src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_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_VdwLJEwSh_GeomW3P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_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              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  92     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  93     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 100     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 101     __m128           c6grid_00;
 102     __m128           c6grid_10;
 103     __m128           c6grid_20;
 104     real             *vdwgridparam;
 105     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 106     __m128           one_half = _mm_set1_ps(0.5);
 107     __m128           minus_one = _mm_set1_ps(-1.0);
 108     __m128i          ewitab;
 109     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 110     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 111     real             *ewtab;
 112     __m128           dummy_mask,cutoff_mask;
 113     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 114     __m128           one     = _mm_set1_ps(1.0);
 115     __m128           two     = _mm_set1_ps(2.0);
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = _mm_set1_ps(fr->ic->epsfac);
 128     charge           = mdatoms->chargeA;
 129     nvdwtype         = fr->ntype;
 130     vdwparam         = fr->nbfp;
 131     vdwtype          = mdatoms->typeA;
 132     vdwgridparam     = fr->ljpme_c6grid;
 133     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 134     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 135     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 136
 137     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 138     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 139     beta2            = _mm_mul_ps(beta,beta);
 140     beta3            = _mm_mul_ps(beta,beta2);
 141     ewtab            = fr->ic->tabq_coul_FDV0;
 142     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 143     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 148     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 149     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 153     rcutoff_scalar   = fr->ic->rcoulomb;
 154     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 155     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 156
 157     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 158     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 159
 160     /* Avoid stupid compiler warnings */
 161     jnrA = jnrB = jnrC = jnrD = 0;
 162     j_coord_offsetA = 0;
 163     j_coord_offsetB = 0;
 164     j_coord_offsetC = 0;
 165     j_coord_offsetD = 0;
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     for(iidx=0;iidx<4*DIM;iidx++)
 171     {
 172         scratch[iidx] = 0.0;
 173     }
 174
 175     /* Start outer loop over neighborlists */
 176     for(iidx=0; iidx<nri; iidx++)
 177     {
 178         /* Load shift vector for this list */
 179         i_shift_offset   = DIM*shiftidx[iidx];
 180
 181         /* Load limits for loop over neighbors */
 182         j_index_start    = jindex[iidx];
 183         j_index_end      = jindex[iidx+1];
 184
 185         /* Get outer coordinate index */
 186         inr              = iinr[iidx];
 187         i_coord_offset   = DIM*inr;
 188
 189         /* Load i particle coords and add shift vector */
 190         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 191                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 192
 193         fix0             = _mm_setzero_ps();
 194         fiy0             = _mm_setzero_ps();
 195         fiz0             = _mm_setzero_ps();
 196         fix1             = _mm_setzero_ps();
 197         fiy1             = _mm_setzero_ps();
 198         fiz1             = _mm_setzero_ps();
 199         fix2             = _mm_setzero_ps();
 200         fiy2             = _mm_setzero_ps();
 201         fiz2             = _mm_setzero_ps();
 202
 203         /* Reset potential sums */
 204         velecsum         = _mm_setzero_ps();
 205         vvdwsum          = _mm_setzero_ps();
 206
 207         /* Start inner kernel loop */
 208         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 209         {
 210
 211             /* Get j neighbor index, and coordinate index */
 212             jnrA             = jjnr[jidx];
 213             jnrB             = jjnr[jidx+1];
 214             jnrC             = jjnr[jidx+2];
 215             jnrD             = jjnr[jidx+3];
 216             j_coord_offsetA  = DIM*jnrA;
 217             j_coord_offsetB  = DIM*jnrB;
 218             j_coord_offsetC  = DIM*jnrC;
 219             j_coord_offsetD  = DIM*jnrD;
 220
 221             /* load j atom coordinates */
 222             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 223                                               x+j_coord_offsetC,x+j_coord_offsetD,
 224                                               &jx0,&jy0,&jz0);
 225
 226             /* Calculate displacement vector */
 227             dx00             = _mm_sub_ps(ix0,jx0);
 228             dy00             = _mm_sub_ps(iy0,jy0);
 229             dz00             = _mm_sub_ps(iz0,jz0);
 230             dx10             = _mm_sub_ps(ix1,jx0);
 231             dy10             = _mm_sub_ps(iy1,jy0);
 232             dz10             = _mm_sub_ps(iz1,jz0);
 233             dx20             = _mm_sub_ps(ix2,jx0);
 234             dy20             = _mm_sub_ps(iy2,jy0);
 235             dz20             = _mm_sub_ps(iz2,jz0);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 239             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 240             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 241
 242             rinv00           = avx128fma_invsqrt_f(rsq00);
 243             rinv10           = avx128fma_invsqrt_f(rsq10);
 244             rinv20           = avx128fma_invsqrt_f(rsq20);
 245
 246             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 247             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 248             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 249
 250             /* Load parameters for j particles */
 251             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 252                                                               charge+jnrC+0,charge+jnrD+0);
 253             vdwjidx0A        = 2*vdwtype[jnrA+0];
 254             vdwjidx0B        = 2*vdwtype[jnrB+0];
 255             vdwjidx0C        = 2*vdwtype[jnrC+0];
 256             vdwjidx0D        = 2*vdwtype[jnrD+0];
 257
 258             fjx0             = _mm_setzero_ps();
 259             fjy0             = _mm_setzero_ps();
 260             fjz0             = _mm_setzero_ps();
 261
 262             /**************************
 263              * CALCULATE INTERACTIONS *
 264              **************************/
 265
 266             if (gmx_mm_any_lt(rsq00,rcutoff2))
 267             {
 268
 269             r00              = _mm_mul_ps(rsq00,rinv00);
 270
 271             /* Compute parameters for interactions between i and j atoms */
 272             qq00             = _mm_mul_ps(iq0,jq0);
 273             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 274                                          vdwparam+vdwioffset0+vdwjidx0B,
 275                                          vdwparam+vdwioffset0+vdwjidx0C,
 276                                          vdwparam+vdwioffset0+vdwjidx0D,
 277                                          &c6_00,&c12_00);
 278
 279             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 280                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 281                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 282                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 283
 284             /* EWALD ELECTROSTATICS */
 285
 286             /* Analytical PME correction */
 287             zeta2            = _mm_mul_ps(beta2,rsq00);
 288             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 289             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 290             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 291             felec            = _mm_mul_ps(qq00,felec);
 292             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 293             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 294             velec            = _mm_mul_ps(qq00,velec);
 295
 296             /* Analytical LJ-PME */
 297             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 298             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 299             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 300             exponent         = avx128fma_exp_f(ewcljrsq);
 301             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 302             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 303             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 304             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 305             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 306             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 307                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 308             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 309             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 310
 311             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 312
 313             /* Update potential sum for this i atom from the interaction with this j atom. */
 314             velec            = _mm_and_ps(velec,cutoff_mask);
 315             velecsum         = _mm_add_ps(velecsum,velec);
 316             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 317             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 318
 319             fscal            = _mm_add_ps(felec,fvdw);
 320
 321             fscal            = _mm_and_ps(fscal,cutoff_mask);
 322
 323              /* Update vectorial force */
 324             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 325             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 326             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 327
 328             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 329             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 330             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 331
 332             }
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             if (gmx_mm_any_lt(rsq10,rcutoff2))
 339             {
 340
 341             r10              = _mm_mul_ps(rsq10,rinv10);
 342
 343             /* Compute parameters for interactions between i and j atoms */
 344             qq10             = _mm_mul_ps(iq1,jq0);
 345
 346             /* EWALD ELECTROSTATICS */
 347
 348             /* Analytical PME correction */
 349             zeta2            = _mm_mul_ps(beta2,rsq10);
 350             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 351             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 352             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 353             felec            = _mm_mul_ps(qq10,felec);
 354             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 355             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 356             velec            = _mm_mul_ps(qq10,velec);
 357
 358             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             velec            = _mm_and_ps(velec,cutoff_mask);
 362             velecsum         = _mm_add_ps(velecsum,velec);
 363
 364             fscal            = felec;
 365
 366             fscal            = _mm_and_ps(fscal,cutoff_mask);
 367
 368              /* Update vectorial force */
 369             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 370             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 371             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 372
 373             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 374             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 375             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 376
 377             }
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             if (gmx_mm_any_lt(rsq20,rcutoff2))
 384             {
 385
 386             r20              = _mm_mul_ps(rsq20,rinv20);
 387
 388             /* Compute parameters for interactions between i and j atoms */
 389             qq20             = _mm_mul_ps(iq2,jq0);
 390
 391             /* EWALD ELECTROSTATICS */
 392
 393             /* Analytical PME correction */
 394             zeta2            = _mm_mul_ps(beta2,rsq20);
 395             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 396             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 397             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 398             felec            = _mm_mul_ps(qq20,felec);
 399             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 400             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 401             velec            = _mm_mul_ps(qq20,velec);
 402
 403             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 404
 405             /* Update potential sum for this i atom from the interaction with this j atom. */
 406             velec            = _mm_and_ps(velec,cutoff_mask);
 407             velecsum         = _mm_add_ps(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             fscal            = _mm_and_ps(fscal,cutoff_mask);
 412
 413              /* Update vectorial force */
 414             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 415             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 416             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 417
 418             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 419             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 420             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 421
 422             }
 423
 424             fjptrA             = f+j_coord_offsetA;
 425             fjptrB             = f+j_coord_offsetB;
 426             fjptrC             = f+j_coord_offsetC;
 427             fjptrD             = f+j_coord_offsetD;
 428
 429             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 430
 431             /* Inner loop uses 129 flops */
 432         }
 433
 434         if(jidx<j_index_end)
 435         {
 436
 437             /* Get j neighbor index, and coordinate index */
 438             jnrlistA         = jjnr[jidx];
 439             jnrlistB         = jjnr[jidx+1];
 440             jnrlistC         = jjnr[jidx+2];
 441             jnrlistD         = jjnr[jidx+3];
 442             /* Sign of each element will be negative for non-real atoms.
 443              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 444              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 445              */
 446             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 447             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 448             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 449             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 450             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 451             j_coord_offsetA  = DIM*jnrA;
 452             j_coord_offsetB  = DIM*jnrB;
 453             j_coord_offsetC  = DIM*jnrC;
 454             j_coord_offsetD  = DIM*jnrD;
 455
 456             /* load j atom coordinates */
 457             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 458                                               x+j_coord_offsetC,x+j_coord_offsetD,
 459                                               &jx0,&jy0,&jz0);
 460
 461             /* Calculate displacement vector */
 462             dx00             = _mm_sub_ps(ix0,jx0);
 463             dy00             = _mm_sub_ps(iy0,jy0);
 464             dz00             = _mm_sub_ps(iz0,jz0);
 465             dx10             = _mm_sub_ps(ix1,jx0);
 466             dy10             = _mm_sub_ps(iy1,jy0);
 467             dz10             = _mm_sub_ps(iz1,jz0);
 468             dx20             = _mm_sub_ps(ix2,jx0);
 469             dy20             = _mm_sub_ps(iy2,jy0);
 470             dz20             = _mm_sub_ps(iz2,jz0);
 471
 472             /* Calculate squared distance and things based on it */
 473             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 474             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 475             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 476
 477             rinv00           = avx128fma_invsqrt_f(rsq00);
 478             rinv10           = avx128fma_invsqrt_f(rsq10);
 479             rinv20           = avx128fma_invsqrt_f(rsq20);
 480
 481             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 482             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 483             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 484
 485             /* Load parameters for j particles */
 486             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 487                                                               charge+jnrC+0,charge+jnrD+0);
 488             vdwjidx0A        = 2*vdwtype[jnrA+0];
 489             vdwjidx0B        = 2*vdwtype[jnrB+0];
 490             vdwjidx0C        = 2*vdwtype[jnrC+0];
 491             vdwjidx0D        = 2*vdwtype[jnrD+0];
 492
 493             fjx0             = _mm_setzero_ps();
 494             fjy0             = _mm_setzero_ps();
 495             fjz0             = _mm_setzero_ps();
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             if (gmx_mm_any_lt(rsq00,rcutoff2))
 502             {
 503
 504             r00              = _mm_mul_ps(rsq00,rinv00);
 505             r00              = _mm_andnot_ps(dummy_mask,r00);
 506
 507             /* Compute parameters for interactions between i and j atoms */
 508             qq00             = _mm_mul_ps(iq0,jq0);
 509             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 510                                          vdwparam+vdwioffset0+vdwjidx0B,
 511                                          vdwparam+vdwioffset0+vdwjidx0C,
 512                                          vdwparam+vdwioffset0+vdwjidx0D,
 513                                          &c6_00,&c12_00);
 514
 515             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 516                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 517                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 518                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 519
 520             /* EWALD ELECTROSTATICS */
 521
 522             /* Analytical PME correction */
 523             zeta2            = _mm_mul_ps(beta2,rsq00);
 524             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 525             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 526             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 527             felec            = _mm_mul_ps(qq00,felec);
 528             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 529             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 530             velec            = _mm_mul_ps(qq00,velec);
 531
 532             /* Analytical LJ-PME */
 533             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 534             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 535             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 536             exponent         = avx128fma_exp_f(ewcljrsq);
 537             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 538             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 539             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 540             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 541             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 542             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 543                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 544             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 545             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 546
 547             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 548
 549             /* Update potential sum for this i atom from the interaction with this j atom. */
 550             velec            = _mm_and_ps(velec,cutoff_mask);
 551             velec            = _mm_andnot_ps(dummy_mask,velec);
 552             velecsum         = _mm_add_ps(velecsum,velec);
 553             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 554             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 555             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 556
 557             fscal            = _mm_add_ps(felec,fvdw);
 558
 559             fscal            = _mm_and_ps(fscal,cutoff_mask);
 560
 561             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 562
 563              /* Update vectorial force */
 564             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 565             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 566             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 567
 568             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 569             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 570             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 571
 572             }
 573
 574             /**************************
 575              * CALCULATE INTERACTIONS *
 576              **************************/
 577
 578             if (gmx_mm_any_lt(rsq10,rcutoff2))
 579             {
 580
 581             r10              = _mm_mul_ps(rsq10,rinv10);
 582             r10              = _mm_andnot_ps(dummy_mask,r10);
 583
 584             /* Compute parameters for interactions between i and j atoms */
 585             qq10             = _mm_mul_ps(iq1,jq0);
 586
 587             /* EWALD ELECTROSTATICS */
 588
 589             /* Analytical PME correction */
 590             zeta2            = _mm_mul_ps(beta2,rsq10);
 591             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 592             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 593             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 594             felec            = _mm_mul_ps(qq10,felec);
 595             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 596             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 597             velec            = _mm_mul_ps(qq10,velec);
 598
 599             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 600
 601             /* Update potential sum for this i atom from the interaction with this j atom. */
 602             velec            = _mm_and_ps(velec,cutoff_mask);
 603             velec            = _mm_andnot_ps(dummy_mask,velec);
 604             velecsum         = _mm_add_ps(velecsum,velec);
 605
 606             fscal            = felec;
 607
 608             fscal            = _mm_and_ps(fscal,cutoff_mask);
 609
 610             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 611
 612              /* Update vectorial force */
 613             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 614             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 615             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 616
 617             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 618             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 619             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 620
 621             }
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             if (gmx_mm_any_lt(rsq20,rcutoff2))
 628             {
 629
 630             r20              = _mm_mul_ps(rsq20,rinv20);
 631             r20              = _mm_andnot_ps(dummy_mask,r20);
 632
 633             /* Compute parameters for interactions between i and j atoms */
 634             qq20             = _mm_mul_ps(iq2,jq0);
 635
 636             /* EWALD ELECTROSTATICS */
 637
 638             /* Analytical PME correction */
 639             zeta2            = _mm_mul_ps(beta2,rsq20);
 640             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 641             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 642             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 643             felec            = _mm_mul_ps(qq20,felec);
 644             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 645             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 646             velec            = _mm_mul_ps(qq20,velec);
 647
 648             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 649
 650             /* Update potential sum for this i atom from the interaction with this j atom. */
 651             velec            = _mm_and_ps(velec,cutoff_mask);
 652             velec            = _mm_andnot_ps(dummy_mask,velec);
 653             velecsum         = _mm_add_ps(velecsum,velec);
 654
 655             fscal            = felec;
 656
 657             fscal            = _mm_and_ps(fscal,cutoff_mask);
 658
 659             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 660
 661              /* Update vectorial force */
 662             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 663             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 664             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 665
 666             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 667             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 668             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 669
 670             }
 671
 672             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 673             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 674             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 675             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 676
 677             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 678
 679             /* Inner loop uses 132 flops */
 680         }
 681
 682         /* End of innermost loop */
 683
 684         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 685                                               f+i_coord_offset,fshift+i_shift_offset);
 686
 687         ggid                        = gid[iidx];
 688         /* Update potential energies */
 689         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 690         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 691
 692         /* Increment number of inner iterations */
 693         inneriter                  += j_index_end - j_index_start;
 694
 695         /* Outer loop uses 20 flops */
 696     }
 697
 698     /* Increment number of outer iterations */
 699     outeriter        += nri;
 700
 701     /* Update outer/inner flops */
 702
 703     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*132);
 704 }
 705 /*
 706  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_F_avx_128_fma_single
 707  * Electrostatics interaction: Ewald
 708  * VdW interaction:            LJEwald
 709  * Geometry:                   Water3-Particle
 710  * Calculate force/pot:        Force
 711  */
 712 void
 713 nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_F_avx_128_fma_single
 714                     (t_nblist                    * gmx_restrict       nlist,
 715                      rvec                        * gmx_restrict          xx,
 716                      rvec                        * gmx_restrict          ff,
 717                      struct t_forcerec           * gmx_restrict          fr,
 718                      t_mdatoms                   * gmx_restrict     mdatoms,
 719                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 720                      t_nrnb                      * gmx_restrict        nrnb)
 721 {
 722     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 723      * just 0 for non-waters.
 724      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 725      * jnr indices corresponding to data put in the four positions in the SIMD register.
 726      */
 727     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 728     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 729     int              jnrA,jnrB,jnrC,jnrD;
 730     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 731     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 732     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 733     real             rcutoff_scalar;
 734     real             *shiftvec,*fshift,*x,*f;
 735     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 736     real             scratch[4*DIM];
 737     __m128           fscal,rcutoff,rcutoff2,jidxall;
 738     int              vdwioffset0;
 739     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 740     int              vdwioffset1;
 741     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 742     int              vdwioffset2;
 743     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 744     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 745     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 746     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 747     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 748     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 749     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 750     real             *charge;
 751     int              nvdwtype;
 752     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 753     int              *vdwtype;
 754     real             *vdwparam;
 755     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 756     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 757     __m128           c6grid_00;
 758     __m128           c6grid_10;
 759     __m128           c6grid_20;
 760     real             *vdwgridparam;
 761     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 762     __m128           one_half = _mm_set1_ps(0.5);
 763     __m128           minus_one = _mm_set1_ps(-1.0);
 764     __m128i          ewitab;
 765     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 766     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 767     real             *ewtab;
 768     __m128           dummy_mask,cutoff_mask;
 769     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 770     __m128           one     = _mm_set1_ps(1.0);
 771     __m128           two     = _mm_set1_ps(2.0);
 772     x                = xx[0];
 773     f                = ff[0];
 774
 775     nri              = nlist->nri;
 776     iinr             = nlist->iinr;
 777     jindex           = nlist->jindex;
 778     jjnr             = nlist->jjnr;
 779     shiftidx         = nlist->shift;
 780     gid              = nlist->gid;
 781     shiftvec         = fr->shift_vec[0];
 782     fshift           = fr->fshift[0];
 783     facel            = _mm_set1_ps(fr->ic->epsfac);
 784     charge           = mdatoms->chargeA;
 785     nvdwtype         = fr->ntype;
 786     vdwparam         = fr->nbfp;
 787     vdwtype          = mdatoms->typeA;
 788     vdwgridparam     = fr->ljpme_c6grid;
 789     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 790     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 791     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 792
 793     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 794     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 795     beta2            = _mm_mul_ps(beta,beta);
 796     beta3            = _mm_mul_ps(beta,beta2);
 797     ewtab            = fr->ic->tabq_coul_F;
 798     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 799     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 800
 801     /* Setup water-specific parameters */
 802     inr              = nlist->iinr[0];
 803     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 804     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 805     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 806     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 807
 808     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 809     rcutoff_scalar   = fr->ic->rcoulomb;
 810     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 811     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 812
 813     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 814     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 815
 816     /* Avoid stupid compiler warnings */
 817     jnrA = jnrB = jnrC = jnrD = 0;
 818     j_coord_offsetA = 0;
 819     j_coord_offsetB = 0;
 820     j_coord_offsetC = 0;
 821     j_coord_offsetD = 0;
 822
 823     outeriter        = 0;
 824     inneriter        = 0;
 825
 826     for(iidx=0;iidx<4*DIM;iidx++)
 827     {
 828         scratch[iidx] = 0.0;
 829     }
 830
 831     /* Start outer loop over neighborlists */
 832     for(iidx=0; iidx<nri; iidx++)
 833     {
 834         /* Load shift vector for this list */
 835         i_shift_offset   = DIM*shiftidx[iidx];
 836
 837         /* Load limits for loop over neighbors */
 838         j_index_start    = jindex[iidx];
 839         j_index_end      = jindex[iidx+1];
 840
 841         /* Get outer coordinate index */
 842         inr              = iinr[iidx];
 843         i_coord_offset   = DIM*inr;
 844
 845         /* Load i particle coords and add shift vector */
 846         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 847                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 848
 849         fix0             = _mm_setzero_ps();
 850         fiy0             = _mm_setzero_ps();
 851         fiz0             = _mm_setzero_ps();
 852         fix1             = _mm_setzero_ps();
 853         fiy1             = _mm_setzero_ps();
 854         fiz1             = _mm_setzero_ps();
 855         fix2             = _mm_setzero_ps();
 856         fiy2             = _mm_setzero_ps();
 857         fiz2             = _mm_setzero_ps();
 858
 859         /* Start inner kernel loop */
 860         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 861         {
 862
 863             /* Get j neighbor index, and coordinate index */
 864             jnrA             = jjnr[jidx];
 865             jnrB             = jjnr[jidx+1];
 866             jnrC             = jjnr[jidx+2];
 867             jnrD             = jjnr[jidx+3];
 868             j_coord_offsetA  = DIM*jnrA;
 869             j_coord_offsetB  = DIM*jnrB;
 870             j_coord_offsetC  = DIM*jnrC;
 871             j_coord_offsetD  = DIM*jnrD;
 872
 873             /* load j atom coordinates */
 874             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 875                                               x+j_coord_offsetC,x+j_coord_offsetD,
 876                                               &jx0,&jy0,&jz0);
 877
 878             /* Calculate displacement vector */
 879             dx00             = _mm_sub_ps(ix0,jx0);
 880             dy00             = _mm_sub_ps(iy0,jy0);
 881             dz00             = _mm_sub_ps(iz0,jz0);
 882             dx10             = _mm_sub_ps(ix1,jx0);
 883             dy10             = _mm_sub_ps(iy1,jy0);
 884             dz10             = _mm_sub_ps(iz1,jz0);
 885             dx20             = _mm_sub_ps(ix2,jx0);
 886             dy20             = _mm_sub_ps(iy2,jy0);
 887             dz20             = _mm_sub_ps(iz2,jz0);
 888
 889             /* Calculate squared distance and things based on it */
 890             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 891             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 892             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 893
 894             rinv00           = avx128fma_invsqrt_f(rsq00);
 895             rinv10           = avx128fma_invsqrt_f(rsq10);
 896             rinv20           = avx128fma_invsqrt_f(rsq20);
 897
 898             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 899             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 900             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 901
 902             /* Load parameters for j particles */
 903             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 904                                                               charge+jnrC+0,charge+jnrD+0);
 905             vdwjidx0A        = 2*vdwtype[jnrA+0];
 906             vdwjidx0B        = 2*vdwtype[jnrB+0];
 907             vdwjidx0C        = 2*vdwtype[jnrC+0];
 908             vdwjidx0D        = 2*vdwtype[jnrD+0];
 909
 910             fjx0             = _mm_setzero_ps();
 911             fjy0             = _mm_setzero_ps();
 912             fjz0             = _mm_setzero_ps();
 913
 914             /**************************
 915              * CALCULATE INTERACTIONS *
 916              **************************/
 917
 918             if (gmx_mm_any_lt(rsq00,rcutoff2))
 919             {
 920
 921             r00              = _mm_mul_ps(rsq00,rinv00);
 922
 923             /* Compute parameters for interactions between i and j atoms */
 924             qq00             = _mm_mul_ps(iq0,jq0);
 925             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 926                                          vdwparam+vdwioffset0+vdwjidx0B,
 927                                          vdwparam+vdwioffset0+vdwjidx0C,
 928                                          vdwparam+vdwioffset0+vdwjidx0D,
 929                                          &c6_00,&c12_00);
 930
 931             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 932                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 933                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 934                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 935
 936             /* EWALD ELECTROSTATICS */
 937
 938             /* Analytical PME correction */
 939             zeta2            = _mm_mul_ps(beta2,rsq00);
 940             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 941             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 942             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 943             felec            = _mm_mul_ps(qq00,felec);
 944
 945             /* Analytical LJ-PME */
 946             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 947             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 948             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 949             exponent         = avx128fma_exp_f(ewcljrsq);
 950             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 951             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 952             /* f6A = 6 * C6grid * (1 - poly) */
 953             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 954             /* f6B = C6grid * exponent * beta^6 */
 955             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 956             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 957             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 958
 959             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 960
 961             fscal            = _mm_add_ps(felec,fvdw);
 962
 963             fscal            = _mm_and_ps(fscal,cutoff_mask);
 964
 965              /* Update vectorial force */
 966             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 967             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 968             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 969
 970             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 971             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 972             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 973
 974             }
 975
 976             /**************************
 977              * CALCULATE INTERACTIONS *
 978              **************************/
 979
 980             if (gmx_mm_any_lt(rsq10,rcutoff2))
 981             {
 982
 983             r10              = _mm_mul_ps(rsq10,rinv10);
 984
 985             /* Compute parameters for interactions between i and j atoms */
 986             qq10             = _mm_mul_ps(iq1,jq0);
 987
 988             /* EWALD ELECTROSTATICS */
 989
 990             /* Analytical PME correction */
 991             zeta2            = _mm_mul_ps(beta2,rsq10);
 992             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 993             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 994             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 995             felec            = _mm_mul_ps(qq10,felec);
 996
 997             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 998
 999             fscal            = felec;
1000
1001             fscal            = _mm_and_ps(fscal,cutoff_mask);
1002
1003              /* Update vectorial force */
1004             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1005             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1006             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1007
1008             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1009             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1010             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1011
1012             }
1013
1014             /**************************
1015              * CALCULATE INTERACTIONS *
1016              **************************/
1017
1018             if (gmx_mm_any_lt(rsq20,rcutoff2))
1019             {
1020
1021             r20              = _mm_mul_ps(rsq20,rinv20);
1022
1023             /* Compute parameters for interactions between i and j atoms */
1024             qq20             = _mm_mul_ps(iq2,jq0);
1025
1026             /* EWALD ELECTROSTATICS */
1027
1028             /* Analytical PME correction */
1029             zeta2            = _mm_mul_ps(beta2,rsq20);
1030             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1031             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1032             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1033             felec            = _mm_mul_ps(qq20,felec);
1034
1035             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1036
1037             fscal            = felec;
1038
1039             fscal            = _mm_and_ps(fscal,cutoff_mask);
1040
1041              /* Update vectorial force */
1042             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1043             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1044             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1045
1046             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1047             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1048             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1049
1050             }
1051
1052             fjptrA             = f+j_coord_offsetA;
1053             fjptrB             = f+j_coord_offsetB;
1054             fjptrC             = f+j_coord_offsetC;
1055             fjptrD             = f+j_coord_offsetD;
1056
1057             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1058
1059             /* Inner loop uses 114 flops */
1060         }
1061
1062         if(jidx<j_index_end)
1063         {
1064
1065             /* Get j neighbor index, and coordinate index */
1066             jnrlistA         = jjnr[jidx];
1067             jnrlistB         = jjnr[jidx+1];
1068             jnrlistC         = jjnr[jidx+2];
1069             jnrlistD         = jjnr[jidx+3];
1070             /* Sign of each element will be negative for non-real atoms.
1071              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1072              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1073              */
1074             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1075             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1076             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1077             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1078             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1079             j_coord_offsetA  = DIM*jnrA;
1080             j_coord_offsetB  = DIM*jnrB;
1081             j_coord_offsetC  = DIM*jnrC;
1082             j_coord_offsetD  = DIM*jnrD;
1083
1084             /* load j atom coordinates */
1085             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1086                                               x+j_coord_offsetC,x+j_coord_offsetD,
1087                                               &jx0,&jy0,&jz0);
1088
1089             /* Calculate displacement vector */
1090             dx00             = _mm_sub_ps(ix0,jx0);
1091             dy00             = _mm_sub_ps(iy0,jy0);
1092             dz00             = _mm_sub_ps(iz0,jz0);
1093             dx10             = _mm_sub_ps(ix1,jx0);
1094             dy10             = _mm_sub_ps(iy1,jy0);
1095             dz10             = _mm_sub_ps(iz1,jz0);
1096             dx20             = _mm_sub_ps(ix2,jx0);
1097             dy20             = _mm_sub_ps(iy2,jy0);
1098             dz20             = _mm_sub_ps(iz2,jz0);
1099
1100             /* Calculate squared distance and things based on it */
1101             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1102             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1103             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1104
1105             rinv00           = avx128fma_invsqrt_f(rsq00);
1106             rinv10           = avx128fma_invsqrt_f(rsq10);
1107             rinv20           = avx128fma_invsqrt_f(rsq20);
1108
1109             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1110             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1111             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1112
1113             /* Load parameters for j particles */
1114             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1115                                                               charge+jnrC+0,charge+jnrD+0);
1116             vdwjidx0A        = 2*vdwtype[jnrA+0];
1117             vdwjidx0B        = 2*vdwtype[jnrB+0];
1118             vdwjidx0C        = 2*vdwtype[jnrC+0];
1119             vdwjidx0D        = 2*vdwtype[jnrD+0];
1120
1121             fjx0             = _mm_setzero_ps();
1122             fjy0             = _mm_setzero_ps();
1123             fjz0             = _mm_setzero_ps();
1124
1125             /**************************
1126              * CALCULATE INTERACTIONS *
1127              **************************/
1128
1129             if (gmx_mm_any_lt(rsq00,rcutoff2))
1130             {
1131
1132             r00              = _mm_mul_ps(rsq00,rinv00);
1133             r00              = _mm_andnot_ps(dummy_mask,r00);
1134
1135             /* Compute parameters for interactions between i and j atoms */
1136             qq00             = _mm_mul_ps(iq0,jq0);
1137             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1138                                          vdwparam+vdwioffset0+vdwjidx0B,
1139                                          vdwparam+vdwioffset0+vdwjidx0C,
1140                                          vdwparam+vdwioffset0+vdwjidx0D,
1141                                          &c6_00,&c12_00);
1142
1143             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
1144                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
1145                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
1146                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
1147
1148             /* EWALD ELECTROSTATICS */
1149
1150             /* Analytical PME correction */
1151             zeta2            = _mm_mul_ps(beta2,rsq00);
1152             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1153             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1154             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1155             felec            = _mm_mul_ps(qq00,felec);
1156
1157             /* Analytical LJ-PME */
1158             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1159             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
1160             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
1161             exponent         = avx128fma_exp_f(ewcljrsq);
1162             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1163             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
1164             /* f6A = 6 * C6grid * (1 - poly) */
1165             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
1166             /* f6B = C6grid * exponent * beta^6 */
1167             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
1168             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1169             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1170
1171             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1172
1173             fscal            = _mm_add_ps(felec,fvdw);
1174
1175             fscal            = _mm_and_ps(fscal,cutoff_mask);
1176
1177             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1178
1179              /* Update vectorial force */
1180             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1181             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1182             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1183
1184             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1185             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1186             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1187
1188             }
1189
1190             /**************************
1191              * CALCULATE INTERACTIONS *
1192              **************************/
1193
1194             if (gmx_mm_any_lt(rsq10,rcutoff2))
1195             {
1196
1197             r10              = _mm_mul_ps(rsq10,rinv10);
1198             r10              = _mm_andnot_ps(dummy_mask,r10);
1199
1200             /* Compute parameters for interactions between i and j atoms */
1201             qq10             = _mm_mul_ps(iq1,jq0);
1202
1203             /* EWALD ELECTROSTATICS */
1204
1205             /* Analytical PME correction */
1206             zeta2            = _mm_mul_ps(beta2,rsq10);
1207             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1208             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1209             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1210             felec            = _mm_mul_ps(qq10,felec);
1211
1212             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1213
1214             fscal            = felec;
1215
1216             fscal            = _mm_and_ps(fscal,cutoff_mask);
1217
1218             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1219
1220              /* Update vectorial force */
1221             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1222             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1223             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1224
1225             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1226             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1227             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1228
1229             }
1230
1231             /**************************
1232              * CALCULATE INTERACTIONS *
1233              **************************/
1234
1235             if (gmx_mm_any_lt(rsq20,rcutoff2))
1236             {
1237
1238             r20              = _mm_mul_ps(rsq20,rinv20);
1239             r20              = _mm_andnot_ps(dummy_mask,r20);
1240
1241             /* Compute parameters for interactions between i and j atoms */
1242             qq20             = _mm_mul_ps(iq2,jq0);
1243
1244             /* EWALD ELECTROSTATICS */
1245
1246             /* Analytical PME correction */
1247             zeta2            = _mm_mul_ps(beta2,rsq20);
1248             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1249             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1250             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1251             felec            = _mm_mul_ps(qq20,felec);
1252
1253             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1254
1255             fscal            = felec;
1256
1257             fscal            = _mm_and_ps(fscal,cutoff_mask);
1258
1259             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1260
1261              /* Update vectorial force */
1262             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1263             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1264             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1265
1266             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1267             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1268             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1269
1270             }
1271
1272             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1273             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1274             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1275             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1276
1277             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1278
1279             /* Inner loop uses 117 flops */
1280         }
1281
1282         /* End of innermost loop */
1283
1284         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1285                                               f+i_coord_offset,fshift+i_shift_offset);
1286
1287         /* Increment number of inner iterations */
1288         inneriter                  += j_index_end - j_index_start;
1289
1290         /* Outer loop uses 18 flops */
1291     }
1292
1293     /* Increment number of outer iterations */
1294     outeriter        += nri;
1295
1296     /* Update outer/inner flops */
1297
1298     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*117);
1299 }