src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LJEwald
  56  * Geometry:                   Water3-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  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     real             *vdwgridparam;
 108     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 109     __m128           one_half = _mm_set1_ps(0.5);
 110     __m128           minus_one = _mm_set1_ps(-1.0);
 111     __m128i          ewitab;
 112     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 113     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 114     real             *ewtab;
 115     __m128           dummy_mask,cutoff_mask;
 116     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 117     __m128           one     = _mm_set1_ps(1.0);
 118     __m128           two     = _mm_set1_ps(2.0);
 119     x                = xx[0];
 120     f                = ff[0];
 121
 122     nri              = nlist->nri;
 123     iinr             = nlist->iinr;
 124     jindex           = nlist->jindex;
 125     jjnr             = nlist->jjnr;
 126     shiftidx         = nlist->shift;
 127     gid              = nlist->gid;
 128     shiftvec         = fr->shift_vec[0];
 129     fshift           = fr->fshift[0];
 130     facel            = _mm_set1_ps(fr->epsfac);
 131     charge           = mdatoms->chargeA;
 132     nvdwtype         = fr->ntype;
 133     vdwparam         = fr->nbfp;
 134     vdwtype          = mdatoms->typeA;
 135     vdwgridparam     = fr->ljpme_c6grid;
 136     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 137     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 138     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 139
 140     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 141     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 142     beta2            = _mm_mul_ps(beta,beta);
 143     beta3            = _mm_mul_ps(beta,beta2);
 144     ewtab            = fr->ic->tabq_coul_FDV0;
 145     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 146     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 147
 148     /* Setup water-specific parameters */
 149     inr              = nlist->iinr[0];
 150     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+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     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 154
 155     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 156     rcutoff_scalar   = fr->rcoulomb;
 157     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 158     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 159
 160     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 161     rvdw             = _mm_set1_ps(fr->rvdw);
 162
 163     /* Avoid stupid compiler warnings */
 164     jnrA = jnrB = jnrC = jnrD = 0;
 165     j_coord_offsetA = 0;
 166     j_coord_offsetB = 0;
 167     j_coord_offsetC = 0;
 168     j_coord_offsetD = 0;
 169
 170     outeriter        = 0;
 171     inneriter        = 0;
 172
 173     for(iidx=0;iidx<4*DIM;iidx++)
 174     {
 175         scratch[iidx] = 0.0;
 176     }
 177
 178     /* Start outer loop over neighborlists */
 179     for(iidx=0; iidx<nri; iidx++)
 180     {
 181         /* Load shift vector for this list */
 182         i_shift_offset   = DIM*shiftidx[iidx];
 183
 184         /* Load limits for loop over neighbors */
 185         j_index_start    = jindex[iidx];
 186         j_index_end      = jindex[iidx+1];
 187
 188         /* Get outer coordinate index */
 189         inr              = iinr[iidx];
 190         i_coord_offset   = DIM*inr;
 191
 192         /* Load i particle coords and add shift vector */
 193         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 194                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 195
 196         fix0             = _mm_setzero_ps();
 197         fiy0             = _mm_setzero_ps();
 198         fiz0             = _mm_setzero_ps();
 199         fix1             = _mm_setzero_ps();
 200         fiy1             = _mm_setzero_ps();
 201         fiz1             = _mm_setzero_ps();
 202         fix2             = _mm_setzero_ps();
 203         fiy2             = _mm_setzero_ps();
 204         fiz2             = _mm_setzero_ps();
 205
 206         /* Reset potential sums */
 207         velecsum         = _mm_setzero_ps();
 208         vvdwsum          = _mm_setzero_ps();
 209
 210         /* Start inner kernel loop */
 211         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 212         {
 213
 214             /* Get j neighbor index, and coordinate index */
 215             jnrA             = jjnr[jidx];
 216             jnrB             = jjnr[jidx+1];
 217             jnrC             = jjnr[jidx+2];
 218             jnrD             = jjnr[jidx+3];
 219             j_coord_offsetA  = DIM*jnrA;
 220             j_coord_offsetB  = DIM*jnrB;
 221             j_coord_offsetC  = DIM*jnrC;
 222             j_coord_offsetD  = DIM*jnrD;
 223
 224             /* load j atom coordinates */
 225             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 226                                               x+j_coord_offsetC,x+j_coord_offsetD,
 227                                               &jx0,&jy0,&jz0);
 228
 229             /* Calculate displacement vector */
 230             dx00             = _mm_sub_ps(ix0,jx0);
 231             dy00             = _mm_sub_ps(iy0,jy0);
 232             dz00             = _mm_sub_ps(iz0,jz0);
 233             dx10             = _mm_sub_ps(ix1,jx0);
 234             dy10             = _mm_sub_ps(iy1,jy0);
 235             dz10             = _mm_sub_ps(iz1,jz0);
 236             dx20             = _mm_sub_ps(ix2,jx0);
 237             dy20             = _mm_sub_ps(iy2,jy0);
 238             dz20             = _mm_sub_ps(iz2,jz0);
 239
 240             /* Calculate squared distance and things based on it */
 241             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 242             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 243             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 244
 245             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 246             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 247             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 248
 249             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 250             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 251             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 252
 253             /* Load parameters for j particles */
 254             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 255                                                               charge+jnrC+0,charge+jnrD+0);
 256             vdwjidx0A        = 2*vdwtype[jnrA+0];
 257             vdwjidx0B        = 2*vdwtype[jnrB+0];
 258             vdwjidx0C        = 2*vdwtype[jnrC+0];
 259             vdwjidx0D        = 2*vdwtype[jnrD+0];
 260
 261             fjx0             = _mm_setzero_ps();
 262             fjy0             = _mm_setzero_ps();
 263             fjz0             = _mm_setzero_ps();
 264
 265             /**************************
 266              * CALCULATE INTERACTIONS *
 267              **************************/
 268
 269             if (gmx_mm_any_lt(rsq00,rcutoff2))
 270             {
 271
 272             r00              = _mm_mul_ps(rsq00,rinv00);
 273
 274             /* Compute parameters for interactions between i and j atoms */
 275             qq00             = _mm_mul_ps(iq0,jq0);
 276             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 277                                          vdwparam+vdwioffset0+vdwjidx0B,
 278                                          vdwparam+vdwioffset0+vdwjidx0C,
 279                                          vdwparam+vdwioffset0+vdwjidx0D,
 280                                          &c6_00,&c12_00);
 281
 282             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 283                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 284                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 285                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 286
 287             /* EWALD ELECTROSTATICS */
 288
 289             /* Analytical PME correction */
 290             zeta2            = _mm_mul_ps(beta2,rsq00);
 291             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 292             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 293             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 294             felec            = _mm_mul_ps(qq00,felec);
 295             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 296             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 297             velec            = _mm_mul_ps(qq00,velec);
 298
 299             /* Analytical LJ-PME */
 300             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 301             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 302             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 303             exponent         = gmx_simd_exp_r(ewcljrsq);
 304             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 305             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 306             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 307             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 308             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 309             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 310                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 311             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 312             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);
 313
 314             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 315
 316             /* Update potential sum for this i atom from the interaction with this j atom. */
 317             velec            = _mm_and_ps(velec,cutoff_mask);
 318             velecsum         = _mm_add_ps(velecsum,velec);
 319             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 320             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 321
 322             fscal            = _mm_add_ps(felec,fvdw);
 323
 324             fscal            = _mm_and_ps(fscal,cutoff_mask);
 325
 326              /* Update vectorial force */
 327             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 328             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 329             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 330
 331             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 332             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 333             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 334
 335             }
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             if (gmx_mm_any_lt(rsq10,rcutoff2))
 342             {
 343
 344             r10              = _mm_mul_ps(rsq10,rinv10);
 345
 346             /* Compute parameters for interactions between i and j atoms */
 347             qq10             = _mm_mul_ps(iq1,jq0);
 348
 349             /* EWALD ELECTROSTATICS */
 350
 351             /* Analytical PME correction */
 352             zeta2            = _mm_mul_ps(beta2,rsq10);
 353             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 354             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 355             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 356             felec            = _mm_mul_ps(qq10,felec);
 357             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 358             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 359             velec            = _mm_mul_ps(qq10,velec);
 360
 361             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velec            = _mm_and_ps(velec,cutoff_mask);
 365             velecsum         = _mm_add_ps(velecsum,velec);
 366
 367             fscal            = felec;
 368
 369             fscal            = _mm_and_ps(fscal,cutoff_mask);
 370
 371              /* Update vectorial force */
 372             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 373             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 374             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 375
 376             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 377             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 378             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 379
 380             }
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             if (gmx_mm_any_lt(rsq20,rcutoff2))
 387             {
 388
 389             r20              = _mm_mul_ps(rsq20,rinv20);
 390
 391             /* Compute parameters for interactions between i and j atoms */
 392             qq20             = _mm_mul_ps(iq2,jq0);
 393
 394             /* EWALD ELECTROSTATICS */
 395
 396             /* Analytical PME correction */
 397             zeta2            = _mm_mul_ps(beta2,rsq20);
 398             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 399             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 400             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 401             felec            = _mm_mul_ps(qq20,felec);
 402             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 403             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 404             velec            = _mm_mul_ps(qq20,velec);
 405
 406             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velec            = _mm_and_ps(velec,cutoff_mask);
 410             velecsum         = _mm_add_ps(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414             fscal            = _mm_and_ps(fscal,cutoff_mask);
 415
 416              /* Update vectorial force */
 417             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 418             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 419             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 420
 421             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 422             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 423             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 424
 425             }
 426
 427             fjptrA             = f+j_coord_offsetA;
 428             fjptrB             = f+j_coord_offsetB;
 429             fjptrC             = f+j_coord_offsetC;
 430             fjptrD             = f+j_coord_offsetD;
 431
 432             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 433
 434             /* Inner loop uses 129 flops */
 435         }
 436
 437         if(jidx<j_index_end)
 438         {
 439
 440             /* Get j neighbor index, and coordinate index */
 441             jnrlistA         = jjnr[jidx];
 442             jnrlistB         = jjnr[jidx+1];
 443             jnrlistC         = jjnr[jidx+2];
 444             jnrlistD         = jjnr[jidx+3];
 445             /* Sign of each element will be negative for non-real atoms.
 446              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 447              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 448              */
 449             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 450             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 451             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 452             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 453             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 454             j_coord_offsetA  = DIM*jnrA;
 455             j_coord_offsetB  = DIM*jnrB;
 456             j_coord_offsetC  = DIM*jnrC;
 457             j_coord_offsetD  = DIM*jnrD;
 458
 459             /* load j atom coordinates */
 460             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 461                                               x+j_coord_offsetC,x+j_coord_offsetD,
 462                                               &jx0,&jy0,&jz0);
 463
 464             /* Calculate displacement vector */
 465             dx00             = _mm_sub_ps(ix0,jx0);
 466             dy00             = _mm_sub_ps(iy0,jy0);
 467             dz00             = _mm_sub_ps(iz0,jz0);
 468             dx10             = _mm_sub_ps(ix1,jx0);
 469             dy10             = _mm_sub_ps(iy1,jy0);
 470             dz10             = _mm_sub_ps(iz1,jz0);
 471             dx20             = _mm_sub_ps(ix2,jx0);
 472             dy20             = _mm_sub_ps(iy2,jy0);
 473             dz20             = _mm_sub_ps(iz2,jz0);
 474
 475             /* Calculate squared distance and things based on it */
 476             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 477             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 478             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 479
 480             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 481             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 482             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 483
 484             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 485             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 486             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 487
 488             /* Load parameters for j particles */
 489             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 490                                                               charge+jnrC+0,charge+jnrD+0);
 491             vdwjidx0A        = 2*vdwtype[jnrA+0];
 492             vdwjidx0B        = 2*vdwtype[jnrB+0];
 493             vdwjidx0C        = 2*vdwtype[jnrC+0];
 494             vdwjidx0D        = 2*vdwtype[jnrD+0];
 495
 496             fjx0             = _mm_setzero_ps();
 497             fjy0             = _mm_setzero_ps();
 498             fjz0             = _mm_setzero_ps();
 499
 500             /**************************
 501              * CALCULATE INTERACTIONS *
 502              **************************/
 503
 504             if (gmx_mm_any_lt(rsq00,rcutoff2))
 505             {
 506
 507             r00              = _mm_mul_ps(rsq00,rinv00);
 508             r00              = _mm_andnot_ps(dummy_mask,r00);
 509
 510             /* Compute parameters for interactions between i and j atoms */
 511             qq00             = _mm_mul_ps(iq0,jq0);
 512             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 513                                          vdwparam+vdwioffset0+vdwjidx0B,
 514                                          vdwparam+vdwioffset0+vdwjidx0C,
 515                                          vdwparam+vdwioffset0+vdwjidx0D,
 516                                          &c6_00,&c12_00);
 517
 518             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 519                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 520                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 521                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 522
 523             /* EWALD ELECTROSTATICS */
 524
 525             /* Analytical PME correction */
 526             zeta2            = _mm_mul_ps(beta2,rsq00);
 527             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 528             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 529             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 530             felec            = _mm_mul_ps(qq00,felec);
 531             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 532             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 533             velec            = _mm_mul_ps(qq00,velec);
 534
 535             /* Analytical LJ-PME */
 536             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 537             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 538             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 539             exponent         = gmx_simd_exp_r(ewcljrsq);
 540             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 541             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 542             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 543             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 544             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 545             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 546                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 547             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 548             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);
 549
 550             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 551
 552             /* Update potential sum for this i atom from the interaction with this j atom. */
 553             velec            = _mm_and_ps(velec,cutoff_mask);
 554             velec            = _mm_andnot_ps(dummy_mask,velec);
 555             velecsum         = _mm_add_ps(velecsum,velec);
 556             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 557             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 558             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 559
 560             fscal            = _mm_add_ps(felec,fvdw);
 561
 562             fscal            = _mm_and_ps(fscal,cutoff_mask);
 563
 564             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 565
 566              /* Update vectorial force */
 567             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 568             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 569             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 570
 571             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 572             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 573             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 574
 575             }
 576
 577             /**************************
 578              * CALCULATE INTERACTIONS *
 579              **************************/
 580
 581             if (gmx_mm_any_lt(rsq10,rcutoff2))
 582             {
 583
 584             r10              = _mm_mul_ps(rsq10,rinv10);
 585             r10              = _mm_andnot_ps(dummy_mask,r10);
 586
 587             /* Compute parameters for interactions between i and j atoms */
 588             qq10             = _mm_mul_ps(iq1,jq0);
 589
 590             /* EWALD ELECTROSTATICS */
 591
 592             /* Analytical PME correction */
 593             zeta2            = _mm_mul_ps(beta2,rsq10);
 594             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 595             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 596             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 597             felec            = _mm_mul_ps(qq10,felec);
 598             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 599             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 600             velec            = _mm_mul_ps(qq10,velec);
 601
 602             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 603
 604             /* Update potential sum for this i atom from the interaction with this j atom. */
 605             velec            = _mm_and_ps(velec,cutoff_mask);
 606             velec            = _mm_andnot_ps(dummy_mask,velec);
 607             velecsum         = _mm_add_ps(velecsum,velec);
 608
 609             fscal            = felec;
 610
 611             fscal            = _mm_and_ps(fscal,cutoff_mask);
 612
 613             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 614
 615              /* Update vectorial force */
 616             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 617             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 618             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 619
 620             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 621             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 622             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 623
 624             }
 625
 626             /**************************
 627              * CALCULATE INTERACTIONS *
 628              **************************/
 629
 630             if (gmx_mm_any_lt(rsq20,rcutoff2))
 631             {
 632
 633             r20              = _mm_mul_ps(rsq20,rinv20);
 634             r20              = _mm_andnot_ps(dummy_mask,r20);
 635
 636             /* Compute parameters for interactions between i and j atoms */
 637             qq20             = _mm_mul_ps(iq2,jq0);
 638
 639             /* EWALD ELECTROSTATICS */
 640
 641             /* Analytical PME correction */
 642             zeta2            = _mm_mul_ps(beta2,rsq20);
 643             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 644             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 645             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 646             felec            = _mm_mul_ps(qq20,felec);
 647             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 648             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 649             velec            = _mm_mul_ps(qq20,velec);
 650
 651             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 652
 653             /* Update potential sum for this i atom from the interaction with this j atom. */
 654             velec            = _mm_and_ps(velec,cutoff_mask);
 655             velec            = _mm_andnot_ps(dummy_mask,velec);
 656             velecsum         = _mm_add_ps(velecsum,velec);
 657
 658             fscal            = felec;
 659
 660             fscal            = _mm_and_ps(fscal,cutoff_mask);
 661
 662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 663
 664              /* Update vectorial force */
 665             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 666             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 667             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 668
 669             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 670             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 671             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 672
 673             }
 674
 675             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 676             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 677             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 678             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 679
 680             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 681
 682             /* Inner loop uses 132 flops */
 683         }
 684
 685         /* End of innermost loop */
 686
 687         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 688                                               f+i_coord_offset,fshift+i_shift_offset);
 689
 690         ggid                        = gid[iidx];
 691         /* Update potential energies */
 692         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 693         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 694
 695         /* Increment number of inner iterations */
 696         inneriter                  += j_index_end - j_index_start;
 697
 698         /* Outer loop uses 20 flops */
 699     }
 700
 701     /* Increment number of outer iterations */
 702     outeriter        += nri;
 703
 704     /* Update outer/inner flops */
 705
 706     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*132);
 707 }
 708 /*
 709  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_F_avx_128_fma_single
 710  * Electrostatics interaction: Ewald
 711  * VdW interaction:            LJEwald
 712  * Geometry:                   Water3-Particle
 713  * Calculate force/pot:        Force
 714  */
 715 void
 716 nb_kernel_ElecEwSh_VdwLJEwSh_GeomW3P1_F_avx_128_fma_single
 717                     (t_nblist                    * gmx_restrict       nlist,
 718                      rvec                        * gmx_restrict          xx,
 719                      rvec                        * gmx_restrict          ff,
 720                      t_forcerec                  * gmx_restrict          fr,
 721                      t_mdatoms                   * gmx_restrict     mdatoms,
 722                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 723                      t_nrnb                      * gmx_restrict        nrnb)
 724 {
 725     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 726      * just 0 for non-waters.
 727      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 728      * jnr indices corresponding to data put in the four positions in the SIMD register.
 729      */
 730     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 731     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 732     int              jnrA,jnrB,jnrC,jnrD;
 733     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 734     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 735     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 736     real             rcutoff_scalar;
 737     real             *shiftvec,*fshift,*x,*f;
 738     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 739     real             scratch[4*DIM];
 740     __m128           fscal,rcutoff,rcutoff2,jidxall;
 741     int              vdwioffset0;
 742     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 743     int              vdwioffset1;
 744     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 745     int              vdwioffset2;
 746     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 747     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 748     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 749     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 750     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 751     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 752     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 753     real             *charge;
 754     int              nvdwtype;
 755     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 756     int              *vdwtype;
 757     real             *vdwparam;
 758     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 759     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 760     __m128           c6grid_00;
 761     __m128           c6grid_10;
 762     __m128           c6grid_20;
 763     real             *vdwgridparam;
 764     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 765     __m128           one_half = _mm_set1_ps(0.5);
 766     __m128           minus_one = _mm_set1_ps(-1.0);
 767     __m128i          ewitab;
 768     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 769     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 770     real             *ewtab;
 771     __m128           dummy_mask,cutoff_mask;
 772     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 773     __m128           one     = _mm_set1_ps(1.0);
 774     __m128           two     = _mm_set1_ps(2.0);
 775     x                = xx[0];
 776     f                = ff[0];
 777
 778     nri              = nlist->nri;
 779     iinr             = nlist->iinr;
 780     jindex           = nlist->jindex;
 781     jjnr             = nlist->jjnr;
 782     shiftidx         = nlist->shift;
 783     gid              = nlist->gid;
 784     shiftvec         = fr->shift_vec[0];
 785     fshift           = fr->fshift[0];
 786     facel            = _mm_set1_ps(fr->epsfac);
 787     charge           = mdatoms->chargeA;
 788     nvdwtype         = fr->ntype;
 789     vdwparam         = fr->nbfp;
 790     vdwtype          = mdatoms->typeA;
 791     vdwgridparam     = fr->ljpme_c6grid;
 792     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 793     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 794     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 795
 796     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 797     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 798     beta2            = _mm_mul_ps(beta,beta);
 799     beta3            = _mm_mul_ps(beta,beta2);
 800     ewtab            = fr->ic->tabq_coul_F;
 801     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 802     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 803
 804     /* Setup water-specific parameters */
 805     inr              = nlist->iinr[0];
 806     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 807     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 808     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 809     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 810
 811     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 812     rcutoff_scalar   = fr->rcoulomb;
 813     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 814     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 815
 816     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 817     rvdw             = _mm_set1_ps(fr->rvdw);
 818
 819     /* Avoid stupid compiler warnings */
 820     jnrA = jnrB = jnrC = jnrD = 0;
 821     j_coord_offsetA = 0;
 822     j_coord_offsetB = 0;
 823     j_coord_offsetC = 0;
 824     j_coord_offsetD = 0;
 825
 826     outeriter        = 0;
 827     inneriter        = 0;
 828
 829     for(iidx=0;iidx<4*DIM;iidx++)
 830     {
 831         scratch[iidx] = 0.0;
 832     }
 833
 834     /* Start outer loop over neighborlists */
 835     for(iidx=0; iidx<nri; iidx++)
 836     {
 837         /* Load shift vector for this list */
 838         i_shift_offset   = DIM*shiftidx[iidx];
 839
 840         /* Load limits for loop over neighbors */
 841         j_index_start    = jindex[iidx];
 842         j_index_end      = jindex[iidx+1];
 843
 844         /* Get outer coordinate index */
 845         inr              = iinr[iidx];
 846         i_coord_offset   = DIM*inr;
 847
 848         /* Load i particle coords and add shift vector */
 849         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 850                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 851
 852         fix0             = _mm_setzero_ps();
 853         fiy0             = _mm_setzero_ps();
 854         fiz0             = _mm_setzero_ps();
 855         fix1             = _mm_setzero_ps();
 856         fiy1             = _mm_setzero_ps();
 857         fiz1             = _mm_setzero_ps();
 858         fix2             = _mm_setzero_ps();
 859         fiy2             = _mm_setzero_ps();
 860         fiz2             = _mm_setzero_ps();
 861
 862         /* Start inner kernel loop */
 863         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 864         {
 865
 866             /* Get j neighbor index, and coordinate index */
 867             jnrA             = jjnr[jidx];
 868             jnrB             = jjnr[jidx+1];
 869             jnrC             = jjnr[jidx+2];
 870             jnrD             = jjnr[jidx+3];
 871             j_coord_offsetA  = DIM*jnrA;
 872             j_coord_offsetB  = DIM*jnrB;
 873             j_coord_offsetC  = DIM*jnrC;
 874             j_coord_offsetD  = DIM*jnrD;
 875
 876             /* load j atom coordinates */
 877             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 878                                               x+j_coord_offsetC,x+j_coord_offsetD,
 879                                               &jx0,&jy0,&jz0);
 880
 881             /* Calculate displacement vector */
 882             dx00             = _mm_sub_ps(ix0,jx0);
 883             dy00             = _mm_sub_ps(iy0,jy0);
 884             dz00             = _mm_sub_ps(iz0,jz0);
 885             dx10             = _mm_sub_ps(ix1,jx0);
 886             dy10             = _mm_sub_ps(iy1,jy0);
 887             dz10             = _mm_sub_ps(iz1,jz0);
 888             dx20             = _mm_sub_ps(ix2,jx0);
 889             dy20             = _mm_sub_ps(iy2,jy0);
 890             dz20             = _mm_sub_ps(iz2,jz0);
 891
 892             /* Calculate squared distance and things based on it */
 893             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 894             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 895             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 896
 897             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 898             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 899             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 900
 901             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 902             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 903             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 904
 905             /* Load parameters for j particles */
 906             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 907                                                               charge+jnrC+0,charge+jnrD+0);
 908             vdwjidx0A        = 2*vdwtype[jnrA+0];
 909             vdwjidx0B        = 2*vdwtype[jnrB+0];
 910             vdwjidx0C        = 2*vdwtype[jnrC+0];
 911             vdwjidx0D        = 2*vdwtype[jnrD+0];
 912
 913             fjx0             = _mm_setzero_ps();
 914             fjy0             = _mm_setzero_ps();
 915             fjz0             = _mm_setzero_ps();
 916
 917             /**************************
 918              * CALCULATE INTERACTIONS *
 919              **************************/
 920
 921             if (gmx_mm_any_lt(rsq00,rcutoff2))
 922             {
 923
 924             r00              = _mm_mul_ps(rsq00,rinv00);
 925
 926             /* Compute parameters for interactions between i and j atoms */
 927             qq00             = _mm_mul_ps(iq0,jq0);
 928             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 929                                          vdwparam+vdwioffset0+vdwjidx0B,
 930                                          vdwparam+vdwioffset0+vdwjidx0C,
 931                                          vdwparam+vdwioffset0+vdwjidx0D,
 932                                          &c6_00,&c12_00);
 933
 934             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 935                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 936                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 937                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 938
 939             /* EWALD ELECTROSTATICS */
 940
 941             /* Analytical PME correction */
 942             zeta2            = _mm_mul_ps(beta2,rsq00);
 943             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 944             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 945             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 946             felec            = _mm_mul_ps(qq00,felec);
 947
 948             /* Analytical LJ-PME */
 949             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 950             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 951             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 952             exponent         = gmx_simd_exp_r(ewcljrsq);
 953             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 954             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 955             /* f6A = 6 * C6grid * (1 - poly) */
 956             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 957             /* f6B = C6grid * exponent * beta^6 */
 958             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 959             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 960             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 961
 962             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 963
 964             fscal            = _mm_add_ps(felec,fvdw);
 965
 966             fscal            = _mm_and_ps(fscal,cutoff_mask);
 967
 968              /* Update vectorial force */
 969             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 970             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 971             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 972
 973             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 974             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 975             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 976
 977             }
 978
 979             /**************************
 980              * CALCULATE INTERACTIONS *
 981              **************************/
 982
 983             if (gmx_mm_any_lt(rsq10,rcutoff2))
 984             {
 985
 986             r10              = _mm_mul_ps(rsq10,rinv10);
 987
 988             /* Compute parameters for interactions between i and j atoms */
 989             qq10             = _mm_mul_ps(iq1,jq0);
 990
 991             /* EWALD ELECTROSTATICS */
 992
 993             /* Analytical PME correction */
 994             zeta2            = _mm_mul_ps(beta2,rsq10);
 995             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 996             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 997             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 998             felec            = _mm_mul_ps(qq10,felec);
 999
1000             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1001
1002             fscal            = felec;
1003
1004             fscal            = _mm_and_ps(fscal,cutoff_mask);
1005
1006              /* Update vectorial force */
1007             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1008             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1009             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1010
1011             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1012             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1013             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1014
1015             }
1016
1017             /**************************
1018              * CALCULATE INTERACTIONS *
1019              **************************/
1020
1021             if (gmx_mm_any_lt(rsq20,rcutoff2))
1022             {
1023
1024             r20              = _mm_mul_ps(rsq20,rinv20);
1025
1026             /* Compute parameters for interactions between i and j atoms */
1027             qq20             = _mm_mul_ps(iq2,jq0);
1028
1029             /* EWALD ELECTROSTATICS */
1030
1031             /* Analytical PME correction */
1032             zeta2            = _mm_mul_ps(beta2,rsq20);
1033             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1034             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1035             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1036             felec            = _mm_mul_ps(qq20,felec);
1037
1038             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1039
1040             fscal            = felec;
1041
1042             fscal            = _mm_and_ps(fscal,cutoff_mask);
1043
1044              /* Update vectorial force */
1045             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1046             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1047             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1048
1049             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1050             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1051             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1052
1053             }
1054
1055             fjptrA             = f+j_coord_offsetA;
1056             fjptrB             = f+j_coord_offsetB;
1057             fjptrC             = f+j_coord_offsetC;
1058             fjptrD             = f+j_coord_offsetD;
1059
1060             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1061
1062             /* Inner loop uses 114 flops */
1063         }
1064
1065         if(jidx<j_index_end)
1066         {
1067
1068             /* Get j neighbor index, and coordinate index */
1069             jnrlistA         = jjnr[jidx];
1070             jnrlistB         = jjnr[jidx+1];
1071             jnrlistC         = jjnr[jidx+2];
1072             jnrlistD         = jjnr[jidx+3];
1073             /* Sign of each element will be negative for non-real atoms.
1074              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1075              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1076              */
1077             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1078             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1079             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1080             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1081             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1082             j_coord_offsetA  = DIM*jnrA;
1083             j_coord_offsetB  = DIM*jnrB;
1084             j_coord_offsetC  = DIM*jnrC;
1085             j_coord_offsetD  = DIM*jnrD;
1086
1087             /* load j atom coordinates */
1088             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1089                                               x+j_coord_offsetC,x+j_coord_offsetD,
1090                                               &jx0,&jy0,&jz0);
1091
1092             /* Calculate displacement vector */
1093             dx00             = _mm_sub_ps(ix0,jx0);
1094             dy00             = _mm_sub_ps(iy0,jy0);
1095             dz00             = _mm_sub_ps(iz0,jz0);
1096             dx10             = _mm_sub_ps(ix1,jx0);
1097             dy10             = _mm_sub_ps(iy1,jy0);
1098             dz10             = _mm_sub_ps(iz1,jz0);
1099             dx20             = _mm_sub_ps(ix2,jx0);
1100             dy20             = _mm_sub_ps(iy2,jy0);
1101             dz20             = _mm_sub_ps(iz2,jz0);
1102
1103             /* Calculate squared distance and things based on it */
1104             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1105             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1106             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1107
1108             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1109             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1110             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1111
1112             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1113             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1114             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1115
1116             /* Load parameters for j particles */
1117             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1118                                                               charge+jnrC+0,charge+jnrD+0);
1119             vdwjidx0A        = 2*vdwtype[jnrA+0];
1120             vdwjidx0B        = 2*vdwtype[jnrB+0];
1121             vdwjidx0C        = 2*vdwtype[jnrC+0];
1122             vdwjidx0D        = 2*vdwtype[jnrD+0];
1123
1124             fjx0             = _mm_setzero_ps();
1125             fjy0             = _mm_setzero_ps();
1126             fjz0             = _mm_setzero_ps();
1127
1128             /**************************
1129              * CALCULATE INTERACTIONS *
1130              **************************/
1131
1132             if (gmx_mm_any_lt(rsq00,rcutoff2))
1133             {
1134
1135             r00              = _mm_mul_ps(rsq00,rinv00);
1136             r00              = _mm_andnot_ps(dummy_mask,r00);
1137
1138             /* Compute parameters for interactions between i and j atoms */
1139             qq00             = _mm_mul_ps(iq0,jq0);
1140             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1141                                          vdwparam+vdwioffset0+vdwjidx0B,
1142                                          vdwparam+vdwioffset0+vdwjidx0C,
1143                                          vdwparam+vdwioffset0+vdwjidx0D,
1144                                          &c6_00,&c12_00);
1145
1146             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
1147                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
1148                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
1149                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
1150
1151             /* EWALD ELECTROSTATICS */
1152
1153             /* Analytical PME correction */
1154             zeta2            = _mm_mul_ps(beta2,rsq00);
1155             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1156             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1157             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1158             felec            = _mm_mul_ps(qq00,felec);
1159
1160             /* Analytical LJ-PME */
1161             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1162             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
1163             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
1164             exponent         = gmx_simd_exp_r(ewcljrsq);
1165             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1166             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
1167             /* f6A = 6 * C6grid * (1 - poly) */
1168             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
1169             /* f6B = C6grid * exponent * beta^6 */
1170             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
1171             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1172             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1173
1174             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1175
1176             fscal            = _mm_add_ps(felec,fvdw);
1177
1178             fscal            = _mm_and_ps(fscal,cutoff_mask);
1179
1180             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1181
1182              /* Update vectorial force */
1183             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1184             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1185             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1186
1187             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1188             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1189             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1190
1191             }
1192
1193             /**************************
1194              * CALCULATE INTERACTIONS *
1195              **************************/
1196
1197             if (gmx_mm_any_lt(rsq10,rcutoff2))
1198             {
1199
1200             r10              = _mm_mul_ps(rsq10,rinv10);
1201             r10              = _mm_andnot_ps(dummy_mask,r10);
1202
1203             /* Compute parameters for interactions between i and j atoms */
1204             qq10             = _mm_mul_ps(iq1,jq0);
1205
1206             /* EWALD ELECTROSTATICS */
1207
1208             /* Analytical PME correction */
1209             zeta2            = _mm_mul_ps(beta2,rsq10);
1210             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1211             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1212             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1213             felec            = _mm_mul_ps(qq10,felec);
1214
1215             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1216
1217             fscal            = felec;
1218
1219             fscal            = _mm_and_ps(fscal,cutoff_mask);
1220
1221             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1222
1223              /* Update vectorial force */
1224             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1225             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1226             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1227
1228             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1229             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1230             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1231
1232             }
1233
1234             /**************************
1235              * CALCULATE INTERACTIONS *
1236              **************************/
1237
1238             if (gmx_mm_any_lt(rsq20,rcutoff2))
1239             {
1240
1241             r20              = _mm_mul_ps(rsq20,rinv20);
1242             r20              = _mm_andnot_ps(dummy_mask,r20);
1243
1244             /* Compute parameters for interactions between i and j atoms */
1245             qq20             = _mm_mul_ps(iq2,jq0);
1246
1247             /* EWALD ELECTROSTATICS */
1248
1249             /* Analytical PME correction */
1250             zeta2            = _mm_mul_ps(beta2,rsq20);
1251             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1252             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1253             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1254             felec            = _mm_mul_ps(qq20,felec);
1255
1256             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1257
1258             fscal            = felec;
1259
1260             fscal            = _mm_and_ps(fscal,cutoff_mask);
1261
1262             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1263
1264              /* Update vectorial force */
1265             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1266             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1267             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1268
1269             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1270             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1271             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1272
1273             }
1274
1275             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1276             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1277             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1278             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1279
1280             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1281
1282             /* Inner loop uses 117 flops */
1283         }
1284
1285         /* End of innermost loop */
1286
1287         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1288                                               f+i_coord_offset,fshift+i_shift_offset);
1289
1290         /* Increment number of inner iterations */
1291         inneriter                  += j_index_end - j_index_start;
1292
1293         /* Outer loop uses 18 flops */
1294     }
1295
1296     /* Increment number of outer iterations */
1297     outeriter        += nri;
1298
1299     /* Update outer/inner flops */
1300
1301     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*117);
1302 }