoms,
 446                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 447                      t_nrnb                      * gmx_restrict        nrnb)
 448 {
 449     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 450      * just 0 for non-waters.
 451      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 452      * jnr indices corresponding to data put in the four positions in the SIMD register.
 453      */
 454     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 455     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 456     int              jnrA,jnrB,jnrC,jnrD;
 457     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 458     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 459     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 460     real             rcutoff_scalar;
 461     real             *shiftvec,*fshift,*x,*f;
 462     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 463     real             scratch[4*DIM];
 464     __m128           fscal,rcutoff,rcutoff2,jidxall;
 465     int              vdwioffset0;
 466     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 467     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 468     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 469     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 470     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 471     real             *charge;
 472     int              nvdwtype;
 473     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 474     int              *vdwtype;
 475     real             *vdwparam;
 476     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 477     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 478     __m128i          ewitab;
 479     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 480     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 481     real             *ewtab;
 482     __m128           dummy_mask,cutoff_mask;
 483     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 484     __m128           one     = _mm_set1_ps(1.0);
 485     __m128           two     = _mm_set1_ps(2.0);
 486     x                = xx[0];
 487     f                = ff[0];
 488
 489     nri              = nlist->nri;
 490     iinr             = nlist->iinr;
 491     jindex           = nlist->jindex;
 492     jjnr             = nlist->jjnr;
 493     shiftidx         = nlist->shift;
 494     gid              = nlist->gid;
 495     shiftvec         = fr->shift_vec[0];
 496     fshift           = fr->fshift[0];
 497     facel            = _mm_set1_ps(fr->ic->epsfac);
 498     charge           = mdatoms->chargeA;
 499     nvdwtype         = fr->ntype;
 500     vdwparam         = fr->nbfp;
 501     vdwtype          = mdatoms->typeA;
 502
 503     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 504     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 505     beta2            = _mm_mul_ps(beta,beta);
 506     beta3            = _mm_mul_ps(beta,beta2);
 507     ewtab            = fr->ic->tabq_coul_F;
 508     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 509     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 510
 511     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 512     rcutoff_scalar   = fr->ic->rcoulomb;
 513     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 514     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 515
 516     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 517     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 518
 519     /* Avoid stupid compiler warnings */
 520     jnrA = jnrB = jnrC = jnrD = 0;
 521     j_coord_offsetA = 0;
 522     j_coord_offsetB = 0;
 523     j_coord_offsetC = 0;
 524     j_coord_offsetD = 0;
 525
 526     outeriter        = 0;
 527     inneriter        = 0;
 528
 529     for(iidx=0;iidx<4*DIM;iidx++)
 530     {
 531         scratch[iidx] = 0.0;
 532     }
 533
 534     /* Start outer loop over neighborlists */
 535     for(iidx=0; iidx<nri; iidx++)
 536     {
 537         /* Load shift vector for this list */
 538         i_shift_offset   = DIM*shiftidx[iidx];
 539
 540         /* Load limits for loop over neighbors */
 541         j_index_start    = jindex[iidx];
 542         j_index_end      = jindex[iidx+1];
 543
 544         /* Get outer coordinate index */
 545         inr              = iinr[iidx];
 546         i_coord_offset   = DIM*inr;
 547
 548         /* Load i particle coords and add shift vector */
 549         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 550
 551         fix0             = _mm_setzero_ps();
 552         fiy0             = _mm_setzero_ps();
 553         fiz0             = _mm_setzero_ps();
 554
 555         /* Load parameters for i particles */
 556         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 557         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 558
 559         /* Start inner kernel loop */
 560         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 561         {
 562
 563             /* Get j neighbor index, and coordinate index */
 564             jnrA             = jjnr[jidx];
 565             jnrB             = jjnr[jidx+1];
 566             jnrC             = jjnr[jidx+2];
 567             jnrD             = jjnr[jidx+3];
 568             j_coord_offsetA  = DIM*jnrA;
 569             j_coord_offsetB  = DIM*jnrB;
 570             j_coord_offsetC  = DIM*jnrC;
 571             j_coord_offsetD  = DIM*jnrD;
 572
 573             /* load j atom coordinates */
 574             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 575                                               x+j_coord_offsetC,x+j_coord_offsetD,
 576                                               &jx0,&jy0,&jz0);
 577
 578             /* Calculate displacement vector */
 579             dx00             = _mm_sub_ps(ix0,jx0);
 580             dy00             = _mm_sub_ps(iy0,jy0);
 581             dz00             = _mm_sub_ps(iz0,jz0);
 582
 583             /* Calculate squared distance and things based on it */
 584             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 585
 586             rinv00           = avx128fma_invsqrt_f(rsq00);
 587
 588             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 589
 590             /* Load parameters for j particles */
 591             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 592                                                               charge+jnrC+0,charge+jnrD+0);
 593             vdwjidx0A        = 2*vdwtype[jnrA+0];
 594             vdwjidx0B        = 2*vdwtype[jnrB+0];
 595             vdwjidx0C        = 2*vdwtype[jnrC+0];
 596             vdwjidx0D        = 2*vdwtype[jnrD+0];
 597
 598             /**************************
 599              * CALCULATE INTERACTIONS *
 600              **************************/
 601
 602             if (gmx_mm_any_lt(rsq00,rcutoff2))
 603             {
 604
 605             r00              = _mm_mul_ps(rsq00,rinv00);
 606
 607             /* Compute parameters for interactions between i and j atoms */
 608             qq00             = _mm_mul_ps(iq0,jq0);
 609             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 610                                          vdwparam+vdwioffset0+vdwjidx0B,
 611                                          vdwparam+vdwioffset0+vdwjidx0C,
 612                                          vdwparam+vdwioffset0+vdwjidx0D,
 613                                          &c6_00,&c12_00);
 614
 615             /* EWALD ELECTROSTATICS */
 616
 617             /* Analytical PME correction */
 618             zeta2            = _mm_mul_ps(beta2,rsq00);
 619             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 620             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 621             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 622             felec            = _mm_mul_ps(qq00,felec);
 623
 624             /* LENNARD-JONES DISPERSION/REPULSION */
 625
 626             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 627             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 628
 629             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 630
 631             fscal            = _mm_add_ps(felec,fvdw);
 632
 633             fscal            = _mm_and_ps(fscal,cutoff_mask);
 634
 635              /* Update vectorial force */
 636             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 637             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 638             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 639
 640             fjptrA             = f+j_coord_offsetA;
 641             fjptrB             = f+j_coord_offsetB;
 642             fjptrC             = f+j_coord_offsetC;
 643             fjptrD             = f+j_coord_offsetD;
 644             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 645                                                    _mm_mul_ps(dx00,fscal),
 646                                                    _mm_mul_ps(dy00,fscal),
 647                                                    _mm_mul_ps(dz00,fscal));
 648
 649             }
 650
 651             /* Inner loop uses 38 flops */
 652         }
 653
 654         if(jidx<j_index_end)
 655         {
 656
 657             /* Get j neighbor index, and coordinate index */
 658             jnrlistA         = jjnr[jidx];
 659             jnrlistB         = jjnr[jidx+1];
 660             jnrlistC         = jjnr[jidx+2];
 661             jnrlistD         = jjnr[jidx+3];
 662             /* Sign of each element will be negative for non-real atoms.
 663              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 664              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 665              */
 666             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 667             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 668             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 669             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 670             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 671             j_coord_offsetA  = DIM*jnrA;
 672             j_coord_offsetB  = DIM*jnrB;
 673             j_coord_offsetC  = DIM*jnrC;
 674             j_coord_offsetD  = DIM*jnrD;
 675
 676             /* load j atom coordinates */
 677             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 678                                               x+j_coord_offsetC,x+j_coord_offsetD,
 679                                               &jx0,&jy0,&jz0);
 680
 681             /* Calculate displacement vector */
 682             dx00             = _mm_sub_ps(ix0,jx0);
 683             dy00             = _mm_sub_ps(iy0,jy0);
 684             dz00             = _mm_sub_ps(iz0,jz0);
 685
 686             /* Calculate squared distance and things based on it */
 687             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 688
 689             rinv00           = avx128fma_invsqrt_f(rsq00);
 690
 691             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 692
 693             /* Load parameters for j particles */
 694             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 695                                                               charge+jnrC+0,charge+jnrD+0);
 696             vdwjidx0A        = 2*vdwtype[jnrA+0];
 697             vdwjidx0B        = 2*vdwtype[jnrB+0];
 698             vdwjidx0C        = 2*vdwtype[jnrC+0];
 699             vdwjidx0D        = 2*vdwtype[jnrD+0];
 700
 701             /**************************
 702              * CALCULATE INTERACTIONS *
 703              **************************/
 704
 705             if (gmx_mm_any_lt(rsq00,rcutoff2))
 706             {
 707
 708             r00              = _mm_mul_ps(rsq00,rinv00);
 709             r00              = _mm_andnot_ps(dummy_mask,r00);
 710
 711             /* Compute parameters for interactions between i and j atoms */
 712             qq00             = _mm_mul_ps(iq0,jq0);
 713             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 714                                          vdwparam+vdwioffset0+vdwjidx0B,
 715                                          vdwparam+vdwioffset0+vdwjidx0C,
 716                                          vdwparam+vdwioffset0+vdwjidx0D,
 717                                          &c6_00,&c12_00);
 718
 719             /* EWALD ELECTROSTATICS */
 720
 721             /* Analytical PME correction */
 722             zeta2            = _mm_mul_ps(beta2,rsq00);
 723             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 724             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 725             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 726             felec            = _mm_mul_ps(qq00,felec);
 727
 728             /* LENNARD-JONES DISPERSION/REPULSION */
 729
 730             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 731             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 732
 733             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 734
 735             fscal            = _mm_add_ps(felec,fvdw);
 736
 737             fscal            = _mm_and_ps(fscal,cutoff_mask);
 738
 739             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 740
 741              /* Update vectorial force */
 742             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 743             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 744             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 745
 746             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 747             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 748             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 749             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 750             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 751                                                    _mm_mul_ps(dx00,fscal),
 752                                                    _mm_mul_ps(dy00,fscal),
 753                                                    _mm_mul_ps(dz00,fscal));
 754
 755             }
 756
 757             /* Inner loop uses 39 flops */
 758         }
 759
 760         /* End of innermost loop */
 761
 762         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 763                                               f+i_coord_offset,fshift+i_shift_offset);
 764
 765         /* Increment number of inner iterations */
 766         inneriter                  += j_index_end - j_index_start;
 767
 768         /* Outer loop uses 7 flops */
 769     }
 770
 771     /* Increment number of outer iterations */
 772     outeriter        += nri;
 773
 774     /* Update outer/inner flops */
 775
 776     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*39);
 777 }