src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwLJEw_GeomW3P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_ElecEw_VdwLJEw_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_ElecEw_VdwLJEw_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     /* Avoid stupid compiler warnings */
 153     jnrA = jnrB = jnrC = jnrD = 0;
 154     j_coord_offsetA = 0;
 155     j_coord_offsetB = 0;
 156     j_coord_offsetC = 0;
 157     j_coord_offsetD = 0;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     for(iidx=0;iidx<4*DIM;iidx++)
 163     {
 164         scratch[iidx] = 0.0;
 165     }
 166
 167     /* Start outer loop over neighborlists */
 168     for(iidx=0; iidx<nri; iidx++)
 169     {
 170         /* Load shift vector for this list */
 171         i_shift_offset   = DIM*shiftidx[iidx];
 172
 173         /* Load limits for loop over neighbors */
 174         j_index_start    = jindex[iidx];
 175         j_index_end      = jindex[iidx+1];
 176
 177         /* Get outer coordinate index */
 178         inr              = iinr[iidx];
 179         i_coord_offset   = DIM*inr;
 180
 181         /* Load i particle coords and add shift vector */
 182         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 183                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 184
 185         fix0             = _mm_setzero_ps();
 186         fiy0             = _mm_setzero_ps();
 187         fiz0             = _mm_setzero_ps();
 188         fix1             = _mm_setzero_ps();
 189         fiy1             = _mm_setzero_ps();
 190         fiz1             = _mm_setzero_ps();
 191         fix2             = _mm_setzero_ps();
 192         fiy2             = _mm_setzero_ps();
 193         fiz2             = _mm_setzero_ps();
 194
 195         /* Reset potential sums */
 196         velecsum         = _mm_setzero_ps();
 197         vvdwsum          = _mm_setzero_ps();
 198
 199         /* Start inner kernel loop */
 200         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 201         {
 202
 203             /* Get j neighbor index, and coordinate index */
 204             jnrA             = jjnr[jidx];
 205             jnrB             = jjnr[jidx+1];
 206             jnrC             = jjnr[jidx+2];
 207             jnrD             = jjnr[jidx+3];
 208             j_coord_offsetA  = DIM*jnrA;
 209             j_coord_offsetB  = DIM*jnrB;
 210             j_coord_offsetC  = DIM*jnrC;
 211             j_coord_offsetD  = DIM*jnrD;
 212
 213             /* load j atom coordinates */
 214             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 215                                               x+j_coord_offsetC,x+j_coord_offsetD,
 216                                               &jx0,&jy0,&jz0);
 217
 218             /* Calculate displacement vector */
 219             dx00             = _mm_sub_ps(ix0,jx0);
 220             dy00             = _mm_sub_ps(iy0,jy0);
 221             dz00             = _mm_sub_ps(iz0,jz0);
 222             dx10             = _mm_sub_ps(ix1,jx0);
 223             dy10             = _mm_sub_ps(iy1,jy0);
 224             dz10             = _mm_sub_ps(iz1,jz0);
 225             dx20             = _mm_sub_ps(ix2,jx0);
 226             dy20             = _mm_sub_ps(iy2,jy0);
 227             dz20             = _mm_sub_ps(iz2,jz0);
 228
 229             /* Calculate squared distance and things based on it */
 230             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 231             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 232             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 233
 234             rinv00           = avx128fma_invsqrt_f(rsq00);
 235             rinv10           = avx128fma_invsqrt_f(rsq10);
 236             rinv20           = avx128fma_invsqrt_f(rsq20);
 237
 238             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 239             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 240             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 241
 242             /* Load parameters for j particles */
 243             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 244                                                               charge+jnrC+0,charge+jnrD+0);
 245             vdwjidx0A        = 2*vdwtype[jnrA+0];
 246             vdwjidx0B        = 2*vdwtype[jnrB+0];
 247             vdwjidx0C        = 2*vdwtype[jnrC+0];
 248             vdwjidx0D        = 2*vdwtype[jnrD+0];
 249
 250             fjx0             = _mm_setzero_ps();
 251             fjy0             = _mm_setzero_ps();
 252             fjz0             = _mm_setzero_ps();
 253
 254             /**************************
 255              * CALCULATE INTERACTIONS *
 256              **************************/
 257
 258             r00              = _mm_mul_ps(rsq00,rinv00);
 259
 260             /* Compute parameters for interactions between i and j atoms */
 261             qq00             = _mm_mul_ps(iq0,jq0);
 262             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 263                                          vdwparam+vdwioffset0+vdwjidx0B,
 264                                          vdwparam+vdwioffset0+vdwjidx0C,
 265                                          vdwparam+vdwioffset0+vdwjidx0D,
 266                                          &c6_00,&c12_00);
 267
 268             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 269                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 270                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 271                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 272
 273             /* EWALD ELECTROSTATICS */
 274
 275             /* Analytical PME correction */
 276             zeta2            = _mm_mul_ps(beta2,rsq00);
 277             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 278             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 279             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 280             felec            = _mm_mul_ps(qq00,felec);
 281             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 282             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 283             velec            = _mm_mul_ps(qq00,velec);
 284
 285             /* Analytical LJ-PME */
 286             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 287             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 288             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 289             exponent         = avx128fma_exp_f(ewcljrsq);
 290             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 291             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 292             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 293             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 294             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 295             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 296             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 297             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);
 298
 299             /* Update potential sum for this i atom from the interaction with this j atom. */
 300             velecsum         = _mm_add_ps(velecsum,velec);
 301             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 302
 303             fscal            = _mm_add_ps(felec,fvdw);
 304
 305              /* Update vectorial force */
 306             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 307             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 308             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 309
 310             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 311             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 312             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 313
 314             /**************************
 315              * CALCULATE INTERACTIONS *
 316              **************************/
 317
 318             r10              = _mm_mul_ps(rsq10,rinv10);
 319
 320             /* Compute parameters for interactions between i and j atoms */
 321             qq10             = _mm_mul_ps(iq1,jq0);
 322
 323             /* EWALD ELECTROSTATICS */
 324
 325             /* Analytical PME correction */
 326             zeta2            = _mm_mul_ps(beta2,rsq10);
 327             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 328             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 329             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 330             felec            = _mm_mul_ps(qq10,felec);
 331             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 332             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 333             velec            = _mm_mul_ps(qq10,velec);
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             velecsum         = _mm_add_ps(velecsum,velec);
 337
 338             fscal            = felec;
 339
 340              /* Update vectorial force */
 341             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 342             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 343             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 344
 345             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 346             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 347             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             r20              = _mm_mul_ps(rsq20,rinv20);
 354
 355             /* Compute parameters for interactions between i and j atoms */
 356             qq20             = _mm_mul_ps(iq2,jq0);
 357
 358             /* EWALD ELECTROSTATICS */
 359
 360             /* Analytical PME correction */
 361             zeta2            = _mm_mul_ps(beta2,rsq20);
 362             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 363             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 364             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 365             felec            = _mm_mul_ps(qq20,felec);
 366             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 367             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 368             velec            = _mm_mul_ps(qq20,velec);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velecsum         = _mm_add_ps(velecsum,velec);
 372
 373             fscal            = felec;
 374
 375              /* Update vectorial force */
 376             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 377             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 378             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 379
 380             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 381             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 382             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 383
 384             fjptrA             = f+j_coord_offsetA;
 385             fjptrB             = f+j_coord_offsetB;
 386             fjptrC             = f+j_coord_offsetC;
 387             fjptrD             = f+j_coord_offsetD;
 388
 389             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 390
 391             /* Inner loop uses 111 flops */
 392         }
 393
 394         if(jidx<j_index_end)
 395         {
 396
 397             /* Get j neighbor index, and coordinate index */
 398             jnrlistA         = jjnr[jidx];
 399             jnrlistB         = jjnr[jidx+1];
 400             jnrlistC         = jjnr[jidx+2];
 401             jnrlistD         = jjnr[jidx+3];
 402             /* Sign of each element will be negative for non-real atoms.
 403              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 404              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 405              */
 406             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 407             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 408             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 409             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 410             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 411             j_coord_offsetA  = DIM*jnrA;
 412             j_coord_offsetB  = DIM*jnrB;
 413             j_coord_offsetC  = DIM*jnrC;
 414             j_coord_offsetD  = DIM*jnrD;
 415
 416             /* load j atom coordinates */
 417             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 418                                               x+j_coord_offsetC,x+j_coord_offsetD,
 419                                               &jx0,&jy0,&jz0);
 420
 421             /* Calculate displacement vector */
 422             dx00             = _mm_sub_ps(ix0,jx0);
 423             dy00             = _mm_sub_ps(iy0,jy0);
 424             dz00             = _mm_sub_ps(iz0,jz0);
 425             dx10             = _mm_sub_ps(ix1,jx0);
 426             dy10             = _mm_sub_ps(iy1,jy0);
 427             dz10             = _mm_sub_ps(iz1,jz0);
 428             dx20             = _mm_sub_ps(ix2,jx0);
 429             dy20             = _mm_sub_ps(iy2,jy0);
 430             dz20             = _mm_sub_ps(iz2,jz0);
 431
 432             /* Calculate squared distance and things based on it */
 433             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 434             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 435             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 436
 437             rinv00           = avx128fma_invsqrt_f(rsq00);
 438             rinv10           = avx128fma_invsqrt_f(rsq10);
 439             rinv20           = avx128fma_invsqrt_f(rsq20);
 440
 441             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 442             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 443             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 444
 445             /* Load parameters for j particles */
 446             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 447                                                               charge+jnrC+0,charge+jnrD+0);
 448             vdwjidx0A        = 2*vdwtype[jnrA+0];
 449             vdwjidx0B        = 2*vdwtype[jnrB+0];
 450             vdwjidx0C        = 2*vdwtype[jnrC+0];
 451             vdwjidx0D        = 2*vdwtype[jnrD+0];
 452
 453             fjx0             = _mm_setzero_ps();
 454             fjy0             = _mm_setzero_ps();
 455             fjz0             = _mm_setzero_ps();
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             r00              = _mm_mul_ps(rsq00,rinv00);
 462             r00              = _mm_andnot_ps(dummy_mask,r00);
 463
 464             /* Compute parameters for interactions between i and j atoms */
 465             qq00             = _mm_mul_ps(iq0,jq0);
 466             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 467                                          vdwparam+vdwioffset0+vdwjidx0B,
 468                                          vdwparam+vdwioffset0+vdwjidx0C,
 469                                          vdwparam+vdwioffset0+vdwjidx0D,
 470                                          &c6_00,&c12_00);
 471
 472             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 473                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 474                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 475                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 476
 477             /* EWALD ELECTROSTATICS */
 478
 479             /* Analytical PME correction */
 480             zeta2            = _mm_mul_ps(beta2,rsq00);
 481             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 482             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 483             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 484             felec            = _mm_mul_ps(qq00,felec);
 485             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 486             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 487             velec            = _mm_mul_ps(qq00,velec);
 488
 489             /* Analytical LJ-PME */
 490             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 491             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 492             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 493             exponent         = avx128fma_exp_f(ewcljrsq);
 494             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 495             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 496             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 497             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 498             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 499             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 500             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 501             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);
 502
 503             /* Update potential sum for this i atom from the interaction with this j atom. */
 504             velec            = _mm_andnot_ps(dummy_mask,velec);
 505             velecsum         = _mm_add_ps(velecsum,velec);
 506             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 507             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 508
 509             fscal            = _mm_add_ps(felec,fvdw);
 510
 511             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 512
 513              /* Update vectorial force */
 514             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 515             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 516             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 517
 518             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 519             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 520             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 521
 522             /**************************
 523              * CALCULATE INTERACTIONS *
 524              **************************/
 525
 526             r10              = _mm_mul_ps(rsq10,rinv10);
 527             r10              = _mm_andnot_ps(dummy_mask,r10);
 528
 529             /* Compute parameters for interactions between i and j atoms */
 530             qq10             = _mm_mul_ps(iq1,jq0);
 531
 532             /* EWALD ELECTROSTATICS */
 533
 534             /* Analytical PME correction */
 535             zeta2            = _mm_mul_ps(beta2,rsq10);
 536             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 537             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 538             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 539             felec            = _mm_mul_ps(qq10,felec);
 540             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 541             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 542             velec            = _mm_mul_ps(qq10,velec);
 543
 544             /* Update potential sum for this i atom from the interaction with this j atom. */
 545             velec            = _mm_andnot_ps(dummy_mask,velec);
 546             velecsum         = _mm_add_ps(velecsum,velec);
 547
 548             fscal            = felec;
 549
 550             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 551
 552              /* Update vectorial force */
 553             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 554             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 555             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 556
 557             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 558             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 559             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 560
 561             /**************************
 562              * CALCULATE INTERACTIONS *
 563              **************************/
 564
 565             r20              = _mm_mul_ps(rsq20,rinv20);
 566             r20              = _mm_andnot_ps(dummy_mask,r20);
 567
 568             /* Compute parameters for interactions between i and j atoms */
 569             qq20             = _mm_mul_ps(iq2,jq0);
 570
 571             /* EWALD ELECTROSTATICS */
 572
 573             /* Analytical PME correction */
 574             zeta2            = _mm_mul_ps(beta2,rsq20);
 575             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 576             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 577             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 578             felec            = _mm_mul_ps(qq20,felec);
 579             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 580             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 581             velec            = _mm_mul_ps(qq20,velec);
 582
 583             /* Update potential sum for this i atom from the interaction with this j atom. */
 584             velec            = _mm_andnot_ps(dummy_mask,velec);
 585             velecsum         = _mm_add_ps(velecsum,velec);
 586
 587             fscal            = felec;
 588
 589             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 590
 591              /* Update vectorial force */
 592             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 593             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 594             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 595
 596             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 597             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 598             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 599
 600             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 601             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 602             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 603             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 604
 605             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 606
 607             /* Inner loop uses 114 flops */
 608         }
 609
 610         /* End of innermost loop */
 611
 612         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 613                                               f+i_coord_offset,fshift+i_shift_offset);
 614
 615         ggid                        = gid[iidx];
 616         /* Update potential energies */
 617         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 618         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 619
 620         /* Increment number of inner iterations */
 621         inneriter                  += j_index_end - j_index_start;
 622
 623         /* Outer loop uses 20 flops */
 624     }
 625
 626     /* Increment number of outer iterations */
 627     outeriter        += nri;
 628
 629     /* Update outer/inner flops */
 630
 631     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*114);
 632 }
 633 /*
 634  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomW3P1_F_avx_128_fma_single
 635  * Electrostatics interaction: Ewald
 636  * VdW interaction:            LJEwald
 637  * Geometry:                   Water3-Particle
 638  * Calculate force/pot:        Force
 639  */
 640 void
 641 nb_kernel_ElecEw_VdwLJEw_GeomW3P1_F_avx_128_fma_single
 642                     (t_nblist                    * gmx_restrict       nlist,
 643                      rvec                        * gmx_restrict          xx,
 644                      rvec                        * gmx_restrict          ff,
 645                      struct t_forcerec           * gmx_restrict          fr,
 646                      t_mdatoms                   * gmx_restrict     mdatoms,
 647                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 648                      t_nrnb                      * gmx_restrict        nrnb)
 649 {
 650     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 651      * just 0 for non-waters.
 652      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 653      * jnr indices corresponding to data put in the four positions in the SIMD register.
 654      */
 655     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 656     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 657     int              jnrA,jnrB,jnrC,jnrD;
 658     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 659     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 660     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 661     real             rcutoff_scalar;
 662     real             *shiftvec,*fshift,*x,*f;
 663     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 664     real             scratch[4*DIM];
 665     __m128           fscal,rcutoff,rcutoff2,jidxall;
 666     int              vdwioffset0;
 667     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 668     int              vdwioffset1;
 669     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 670     int              vdwioffset2;
 671     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 672     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 673     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 674     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 675     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 676     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 677     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 678     real             *charge;
 679     int              nvdwtype;
 680     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 681     int              *vdwtype;
 682     real             *vdwparam;
 683     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 684     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 685     __m128           c6grid_00;
 686     __m128           c6grid_10;
 687     __m128           c6grid_20;
 688     real             *vdwgridparam;
 689     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 690     __m128           one_half = _mm_set1_ps(0.5);
 691     __m128           minus_one = _mm_set1_ps(-1.0);
 692     __m128i          ewitab;
 693     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 694     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 695     real             *ewtab;
 696     __m128           dummy_mask,cutoff_mask;
 697     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 698     __m128           one     = _mm_set1_ps(1.0);
 699     __m128           two     = _mm_set1_ps(2.0);
 700     x                = xx[0];
 701     f                = ff[0];
 702
 703     nri              = nlist->nri;
 704     iinr             = nlist->iinr;
 705     jindex           = nlist->jindex;
 706     jjnr             = nlist->jjnr;
 707     shiftidx         = nlist->shift;
 708     gid              = nlist->gid;
 709     shiftvec         = fr->shift_vec[0];
 710     fshift           = fr->fshift[0];
 711     facel            = _mm_set1_ps(fr->ic->epsfac);
 712     charge           = mdatoms->chargeA;
 713     nvdwtype         = fr->ntype;
 714     vdwparam         = fr->nbfp;
 715     vdwtype          = mdatoms->typeA;
 716     vdwgridparam     = fr->ljpme_c6grid;
 717     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 718     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 719     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 720
 721     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 722     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 723     beta2            = _mm_mul_ps(beta,beta);
 724     beta3            = _mm_mul_ps(beta,beta2);
 725     ewtab            = fr->ic->tabq_coul_F;
 726     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 727     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 728
 729     /* Setup water-specific parameters */
 730     inr              = nlist->iinr[0];
 731     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 732     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 733     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 734     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 735
 736     /* Avoid stupid compiler warnings */
 737     jnrA = jnrB = jnrC = jnrD = 0;
 738     j_coord_offsetA = 0;
 739     j_coord_offsetB = 0;
 740     j_coord_offsetC = 0;
 741     j_coord_offsetD = 0;
 742
 743     outeriter        = 0;
 744     inneriter        = 0;
 745
 746     for(iidx=0;iidx<4*DIM;iidx++)
 747     {
 748         scratch[iidx] = 0.0;
 749     }
 750
 751     /* Start outer loop over neighborlists */
 752     for(iidx=0; iidx<nri; iidx++)
 753     {
 754         /* Load shift vector for this list */
 755         i_shift_offset   = DIM*shiftidx[iidx];
 756
 757         /* Load limits for loop over neighbors */
 758         j_index_start    = jindex[iidx];
 759         j_index_end      = jindex[iidx+1];
 760
 761         /* Get outer coordinate index */
 762         inr              = iinr[iidx];
 763         i_coord_offset   = DIM*inr;
 764
 765         /* Load i particle coords and add shift vector */
 766         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 767                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 768
 769         fix0             = _mm_setzero_ps();
 770         fiy0             = _mm_setzero_ps();
 771         fiz0             = _mm_setzero_ps();
 772         fix1             = _mm_setzero_ps();
 773         fiy1             = _mm_setzero_ps();
 774         fiz1             = _mm_setzero_ps();
 775         fix2             = _mm_setzero_ps();
 776         fiy2             = _mm_setzero_ps();
 777         fiz2             = _mm_setzero_ps();
 778
 779         /* Start inner kernel loop */
 780         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 781         {
 782
 783             /* Get j neighbor index, and coordinate index */
 784             jnrA             = jjnr[jidx];
 785             jnrB             = jjnr[jidx+1];
 786             jnrC             = jjnr[jidx+2];
 787             jnrD             = jjnr[jidx+3];
 788             j_coord_offsetA  = DIM*jnrA;
 789             j_coord_offsetB  = DIM*jnrB;
 790             j_coord_offsetC  = DIM*jnrC;
 791             j_coord_offsetD  = DIM*jnrD;
 792
 793             /* load j atom coordinates */
 794             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 795                                               x+j_coord_offsetC,x+j_coord_offsetD,
 796                                               &jx0,&jy0,&jz0);
 797
 798             /* Calculate displacement vector */
 799             dx00             = _mm_sub_ps(ix0,jx0);
 800             dy00             = _mm_sub_ps(iy0,jy0);
 801             dz00             = _mm_sub_ps(iz0,jz0);
 802             dx10             = _mm_sub_ps(ix1,jx0);
 803             dy10             = _mm_sub_ps(iy1,jy0);
 804             dz10             = _mm_sub_ps(iz1,jz0);
 805             dx20             = _mm_sub_ps(ix2,jx0);
 806             dy20             = _mm_sub_ps(iy2,jy0);
 807             dz20             = _mm_sub_ps(iz2,jz0);
 808
 809             /* Calculate squared distance and things based on it */
 810             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 811             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 812             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 813
 814             rinv00           = avx128fma_invsqrt_f(rsq00);
 815             rinv10           = avx128fma_invsqrt_f(rsq10);
 816             rinv20           = avx128fma_invsqrt_f(rsq20);
 817
 818             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 819             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 820             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 821
 822             /* Load parameters for j particles */
 823             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 824                                                               charge+jnrC+0,charge+jnrD+0);
 825             vdwjidx0A        = 2*vdwtype[jnrA+0];
 826             vdwjidx0B        = 2*vdwtype[jnrB+0];
 827             vdwjidx0C        = 2*vdwtype[jnrC+0];
 828             vdwjidx0D        = 2*vdwtype[jnrD+0];
 829
 830             fjx0             = _mm_setzero_ps();
 831             fjy0             = _mm_setzero_ps();
 832             fjz0             = _mm_setzero_ps();
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             r00              = _mm_mul_ps(rsq00,rinv00);
 839
 840             /* Compute parameters for interactions between i and j atoms */
 841             qq00             = _mm_mul_ps(iq0,jq0);
 842             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 843                                          vdwparam+vdwioffset0+vdwjidx0B,
 844                                          vdwparam+vdwioffset0+vdwjidx0C,
 845                                          vdwparam+vdwioffset0+vdwjidx0D,
 846                                          &c6_00,&c12_00);
 847
 848             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 849                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 850                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 851                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 852
 853             /* EWALD ELECTROSTATICS */
 854
 855             /* Analytical PME correction */
 856             zeta2            = _mm_mul_ps(beta2,rsq00);
 857             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 858             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 859             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 860             felec            = _mm_mul_ps(qq00,felec);
 861
 862             /* Analytical LJ-PME */
 863             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 864             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 865             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 866             exponent         = avx128fma_exp_f(ewcljrsq);
 867             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 868             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 869             /* f6A = 6 * C6grid * (1 - poly) */
 870             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 871             /* f6B = C6grid * exponent * beta^6 */
 872             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 873             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 874             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 875
 876             fscal            = _mm_add_ps(felec,fvdw);
 877
 878              /* Update vectorial force */
 879             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 880             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 881             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 882
 883             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 884             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 885             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 886
 887             /**************************
 888              * CALCULATE INTERACTIONS *
 889              **************************/
 890
 891             r10              = _mm_mul_ps(rsq10,rinv10);
 892
 893             /* Compute parameters for interactions between i and j atoms */
 894             qq10             = _mm_mul_ps(iq1,jq0);
 895
 896             /* EWALD ELECTROSTATICS */
 897
 898             /* Analytical PME correction */
 899             zeta2            = _mm_mul_ps(beta2,rsq10);
 900             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 901             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 902             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 903             felec            = _mm_mul_ps(qq10,felec);
 904
 905             fscal            = felec;
 906
 907              /* Update vectorial force */
 908             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 909             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 910             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 911
 912             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 913             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 914             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 915
 916             /**************************
 917              * CALCULATE INTERACTIONS *
 918              **************************/
 919
 920             r20              = _mm_mul_ps(rsq20,rinv20);
 921
 922             /* Compute parameters for interactions between i and j atoms */
 923             qq20             = _mm_mul_ps(iq2,jq0);
 924
 925             /* EWALD ELECTROSTATICS */
 926
 927             /* Analytical PME correction */
 928             zeta2            = _mm_mul_ps(beta2,rsq20);
 929             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 930             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 931             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 932             felec            = _mm_mul_ps(qq20,felec);
 933
 934             fscal            = felec;
 935
 936              /* Update vectorial force */
 937             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 938             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 939             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 940
 941             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 942             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 943             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 944
 945             fjptrA             = f+j_coord_offsetA;
 946             fjptrB             = f+j_coord_offsetB;
 947             fjptrC             = f+j_coord_offsetC;
 948             fjptrD             = f+j_coord_offsetD;
 949
 950             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 951
 952             /* Inner loop uses 105 flops */
 953         }
 954
 955         if(jidx<j_index_end)
 956         {
 957
 958             /* Get j neighbor index, and coordinate index */
 959             jnrlistA         = jjnr[jidx];
 960             jnrlistB         = jjnr[jidx+1];
 961             jnrlistC         = jjnr[jidx+2];
 962             jnrlistD         = jjnr[jidx+3];
 963             /* Sign of each element will be negative for non-real atoms.
 964              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 965              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 966              */
 967             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 968             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 969             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 970             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 971             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 972             j_coord_offsetA  = DIM*jnrA;
 973             j_coord_offsetB  = DIM*jnrB;
 974             j_coord_offsetC  = DIM*jnrC;
 975             j_coord_offsetD  = DIM*jnrD;
 976
 977             /* load j atom coordinates */
 978             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 979                                               x+j_coord_offsetC,x+j_coord_offsetD,
 980                                               &jx0,&jy0,&jz0);
 981
 982             /* Calculate displacement vector */
 983             dx00             = _mm_sub_ps(ix0,jx0);
 984             dy00             = _mm_sub_ps(iy0,jy0);
 985             dz00             = _mm_sub_ps(iz0,jz0);
 986             dx10             = _mm_sub_ps(ix1,jx0);
 987             dy10             = _mm_sub_ps(iy1,jy0);
 988             dz10             = _mm_sub_ps(iz1,jz0);
 989             dx20             = _mm_sub_ps(ix2,jx0);
 990             dy20             = _mm_sub_ps(iy2,jy0);
 991             dz20             = _mm_sub_ps(iz2,jz0);
 992
 993             /* Calculate squared distance and things based on it */
 994             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 995             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 996             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 997
 998             rinv00           = avx128fma_invsqrt_f(rsq00);
 999             rinv10           = avx128fma_invsqrt_f(rsq10);
1000             rinv20           = avx128fma_invsqrt_f(rsq20);
1001
1002             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1003             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1004             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1005
1006             /* Load parameters for j particles */
1007             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1008                                                               charge+jnrC+0,charge+jnrD+0);
1009             vdwjidx0A        = 2*vdwtype[jnrA+0];
1010             vdwjidx0B        = 2*vdwtype[jnrB+0];
1011             vdwjidx0C        = 2*vdwtype[jnrC+0];
1012             vdwjidx0D        = 2*vdwtype[jnrD+0];
1013
1014             fjx0             = _mm_setzero_ps();
1015             fjy0             = _mm_setzero_ps();
1016             fjz0             = _mm_setzero_ps();
1017
1018             /**************************
1019              * CALCULATE INTERACTIONS *
1020              **************************/
1021
1022             r00              = _mm_mul_ps(rsq00,rinv00);
1023             r00              = _mm_andnot_ps(dummy_mask,r00);
1024
1025             /* Compute parameters for interactions between i and j atoms */
1026             qq00             = _mm_mul_ps(iq0,jq0);
1027             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1028                                          vdwparam+vdwioffset0+vdwjidx0B,
1029                                          vdwparam+vdwioffset0+vdwjidx0C,
1030                                          vdwparam+vdwioffset0+vdwjidx0D,
1031                                          &c6_00,&c12_00);
1032
1033             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
1034                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
1035                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
1036                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
1037
1038             /* EWALD ELECTROSTATICS */
1039
1040             /* Analytical PME correction */
1041             zeta2            = _mm_mul_ps(beta2,rsq00);
1042             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1043             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1044             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1045             felec            = _mm_mul_ps(qq00,felec);
1046
1047             /* Analytical LJ-PME */
1048             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1049             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
1050             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
1051             exponent         = avx128fma_exp_f(ewcljrsq);
1052             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1053             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
1054             /* f6A = 6 * C6grid * (1 - poly) */
1055             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
1056             /* f6B = C6grid * exponent * beta^6 */
1057             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
1058             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1059             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1060
1061             fscal            = _mm_add_ps(felec,fvdw);
1062
1063             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1064
1065              /* Update vectorial force */
1066             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1067             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1068             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1069
1070             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1071             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1072             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1073
1074             /**************************
1075              * CALCULATE INTERACTIONS *
1076              **************************/
1077
1078             r10              = _mm_mul_ps(rsq10,rinv10);
1079             r10              = _mm_andnot_ps(dummy_mask,r10);
1080
1081             /* Compute parameters for interactions between i and j atoms */
1082             qq10             = _mm_mul_ps(iq1,jq0);
1083
1084             /* EWALD ELECTROSTATICS */
1085
1086             /* Analytical PME correction */
1087             zeta2            = _mm_mul_ps(beta2,rsq10);
1088             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1089             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1090             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1091             felec            = _mm_mul_ps(qq10,felec);
1092
1093             fscal            = felec;
1094
1095             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1096
1097              /* Update vectorial force */
1098             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1099             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1100             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1101
1102             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1103             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1104             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1105
1106             /**************************
1107              * CALCULATE INTERACTIONS *
1108              **************************/
1109
1110             r20              = _mm_mul_ps(rsq20,rinv20);
1111             r20              = _mm_andnot_ps(dummy_mask,r20);
1112
1113             /* Compute parameters for interactions between i and j atoms */
1114             qq20             = _mm_mul_ps(iq2,jq0);
1115
1116             /* EWALD ELECTROSTATICS */
1117
1118             /* Analytical PME correction */
1119             zeta2            = _mm_mul_ps(beta2,rsq20);
1120             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1121             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1122             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1123             felec            = _mm_mul_ps(qq20,felec);
1124
1125             fscal            = felec;
1126
1127             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1128
1129              /* Update vectorial force */
1130             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1131             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1132             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1133
1134             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1135             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1136             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1137
1138             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1139             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1140             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1141             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1142
1143             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1144
1145             /* Inner loop uses 108 flops */
1146         }
1147
1148         /* End of innermost loop */
1149
1150         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1151                                               f+i_coord_offset,fshift+i_shift_offset);
1152
1153         /* Increment number of inner iterations */
1154         inneriter                  += j_index_end - j_index_start;
1155
1156         /* Outer loop uses 18 flops */
1157     }
1158
1159     /* Increment number of outer iterations */
1160     outeriter        += nri;
1161
1162     /* Update outer/inner flops */
1163
1164     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*108);
1165 }