src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwLJEw_GeomW4P1_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_GeomW4P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJEw_GeomW4P1_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              vdwioffset3;
  89     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  91     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  94     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  95     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  96     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     int              nvdwtype;
  99     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 100     int              *vdwtype;
 101     real             *vdwparam;
 102     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 103     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 104     __m128           c6grid_00;
 105     __m128           c6grid_10;
 106     __m128           c6grid_20;
 107     __m128           c6grid_30;
 108     real             *vdwgridparam;
 109     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 110     __m128           one_half = _mm_set1_ps(0.5);
 111     __m128           minus_one = _mm_set1_ps(-1.0);
 112     __m128i          ewitab;
 113     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 114     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 115     real             *ewtab;
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->ic->epsfac);
 132     charge           = mdatoms->chargeA;
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136     vdwgridparam     = fr->ljpme_c6grid;
 137     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 138     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 139     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 140
 141     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 142     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 143     beta2            = _mm_mul_ps(beta,beta);
 144     beta3            = _mm_mul_ps(beta,beta2);
 145     ewtab            = fr->ic->tabq_coul_FDV0;
 146     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 147     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 148
 149     /* Setup water-specific parameters */
 150     inr              = nlist->iinr[0];
 151     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 152     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 153     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 154     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 155
 156     /* Avoid stupid compiler warnings */
 157     jnrA = jnrB = jnrC = jnrD = 0;
 158     j_coord_offsetA = 0;
 159     j_coord_offsetB = 0;
 160     j_coord_offsetC = 0;
 161     j_coord_offsetD = 0;
 162
 163     outeriter        = 0;
 164     inneriter        = 0;
 165
 166     for(iidx=0;iidx<4*DIM;iidx++)
 167     {
 168         scratch[iidx] = 0.0;
 169     }
 170
 171     /* Start outer loop over neighborlists */
 172     for(iidx=0; iidx<nri; iidx++)
 173     {
 174         /* Load shift vector for this list */
 175         i_shift_offset   = DIM*shiftidx[iidx];
 176
 177         /* Load limits for loop over neighbors */
 178         j_index_start    = jindex[iidx];
 179         j_index_end      = jindex[iidx+1];
 180
 181         /* Get outer coordinate index */
 182         inr              = iinr[iidx];
 183         i_coord_offset   = DIM*inr;
 184
 185         /* Load i particle coords and add shift vector */
 186         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 187                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 188
 189         fix0             = _mm_setzero_ps();
 190         fiy0             = _mm_setzero_ps();
 191         fiz0             = _mm_setzero_ps();
 192         fix1             = _mm_setzero_ps();
 193         fiy1             = _mm_setzero_ps();
 194         fiz1             = _mm_setzero_ps();
 195         fix2             = _mm_setzero_ps();
 196         fiy2             = _mm_setzero_ps();
 197         fiz2             = _mm_setzero_ps();
 198         fix3             = _mm_setzero_ps();
 199         fiy3             = _mm_setzero_ps();
 200         fiz3             = _mm_setzero_ps();
 201
 202         /* Reset potential sums */
 203         velecsum         = _mm_setzero_ps();
 204         vvdwsum          = _mm_setzero_ps();
 205
 206         /* Start inner kernel loop */
 207         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 208         {
 209
 210             /* Get j neighbor index, and coordinate index */
 211             jnrA             = jjnr[jidx];
 212             jnrB             = jjnr[jidx+1];
 213             jnrC             = jjnr[jidx+2];
 214             jnrD             = jjnr[jidx+3];
 215             j_coord_offsetA  = DIM*jnrA;
 216             j_coord_offsetB  = DIM*jnrB;
 217             j_coord_offsetC  = DIM*jnrC;
 218             j_coord_offsetD  = DIM*jnrD;
 219
 220             /* load j atom coordinates */
 221             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 222                                               x+j_coord_offsetC,x+j_coord_offsetD,
 223                                               &jx0,&jy0,&jz0);
 224
 225             /* Calculate displacement vector */
 226             dx00             = _mm_sub_ps(ix0,jx0);
 227             dy00             = _mm_sub_ps(iy0,jy0);
 228             dz00             = _mm_sub_ps(iz0,jz0);
 229             dx10             = _mm_sub_ps(ix1,jx0);
 230             dy10             = _mm_sub_ps(iy1,jy0);
 231             dz10             = _mm_sub_ps(iz1,jz0);
 232             dx20             = _mm_sub_ps(ix2,jx0);
 233             dy20             = _mm_sub_ps(iy2,jy0);
 234             dz20             = _mm_sub_ps(iz2,jz0);
 235             dx30             = _mm_sub_ps(ix3,jx0);
 236             dy30             = _mm_sub_ps(iy3,jy0);
 237             dz30             = _mm_sub_ps(iz3,jz0);
 238
 239             /* Calculate squared distance and things based on it */
 240             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 241             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 242             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 243             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 244
 245             rinv00           = avx128fma_invsqrt_f(rsq00);
 246             rinv10           = avx128fma_invsqrt_f(rsq10);
 247             rinv20           = avx128fma_invsqrt_f(rsq20);
 248             rinv30           = avx128fma_invsqrt_f(rsq30);
 249
 250             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 251             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 252             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 253             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 254
 255             /* Load parameters for j particles */
 256             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 257                                                               charge+jnrC+0,charge+jnrD+0);
 258             vdwjidx0A        = 2*vdwtype[jnrA+0];
 259             vdwjidx0B        = 2*vdwtype[jnrB+0];
 260             vdwjidx0C        = 2*vdwtype[jnrC+0];
 261             vdwjidx0D        = 2*vdwtype[jnrD+0];
 262
 263             fjx0             = _mm_setzero_ps();
 264             fjy0             = _mm_setzero_ps();
 265             fjz0             = _mm_setzero_ps();
 266
 267             /**************************
 268              * CALCULATE INTERACTIONS *
 269              **************************/
 270
 271             r00              = _mm_mul_ps(rsq00,rinv00);
 272
 273             /* Compute parameters for interactions between i and j atoms */
 274             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 275                                          vdwparam+vdwioffset0+vdwjidx0B,
 276                                          vdwparam+vdwioffset0+vdwjidx0C,
 277                                          vdwparam+vdwioffset0+vdwjidx0D,
 278                                          &c6_00,&c12_00);
 279
 280             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 281                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 282                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 283                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 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             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 301
 302             fscal            = fvdw;
 303
 304              /* Update vectorial force */
 305             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 306             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 307             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 308
 309             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 310             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 311             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             r10              = _mm_mul_ps(rsq10,rinv10);
 318
 319             /* Compute parameters for interactions between i and j atoms */
 320             qq10             = _mm_mul_ps(iq1,jq0);
 321
 322             /* EWALD ELECTROSTATICS */
 323
 324             /* Analytical PME correction */
 325             zeta2            = _mm_mul_ps(beta2,rsq10);
 326             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 327             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 328             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 329             felec            = _mm_mul_ps(qq10,felec);
 330             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 331             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 332             velec            = _mm_mul_ps(qq10,velec);
 333
 334             /* Update potential sum for this i atom from the interaction with this j atom. */
 335             velecsum         = _mm_add_ps(velecsum,velec);
 336
 337             fscal            = felec;
 338
 339              /* Update vectorial force */
 340             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 341             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 342             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 343
 344             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 345             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 346             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 347
 348             /**************************
 349              * CALCULATE INTERACTIONS *
 350              **************************/
 351
 352             r20              = _mm_mul_ps(rsq20,rinv20);
 353
 354             /* Compute parameters for interactions between i and j atoms */
 355             qq20             = _mm_mul_ps(iq2,jq0);
 356
 357             /* EWALD ELECTROSTATICS */
 358
 359             /* Analytical PME correction */
 360             zeta2            = _mm_mul_ps(beta2,rsq20);
 361             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 362             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 363             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 364             felec            = _mm_mul_ps(qq20,felec);
 365             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 366             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 367             velec            = _mm_mul_ps(qq20,velec);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velecsum         = _mm_add_ps(velecsum,velec);
 371
 372             fscal            = felec;
 373
 374              /* Update vectorial force */
 375             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 376             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 377             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 378
 379             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 380             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 381             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             r30              = _mm_mul_ps(rsq30,rinv30);
 388
 389             /* Compute parameters for interactions between i and j atoms */
 390             qq30             = _mm_mul_ps(iq3,jq0);
 391
 392             /* EWALD ELECTROSTATICS */
 393
 394             /* Analytical PME correction */
 395             zeta2            = _mm_mul_ps(beta2,rsq30);
 396             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
 397             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 398             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 399             felec            = _mm_mul_ps(qq30,felec);
 400             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 401             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv30);
 402             velec            = _mm_mul_ps(qq30,velec);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velecsum         = _mm_add_ps(velecsum,velec);
 406
 407             fscal            = felec;
 408
 409              /* Update vectorial force */
 410             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 411             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 412             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 413
 414             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 415             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 416             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 417
 418             fjptrA             = f+j_coord_offsetA;
 419             fjptrB             = f+j_coord_offsetB;
 420             fjptrC             = f+j_coord_offsetC;
 421             fjptrD             = f+j_coord_offsetD;
 422
 423             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 424
 425             /* Inner loop uses 137 flops */
 426         }
 427
 428         if(jidx<j_index_end)
 429         {
 430
 431             /* Get j neighbor index, and coordinate index */
 432             jnrlistA         = jjnr[jidx];
 433             jnrlistB         = jjnr[jidx+1];
 434             jnrlistC         = jjnr[jidx+2];
 435             jnrlistD         = jjnr[jidx+3];
 436             /* Sign of each element will be negative for non-real atoms.
 437              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 438              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 439              */
 440             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 441             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 442             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 443             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 444             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 445             j_coord_offsetA  = DIM*jnrA;
 446             j_coord_offsetB  = DIM*jnrB;
 447             j_coord_offsetC  = DIM*jnrC;
 448             j_coord_offsetD  = DIM*jnrD;
 449
 450             /* load j atom coordinates */
 451             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 452                                               x+j_coord_offsetC,x+j_coord_offsetD,
 453                                               &jx0,&jy0,&jz0);
 454
 455             /* Calculate displacement vector */
 456             dx00             = _mm_sub_ps(ix0,jx0);
 457             dy00             = _mm_sub_ps(iy0,jy0);
 458             dz00             = _mm_sub_ps(iz0,jz0);
 459             dx10             = _mm_sub_ps(ix1,jx0);
 460             dy10             = _mm_sub_ps(iy1,jy0);
 461             dz10             = _mm_sub_ps(iz1,jz0);
 462             dx20             = _mm_sub_ps(ix2,jx0);
 463             dy20             = _mm_sub_ps(iy2,jy0);
 464             dz20             = _mm_sub_ps(iz2,jz0);
 465             dx30             = _mm_sub_ps(ix3,jx0);
 466             dy30             = _mm_sub_ps(iy3,jy0);
 467             dz30             = _mm_sub_ps(iz3,jz0);
 468
 469             /* Calculate squared distance and things based on it */
 470             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 471             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 472             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 473             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 474
 475             rinv00           = avx128fma_invsqrt_f(rsq00);
 476             rinv10           = avx128fma_invsqrt_f(rsq10);
 477             rinv20           = avx128fma_invsqrt_f(rsq20);
 478             rinv30           = avx128fma_invsqrt_f(rsq30);
 479
 480             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 481             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 482             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 483             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 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             r00              = _mm_mul_ps(rsq00,rinv00);
 502             r00              = _mm_andnot_ps(dummy_mask,r00);
 503
 504             /* Compute parameters for interactions between i and j atoms */
 505             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 506                                          vdwparam+vdwioffset0+vdwjidx0B,
 507                                          vdwparam+vdwioffset0+vdwjidx0C,
 508                                          vdwparam+vdwioffset0+vdwjidx0D,
 509                                          &c6_00,&c12_00);
 510
 511             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 512                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 513                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 514                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 515
 516             /* Analytical LJ-PME */
 517             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 518             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 519             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 520             exponent         = avx128fma_exp_f(ewcljrsq);
 521             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 522             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 523             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 524             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 525             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 526             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 527             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 528             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);
 529
 530             /* Update potential sum for this i atom from the interaction with this j atom. */
 531             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 532             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 533
 534             fscal            = fvdw;
 535
 536             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 537
 538              /* Update vectorial force */
 539             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 540             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 541             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 542
 543             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 544             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 545             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             r10              = _mm_mul_ps(rsq10,rinv10);
 552             r10              = _mm_andnot_ps(dummy_mask,r10);
 553
 554             /* Compute parameters for interactions between i and j atoms */
 555             qq10             = _mm_mul_ps(iq1,jq0);
 556
 557             /* EWALD ELECTROSTATICS */
 558
 559             /* Analytical PME correction */
 560             zeta2            = _mm_mul_ps(beta2,rsq10);
 561             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 562             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 563             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 564             felec            = _mm_mul_ps(qq10,felec);
 565             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 566             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 567             velec            = _mm_mul_ps(qq10,velec);
 568
 569             /* Update potential sum for this i atom from the interaction with this j atom. */
 570             velec            = _mm_andnot_ps(dummy_mask,velec);
 571             velecsum         = _mm_add_ps(velecsum,velec);
 572
 573             fscal            = felec;
 574
 575             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 576
 577              /* Update vectorial force */
 578             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 579             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 580             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 581
 582             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 583             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 584             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             r20              = _mm_mul_ps(rsq20,rinv20);
 591             r20              = _mm_andnot_ps(dummy_mask,r20);
 592
 593             /* Compute parameters for interactions between i and j atoms */
 594             qq20             = _mm_mul_ps(iq2,jq0);
 595
 596             /* EWALD ELECTROSTATICS */
 597
 598             /* Analytical PME correction */
 599             zeta2            = _mm_mul_ps(beta2,rsq20);
 600             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 601             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 602             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 603             felec            = _mm_mul_ps(qq20,felec);
 604             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 605             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 606             velec            = _mm_mul_ps(qq20,velec);
 607
 608             /* Update potential sum for this i atom from the interaction with this j atom. */
 609             velec            = _mm_andnot_ps(dummy_mask,velec);
 610             velecsum         = _mm_add_ps(velecsum,velec);
 611
 612             fscal            = felec;
 613
 614             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 615
 616              /* Update vectorial force */
 617             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 618             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 619             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 620
 621             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 622             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 623             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 624
 625             /**************************
 626              * CALCULATE INTERACTIONS *
 627              **************************/
 628
 629             r30              = _mm_mul_ps(rsq30,rinv30);
 630             r30              = _mm_andnot_ps(dummy_mask,r30);
 631
 632             /* Compute parameters for interactions between i and j atoms */
 633             qq30             = _mm_mul_ps(iq3,jq0);
 634
 635             /* EWALD ELECTROSTATICS */
 636
 637             /* Analytical PME correction */
 638             zeta2            = _mm_mul_ps(beta2,rsq30);
 639             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
 640             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 641             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 642             felec            = _mm_mul_ps(qq30,felec);
 643             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 644             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv30);
 645             velec            = _mm_mul_ps(qq30,velec);
 646
 647             /* Update potential sum for this i atom from the interaction with this j atom. */
 648             velec            = _mm_andnot_ps(dummy_mask,velec);
 649             velecsum         = _mm_add_ps(velecsum,velec);
 650
 651             fscal            = felec;
 652
 653             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 654
 655              /* Update vectorial force */
 656             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 657             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 658             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 659
 660             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 661             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 662             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 663
 664             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 665             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 666             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 667             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 668
 669             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 670
 671             /* Inner loop uses 141 flops */
 672         }
 673
 674         /* End of innermost loop */
 675
 676         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 677                                               f+i_coord_offset,fshift+i_shift_offset);
 678
 679         ggid                        = gid[iidx];
 680         /* Update potential energies */
 681         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 682         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 683
 684         /* Increment number of inner iterations */
 685         inneriter                  += j_index_end - j_index_start;
 686
 687         /* Outer loop uses 26 flops */
 688     }
 689
 690     /* Increment number of outer iterations */
 691     outeriter        += nri;
 692
 693     /* Update outer/inner flops */
 694
 695     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*141);
 696 }
 697 /*
 698  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomW4P1_F_avx_128_fma_single
 699  * Electrostatics interaction: Ewald
 700  * VdW interaction:            LJEwald
 701  * Geometry:                   Water4-Particle
 702  * Calculate force/pot:        Force
 703  */
 704 void
 705 nb_kernel_ElecEw_VdwLJEw_GeomW4P1_F_avx_128_fma_single
 706                     (t_nblist                    * gmx_restrict       nlist,
 707                      rvec                        * gmx_restrict          xx,
 708                      rvec                        * gmx_restrict          ff,
 709                      struct t_forcerec           * gmx_restrict          fr,
 710                      t_mdatoms                   * gmx_restrict     mdatoms,
 711                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 712                      t_nrnb                      * gmx_restrict        nrnb)
 713 {
 714     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 715      * just 0 for non-waters.
 716      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 717      * jnr indices corresponding to data put in the four positions in the SIMD register.
 718      */
 719     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 720     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 721     int              jnrA,jnrB,jnrC,jnrD;
 722     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 723     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 724     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 725     real             rcutoff_scalar;
 726     real             *shiftvec,*fshift,*x,*f;
 727     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 728     real             scratch[4*DIM];
 729     __m128           fscal,rcutoff,rcutoff2,jidxall;
 730     int              vdwioffset0;
 731     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 732     int              vdwioffset1;
 733     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 734     int              vdwioffset2;
 735     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 736     int              vdwioffset3;
 737     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 738     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 739     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 740     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 741     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 742     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 743     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 744     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 745     real             *charge;
 746     int              nvdwtype;
 747     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 748     int              *vdwtype;
 749     real             *vdwparam;
 750     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 751     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 752     __m128           c6grid_00;
 753     __m128           c6grid_10;
 754     __m128           c6grid_20;
 755     __m128           c6grid_30;
 756     real             *vdwgridparam;
 757     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 758     __m128           one_half = _mm_set1_ps(0.5);
 759     __m128           minus_one = _mm_set1_ps(-1.0);
 760     __m128i          ewitab;
 761     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 762     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 763     real             *ewtab;
 764     __m128           dummy_mask,cutoff_mask;
 765     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 766     __m128           one     = _mm_set1_ps(1.0);
 767     __m128           two     = _mm_set1_ps(2.0);
 768     x                = xx[0];
 769     f                = ff[0];
 770
 771     nri              = nlist->nri;
 772     iinr             = nlist->iinr;
 773     jindex           = nlist->jindex;
 774     jjnr             = nlist->jjnr;
 775     shiftidx         = nlist->shift;
 776     gid              = nlist->gid;
 777     shiftvec         = fr->shift_vec[0];
 778     fshift           = fr->fshift[0];
 779     facel            = _mm_set1_ps(fr->ic->epsfac);
 780     charge           = mdatoms->chargeA;
 781     nvdwtype         = fr->ntype;
 782     vdwparam         = fr->nbfp;
 783     vdwtype          = mdatoms->typeA;
 784     vdwgridparam     = fr->ljpme_c6grid;
 785     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 786     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 787     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 788
 789     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 790     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 791     beta2            = _mm_mul_ps(beta,beta);
 792     beta3            = _mm_mul_ps(beta,beta2);
 793     ewtab            = fr->ic->tabq_coul_F;
 794     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 795     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 796
 797     /* Setup water-specific parameters */
 798     inr              = nlist->iinr[0];
 799     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 800     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 801     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 802     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 803
 804     /* Avoid stupid compiler warnings */
 805     jnrA = jnrB = jnrC = jnrD = 0;
 806     j_coord_offsetA = 0;
 807     j_coord_offsetB = 0;
 808     j_coord_offsetC = 0;
 809     j_coord_offsetD = 0;
 810
 811     outeriter        = 0;
 812     inneriter        = 0;
 813
 814     for(iidx=0;iidx<4*DIM;iidx++)
 815     {
 816         scratch[iidx] = 0.0;
 817     }
 818
 819     /* Start outer loop over neighborlists */
 820     for(iidx=0; iidx<nri; iidx++)
 821     {
 822         /* Load shift vector for this list */
 823         i_shift_offset   = DIM*shiftidx[iidx];
 824
 825         /* Load limits for loop over neighbors */
 826         j_index_start    = jindex[iidx];
 827         j_index_end      = jindex[iidx+1];
 828
 829         /* Get outer coordinate index */
 830         inr              = iinr[iidx];
 831         i_coord_offset   = DIM*inr;
 832
 833         /* Load i particle coords and add shift vector */
 834         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 835                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 836
 837         fix0             = _mm_setzero_ps();
 838         fiy0             = _mm_setzero_ps();
 839         fiz0             = _mm_setzero_ps();
 840         fix1             = _mm_setzero_ps();
 841         fiy1             = _mm_setzero_ps();
 842         fiz1             = _mm_setzero_ps();
 843         fix2             = _mm_setzero_ps();
 844         fiy2             = _mm_setzero_ps();
 845         fiz2             = _mm_setzero_ps();
 846         fix3             = _mm_setzero_ps();
 847         fiy3             = _mm_setzero_ps();
 848         fiz3             = _mm_setzero_ps();
 849
 850         /* Start inner kernel loop */
 851         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 852         {
 853
 854             /* Get j neighbor index, and coordinate index */
 855             jnrA             = jjnr[jidx];
 856             jnrB             = jjnr[jidx+1];
 857             jnrC             = jjnr[jidx+2];
 858             jnrD             = jjnr[jidx+3];
 859             j_coord_offsetA  = DIM*jnrA;
 860             j_coord_offsetB  = DIM*jnrB;
 861             j_coord_offsetC  = DIM*jnrC;
 862             j_coord_offsetD  = DIM*jnrD;
 863
 864             /* load j atom coordinates */
 865             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 866                                               x+j_coord_offsetC,x+j_coord_offsetD,
 867                                               &jx0,&jy0,&jz0);
 868
 869             /* Calculate displacement vector */
 870             dx00             = _mm_sub_ps(ix0,jx0);
 871             dy00             = _mm_sub_ps(iy0,jy0);
 872             dz00             = _mm_sub_ps(iz0,jz0);
 873             dx10             = _mm_sub_ps(ix1,jx0);
 874             dy10             = _mm_sub_ps(iy1,jy0);
 875             dz10             = _mm_sub_ps(iz1,jz0);
 876             dx20             = _mm_sub_ps(ix2,jx0);
 877             dy20             = _mm_sub_ps(iy2,jy0);
 878             dz20             = _mm_sub_ps(iz2,jz0);
 879             dx30             = _mm_sub_ps(ix3,jx0);
 880             dy30             = _mm_sub_ps(iy3,jy0);
 881             dz30             = _mm_sub_ps(iz3,jz0);
 882
 883             /* Calculate squared distance and things based on it */
 884             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 885             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 886             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 887             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 888
 889             rinv00           = avx128fma_invsqrt_f(rsq00);
 890             rinv10           = avx128fma_invsqrt_f(rsq10);
 891             rinv20           = avx128fma_invsqrt_f(rsq20);
 892             rinv30           = avx128fma_invsqrt_f(rsq30);
 893
 894             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 895             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 896             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 897             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 898
 899             /* Load parameters for j particles */
 900             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 901                                                               charge+jnrC+0,charge+jnrD+0);
 902             vdwjidx0A        = 2*vdwtype[jnrA+0];
 903             vdwjidx0B        = 2*vdwtype[jnrB+0];
 904             vdwjidx0C        = 2*vdwtype[jnrC+0];
 905             vdwjidx0D        = 2*vdwtype[jnrD+0];
 906
 907             fjx0             = _mm_setzero_ps();
 908             fjy0             = _mm_setzero_ps();
 909             fjz0             = _mm_setzero_ps();
 910
 911             /**************************
 912              * CALCULATE INTERACTIONS *
 913              **************************/
 914
 915             r00              = _mm_mul_ps(rsq00,rinv00);
 916
 917             /* Compute parameters for interactions between i and j atoms */
 918             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 919                                          vdwparam+vdwioffset0+vdwjidx0B,
 920                                          vdwparam+vdwioffset0+vdwjidx0C,
 921                                          vdwparam+vdwioffset0+vdwjidx0D,
 922                                          &c6_00,&c12_00);
 923
 924             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 925                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 926                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 927                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 928
 929             /* Analytical LJ-PME */
 930             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 931             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 932             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 933             exponent         = avx128fma_exp_f(ewcljrsq);
 934             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 935             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 936             /* f6A = 6 * C6grid * (1 - poly) */
 937             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 938             /* f6B = C6grid * exponent * beta^6 */
 939             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 940             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 941             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 942
 943             fscal            = fvdw;
 944
 945              /* Update vectorial force */
 946             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 947             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 948             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 949
 950             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 951             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 952             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 953
 954             /**************************
 955              * CALCULATE INTERACTIONS *
 956              **************************/
 957
 958             r10              = _mm_mul_ps(rsq10,rinv10);
 959
 960             /* Compute parameters for interactions between i and j atoms */
 961             qq10             = _mm_mul_ps(iq1,jq0);
 962
 963             /* EWALD ELECTROSTATICS */
 964
 965             /* Analytical PME correction */
 966             zeta2            = _mm_mul_ps(beta2,rsq10);
 967             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 968             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 969             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 970             felec            = _mm_mul_ps(qq10,felec);
 971
 972             fscal            = felec;
 973
 974              /* Update vectorial force */
 975             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 976             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 977             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 978
 979             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 980             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 981             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 982
 983             /**************************
 984              * CALCULATE INTERACTIONS *
 985              **************************/
 986
 987             r20              = _mm_mul_ps(rsq20,rinv20);
 988
 989             /* Compute parameters for interactions between i and j atoms */
 990             qq20             = _mm_mul_ps(iq2,jq0);
 991
 992             /* EWALD ELECTROSTATICS */
 993
 994             /* Analytical PME correction */
 995             zeta2            = _mm_mul_ps(beta2,rsq20);
 996             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 997             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 998             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 999             felec            = _mm_mul_ps(qq20,felec);
1000
1001             fscal            = felec;
1002
1003              /* Update vectorial force */
1004             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1005             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1006             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1007
1008             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1009             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1010             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1011
1012             /**************************
1013              * CALCULATE INTERACTIONS *
1014              **************************/
1015
1016             r30              = _mm_mul_ps(rsq30,rinv30);
1017
1018             /* Compute parameters for interactions between i and j atoms */
1019             qq30             = _mm_mul_ps(iq3,jq0);
1020
1021             /* EWALD ELECTROSTATICS */
1022
1023             /* Analytical PME correction */
1024             zeta2            = _mm_mul_ps(beta2,rsq30);
1025             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
1026             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1027             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1028             felec            = _mm_mul_ps(qq30,felec);
1029
1030             fscal            = felec;
1031
1032              /* Update vectorial force */
1033             fix3             = _mm_macc_ps(dx30,fscal,fix3);
1034             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
1035             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
1036
1037             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
1038             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
1039             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
1040
1041             fjptrA             = f+j_coord_offsetA;
1042             fjptrB             = f+j_coord_offsetB;
1043             fjptrC             = f+j_coord_offsetC;
1044             fjptrD             = f+j_coord_offsetD;
1045
1046             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1047
1048             /* Inner loop uses 131 flops */
1049         }
1050
1051         if(jidx<j_index_end)
1052         {
1053
1054             /* Get j neighbor index, and coordinate index */
1055             jnrlistA         = jjnr[jidx];
1056             jnrlistB         = jjnr[jidx+1];
1057             jnrlistC         = jjnr[jidx+2];
1058             jnrlistD         = jjnr[jidx+3];
1059             /* Sign of each element will be negative for non-real atoms.
1060              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1061              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1062              */
1063             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1064             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1065             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1066             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1067             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1068             j_coord_offsetA  = DIM*jnrA;
1069             j_coord_offsetB  = DIM*jnrB;
1070             j_coord_offsetC  = DIM*jnrC;
1071             j_coord_offsetD  = DIM*jnrD;
1072
1073             /* load j atom coordinates */
1074             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1075                                               x+j_coord_offsetC,x+j_coord_offsetD,
1076                                               &jx0,&jy0,&jz0);
1077
1078             /* Calculate displacement vector */
1079             dx00             = _mm_sub_ps(ix0,jx0);
1080             dy00             = _mm_sub_ps(iy0,jy0);
1081             dz00             = _mm_sub_ps(iz0,jz0);
1082             dx10             = _mm_sub_ps(ix1,jx0);
1083             dy10             = _mm_sub_ps(iy1,jy0);
1084             dz10             = _mm_sub_ps(iz1,jz0);
1085             dx20             = _mm_sub_ps(ix2,jx0);
1086             dy20             = _mm_sub_ps(iy2,jy0);
1087             dz20             = _mm_sub_ps(iz2,jz0);
1088             dx30             = _mm_sub_ps(ix3,jx0);
1089             dy30             = _mm_sub_ps(iy3,jy0);
1090             dz30             = _mm_sub_ps(iz3,jz0);
1091
1092             /* Calculate squared distance and things based on it */
1093             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1094             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1095             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1096             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
1097
1098             rinv00           = avx128fma_invsqrt_f(rsq00);
1099             rinv10           = avx128fma_invsqrt_f(rsq10);
1100             rinv20           = avx128fma_invsqrt_f(rsq20);
1101             rinv30           = avx128fma_invsqrt_f(rsq30);
1102
1103             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1104             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1105             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1106             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
1107
1108             /* Load parameters for j particles */
1109             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1110                                                               charge+jnrC+0,charge+jnrD+0);
1111             vdwjidx0A        = 2*vdwtype[jnrA+0];
1112             vdwjidx0B        = 2*vdwtype[jnrB+0];
1113             vdwjidx0C        = 2*vdwtype[jnrC+0];
1114             vdwjidx0D        = 2*vdwtype[jnrD+0];
1115
1116             fjx0             = _mm_setzero_ps();
1117             fjy0             = _mm_setzero_ps();
1118             fjz0             = _mm_setzero_ps();
1119
1120             /**************************
1121              * CALCULATE INTERACTIONS *
1122              **************************/
1123
1124             r00              = _mm_mul_ps(rsq00,rinv00);
1125             r00              = _mm_andnot_ps(dummy_mask,r00);
1126
1127             /* Compute parameters for interactions between i and j atoms */
1128             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1129                                          vdwparam+vdwioffset0+vdwjidx0B,
1130                                          vdwparam+vdwioffset0+vdwjidx0C,
1131                                          vdwparam+vdwioffset0+vdwjidx0D,
1132                                          &c6_00,&c12_00);
1133
1134             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
1135                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
1136                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
1137                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
1138
1139             /* Analytical LJ-PME */
1140             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1141             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
1142             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
1143             exponent         = avx128fma_exp_f(ewcljrsq);
1144             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1145             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
1146             /* f6A = 6 * C6grid * (1 - poly) */
1147             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
1148             /* f6B = C6grid * exponent * beta^6 */
1149             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
1150             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1151             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1152
1153             fscal            = fvdw;
1154
1155             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1156
1157              /* Update vectorial force */
1158             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1159             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1160             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1161
1162             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1163             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1164             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1165
1166             /**************************
1167              * CALCULATE INTERACTIONS *
1168              **************************/
1169
1170             r10              = _mm_mul_ps(rsq10,rinv10);
1171             r10              = _mm_andnot_ps(dummy_mask,r10);
1172
1173             /* Compute parameters for interactions between i and j atoms */
1174             qq10             = _mm_mul_ps(iq1,jq0);
1175
1176             /* EWALD ELECTROSTATICS */
1177
1178             /* Analytical PME correction */
1179             zeta2            = _mm_mul_ps(beta2,rsq10);
1180             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1181             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1182             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1183             felec            = _mm_mul_ps(qq10,felec);
1184
1185             fscal            = felec;
1186
1187             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1188
1189              /* Update vectorial force */
1190             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1191             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1192             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1193
1194             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1195             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1196             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1197
1198             /**************************
1199              * CALCULATE INTERACTIONS *
1200              **************************/
1201
1202             r20              = _mm_mul_ps(rsq20,rinv20);
1203             r20              = _mm_andnot_ps(dummy_mask,r20);
1204
1205             /* Compute parameters for interactions between i and j atoms */
1206             qq20             = _mm_mul_ps(iq2,jq0);
1207
1208             /* EWALD ELECTROSTATICS */
1209
1210             /* Analytical PME correction */
1211             zeta2            = _mm_mul_ps(beta2,rsq20);
1212             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1213             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1214             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1215             felec            = _mm_mul_ps(qq20,felec);
1216
1217             fscal            = felec;
1218
1219             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1220
1221              /* Update vectorial force */
1222             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1223             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1224             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1225
1226             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1227             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1228             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1229
1230             /**************************
1231              * CALCULATE INTERACTIONS *
1232              **************************/
1233
1234             r30              = _mm_mul_ps(rsq30,rinv30);
1235             r30              = _mm_andnot_ps(dummy_mask,r30);
1236
1237             /* Compute parameters for interactions between i and j atoms */
1238             qq30             = _mm_mul_ps(iq3,jq0);
1239
1240             /* EWALD ELECTROSTATICS */
1241
1242             /* Analytical PME correction */
1243             zeta2            = _mm_mul_ps(beta2,rsq30);
1244             rinv3            = _mm_mul_ps(rinvsq30,rinv30);
1245             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1246             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1247             felec            = _mm_mul_ps(qq30,felec);
1248
1249             fscal            = felec;
1250
1251             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1252
1253              /* Update vectorial force */
1254             fix3             = _mm_macc_ps(dx30,fscal,fix3);
1255             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
1256             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
1257
1258             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
1259             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
1260             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
1261
1262             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1263             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1264             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1265             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1266
1267             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1268
1269             /* Inner loop uses 135 flops */
1270         }
1271
1272         /* End of innermost loop */
1273
1274         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1275                                               f+i_coord_offset,fshift+i_shift_offset);
1276
1277         /* Increment number of inner iterations */
1278         inneriter                  += j_index_end - j_index_start;
1279
1280         /* Outer loop uses 24 flops */
1281     }
1282
1283     /* Increment number of outer iterations */
1284     outeriter        += nri;
1285
1286     /* Update outer/inner flops */
1287
1288     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*135);
1289 }