src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwCSTab_GeomW3P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW3P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwCSTab_GeomW3P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  92     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  93     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 100     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 101     __m128i          vfitab;
 102     __m128i          ifour       = _mm_set1_epi32(4);
 103     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 104     real             *vftab;
 105     __m128i          ewitab;
 106     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 107     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 108     real             *ewtab;
 109     __m128           dummy_mask,cutoff_mask;
 110     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 111     __m128           one     = _mm_set1_ps(1.0);
 112     __m128           two     = _mm_set1_ps(2.0);
 113     x                = xx[0];
 114     f                = ff[0];
 115
 116     nri              = nlist->nri;
 117     iinr             = nlist->iinr;
 118     jindex           = nlist->jindex;
 119     jjnr             = nlist->jjnr;
 120     shiftidx         = nlist->shift;
 121     gid              = nlist->gid;
 122     shiftvec         = fr->shift_vec[0];
 123     fshift           = fr->fshift[0];
 124     facel            = _mm_set1_ps(fr->ic->epsfac);
 125     charge           = mdatoms->chargeA;
 126     nvdwtype         = fr->ntype;
 127     vdwparam         = fr->nbfp;
 128     vdwtype          = mdatoms->typeA;
 129
 130     vftab            = kernel_data->table_vdw->data;
 131     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 132
 133     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 134     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 135     beta2            = _mm_mul_ps(beta,beta);
 136     beta3            = _mm_mul_ps(beta,beta2);
 137     ewtab            = fr->ic->tabq_coul_FDV0;
 138     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 139     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 140
 141     /* Setup water-specific parameters */
 142     inr              = nlist->iinr[0];
 143     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 144     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 145     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 146     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 147
 148     /* Avoid stupid compiler warnings */
 149     jnrA = jnrB = jnrC = jnrD = 0;
 150     j_coord_offsetA = 0;
 151     j_coord_offsetB = 0;
 152     j_coord_offsetC = 0;
 153     j_coord_offsetD = 0;
 154
 155     outeriter        = 0;
 156     inneriter        = 0;
 157
 158     for(iidx=0;iidx<4*DIM;iidx++)
 159     {
 160         scratch[iidx] = 0.0;
 161     }
 162
 163     /* Start outer loop over neighborlists */
 164     for(iidx=0; iidx<nri; iidx++)
 165     {
 166         /* Load shift vector for this list */
 167         i_shift_offset   = DIM*shiftidx[iidx];
 168
 169         /* Load limits for loop over neighbors */
 170         j_index_start    = jindex[iidx];
 171         j_index_end      = jindex[iidx+1];
 172
 173         /* Get outer coordinate index */
 174         inr              = iinr[iidx];
 175         i_coord_offset   = DIM*inr;
 176
 177         /* Load i particle coords and add shift vector */
 178         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 179                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 180
 181         fix0             = _mm_setzero_ps();
 182         fiy0             = _mm_setzero_ps();
 183         fiz0             = _mm_setzero_ps();
 184         fix1             = _mm_setzero_ps();
 185         fiy1             = _mm_setzero_ps();
 186         fiz1             = _mm_setzero_ps();
 187         fix2             = _mm_setzero_ps();
 188         fiy2             = _mm_setzero_ps();
 189         fiz2             = _mm_setzero_ps();
 190
 191         /* Reset potential sums */
 192         velecsum         = _mm_setzero_ps();
 193         vvdwsum          = _mm_setzero_ps();
 194
 195         /* Start inner kernel loop */
 196         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 197         {
 198
 199             /* Get j neighbor index, and coordinate index */
 200             jnrA             = jjnr[jidx];
 201             jnrB             = jjnr[jidx+1];
 202             jnrC             = jjnr[jidx+2];
 203             jnrD             = jjnr[jidx+3];
 204             j_coord_offsetA  = DIM*jnrA;
 205             j_coord_offsetB  = DIM*jnrB;
 206             j_coord_offsetC  = DIM*jnrC;
 207             j_coord_offsetD  = DIM*jnrD;
 208
 209             /* load j atom coordinates */
 210             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 211                                               x+j_coord_offsetC,x+j_coord_offsetD,
 212                                               &jx0,&jy0,&jz0);
 213
 214             /* Calculate displacement vector */
 215             dx00             = _mm_sub_ps(ix0,jx0);
 216             dy00             = _mm_sub_ps(iy0,jy0);
 217             dz00             = _mm_sub_ps(iz0,jz0);
 218             dx10             = _mm_sub_ps(ix1,jx0);
 219             dy10             = _mm_sub_ps(iy1,jy0);
 220             dz10             = _mm_sub_ps(iz1,jz0);
 221             dx20             = _mm_sub_ps(ix2,jx0);
 222             dy20             = _mm_sub_ps(iy2,jy0);
 223             dz20             = _mm_sub_ps(iz2,jz0);
 224
 225             /* Calculate squared distance and things based on it */
 226             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 227             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 228             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 229
 230             rinv00           = avx128fma_invsqrt_f(rsq00);
 231             rinv10           = avx128fma_invsqrt_f(rsq10);
 232             rinv20           = avx128fma_invsqrt_f(rsq20);
 233
 234             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 235             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 236             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 237
 238             /* Load parameters for j particles */
 239             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 240                                                               charge+jnrC+0,charge+jnrD+0);
 241             vdwjidx0A        = 2*vdwtype[jnrA+0];
 242             vdwjidx0B        = 2*vdwtype[jnrB+0];
 243             vdwjidx0C        = 2*vdwtype[jnrC+0];
 244             vdwjidx0D        = 2*vdwtype[jnrD+0];
 245
 246             fjx0             = _mm_setzero_ps();
 247             fjy0             = _mm_setzero_ps();
 248             fjz0             = _mm_setzero_ps();
 249
 250             /**************************
 251              * CALCULATE INTERACTIONS *
 252              **************************/
 253
 254             r00              = _mm_mul_ps(rsq00,rinv00);
 255
 256             /* Compute parameters for interactions between i and j atoms */
 257             qq00             = _mm_mul_ps(iq0,jq0);
 258             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 259                                          vdwparam+vdwioffset0+vdwjidx0B,
 260                                          vdwparam+vdwioffset0+vdwjidx0C,
 261                                          vdwparam+vdwioffset0+vdwjidx0D,
 262                                          &c6_00,&c12_00);
 263
 264             /* Calculate table index by multiplying r with table scale and truncate to integer */
 265             rt               = _mm_mul_ps(r00,vftabscale);
 266             vfitab           = _mm_cvttps_epi32(rt);
 267 #ifdef __XOP__
 268             vfeps            = _mm_frcz_ps(rt);
 269 #else
 270             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 271 #endif
 272             twovfeps         = _mm_add_ps(vfeps,vfeps);
 273             vfitab           = _mm_slli_epi32(vfitab,3);
 274
 275             /* EWALD ELECTROSTATICS */
 276
 277             /* Analytical PME correction */
 278             zeta2            = _mm_mul_ps(beta2,rsq00);
 279             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 280             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 281             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 282             felec            = _mm_mul_ps(qq00,felec);
 283             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 284             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 285             velec            = _mm_mul_ps(qq00,velec);
 286
 287             /* CUBIC SPLINE TABLE DISPERSION */
 288             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 289             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 290             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 291             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 292             _MM_TRANSPOSE4_PS(Y,F,G,H);
 293             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 294             VV               = _mm_macc_ps(vfeps,Fp,Y);
 295             vvdw6            = _mm_mul_ps(c6_00,VV);
 296             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 297             fvdw6            = _mm_mul_ps(c6_00,FF);
 298
 299             /* CUBIC SPLINE TABLE REPULSION */
 300             vfitab           = _mm_add_epi32(vfitab,ifour);
 301             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 302             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 303             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 304             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 305             _MM_TRANSPOSE4_PS(Y,F,G,H);
 306             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 307             VV               = _mm_macc_ps(vfeps,Fp,Y);
 308             vvdw12           = _mm_mul_ps(c12_00,VV);
 309             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 310             fvdw12           = _mm_mul_ps(c12_00,FF);
 311             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 312             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 313
 314             /* Update potential sum for this i atom from the interaction with this j atom. */
 315             velecsum         = _mm_add_ps(velecsum,velec);
 316             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 317
 318             fscal            = _mm_add_ps(felec,fvdw);
 319
 320              /* Update vectorial force */
 321             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 322             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 323             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 324
 325             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 326             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 327             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             r10              = _mm_mul_ps(rsq10,rinv10);
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq10             = _mm_mul_ps(iq1,jq0);
 337
 338             /* EWALD ELECTROSTATICS */
 339
 340             /* Analytical PME correction */
 341             zeta2            = _mm_mul_ps(beta2,rsq10);
 342             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 343             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 344             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 345             felec            = _mm_mul_ps(qq10,felec);
 346             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 347             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 348             velec            = _mm_mul_ps(qq10,velec);
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             velecsum         = _mm_add_ps(velecsum,velec);
 352
 353             fscal            = felec;
 354
 355              /* Update vectorial force */
 356             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 357             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 358             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 359
 360             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 361             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 362             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 363
 364             /**************************
 365              * CALCULATE INTERACTIONS *
 366              **************************/
 367
 368             r20              = _mm_mul_ps(rsq20,rinv20);
 369
 370             /* Compute parameters for interactions between i and j atoms */
 371             qq20             = _mm_mul_ps(iq2,jq0);
 372
 373             /* EWALD ELECTROSTATICS */
 374
 375             /* Analytical PME correction */
 376             zeta2            = _mm_mul_ps(beta2,rsq20);
 377             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 378             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 379             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 380             felec            = _mm_mul_ps(qq20,felec);
 381             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 382             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 383             velec            = _mm_mul_ps(qq20,velec);
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velecsum         = _mm_add_ps(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390              /* Update vectorial force */
 391             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 392             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 393             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 394
 395             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 396             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 397             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 398
 399             fjptrA             = f+j_coord_offsetA;
 400             fjptrB             = f+j_coord_offsetB;
 401             fjptrC             = f+j_coord_offsetC;
 402             fjptrD             = f+j_coord_offsetD;
 403
 404             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 405
 406             /* Inner loop uses 121 flops */
 407         }
 408
 409         if(jidx<j_index_end)
 410         {
 411
 412             /* Get j neighbor index, and coordinate index */
 413             jnrlistA         = jjnr[jidx];
 414             jnrlistB         = jjnr[jidx+1];
 415             jnrlistC         = jjnr[jidx+2];
 416             jnrlistD         = jjnr[jidx+3];
 417             /* Sign of each element will be negative for non-real atoms.
 418              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 419              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 420              */
 421             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 422             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 423             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 424             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 425             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 426             j_coord_offsetA  = DIM*jnrA;
 427             j_coord_offsetB  = DIM*jnrB;
 428             j_coord_offsetC  = DIM*jnrC;
 429             j_coord_offsetD  = DIM*jnrD;
 430
 431             /* load j atom coordinates */
 432             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 433                                               x+j_coord_offsetC,x+j_coord_offsetD,
 434                                               &jx0,&jy0,&jz0);
 435
 436             /* Calculate displacement vector */
 437             dx00             = _mm_sub_ps(ix0,jx0);
 438             dy00             = _mm_sub_ps(iy0,jy0);
 439             dz00             = _mm_sub_ps(iz0,jz0);
 440             dx10             = _mm_sub_ps(ix1,jx0);
 441             dy10             = _mm_sub_ps(iy1,jy0);
 442             dz10             = _mm_sub_ps(iz1,jz0);
 443             dx20             = _mm_sub_ps(ix2,jx0);
 444             dy20             = _mm_sub_ps(iy2,jy0);
 445             dz20             = _mm_sub_ps(iz2,jz0);
 446
 447             /* Calculate squared distance and things based on it */
 448             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 449             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 450             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 451
 452             rinv00           = avx128fma_invsqrt_f(rsq00);
 453             rinv10           = avx128fma_invsqrt_f(rsq10);
 454             rinv20           = avx128fma_invsqrt_f(rsq20);
 455
 456             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 457             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 458             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 459
 460             /* Load parameters for j particles */
 461             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 462                                                               charge+jnrC+0,charge+jnrD+0);
 463             vdwjidx0A        = 2*vdwtype[jnrA+0];
 464             vdwjidx0B        = 2*vdwtype[jnrB+0];
 465             vdwjidx0C        = 2*vdwtype[jnrC+0];
 466             vdwjidx0D        = 2*vdwtype[jnrD+0];
 467
 468             fjx0             = _mm_setzero_ps();
 469             fjy0             = _mm_setzero_ps();
 470             fjz0             = _mm_setzero_ps();
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             r00              = _mm_mul_ps(rsq00,rinv00);
 477             r00              = _mm_andnot_ps(dummy_mask,r00);
 478
 479             /* Compute parameters for interactions between i and j atoms */
 480             qq00             = _mm_mul_ps(iq0,jq0);
 481             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 482                                          vdwparam+vdwioffset0+vdwjidx0B,
 483                                          vdwparam+vdwioffset0+vdwjidx0C,
 484                                          vdwparam+vdwioffset0+vdwjidx0D,
 485                                          &c6_00,&c12_00);
 486
 487             /* Calculate table index by multiplying r with table scale and truncate to integer */
 488             rt               = _mm_mul_ps(r00,vftabscale);
 489             vfitab           = _mm_cvttps_epi32(rt);
 490 #ifdef __XOP__
 491             vfeps            = _mm_frcz_ps(rt);
 492 #else
 493             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 494 #endif
 495             twovfeps         = _mm_add_ps(vfeps,vfeps);
 496             vfitab           = _mm_slli_epi32(vfitab,3);
 497
 498             /* EWALD ELECTROSTATICS */
 499
 500             /* Analytical PME correction */
 501             zeta2            = _mm_mul_ps(beta2,rsq00);
 502             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 503             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 504             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 505             felec            = _mm_mul_ps(qq00,felec);
 506             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 507             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 508             velec            = _mm_mul_ps(qq00,velec);
 509
 510             /* CUBIC SPLINE TABLE DISPERSION */
 511             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 512             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 513             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 514             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 515             _MM_TRANSPOSE4_PS(Y,F,G,H);
 516             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 517             VV               = _mm_macc_ps(vfeps,Fp,Y);
 518             vvdw6            = _mm_mul_ps(c6_00,VV);
 519             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 520             fvdw6            = _mm_mul_ps(c6_00,FF);
 521
 522             /* CUBIC SPLINE TABLE REPULSION */
 523             vfitab           = _mm_add_epi32(vfitab,ifour);
 524             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 525             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 526             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 527             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 528             _MM_TRANSPOSE4_PS(Y,F,G,H);
 529             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 530             VV               = _mm_macc_ps(vfeps,Fp,Y);
 531             vvdw12           = _mm_mul_ps(c12_00,VV);
 532             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 533             fvdw12           = _mm_mul_ps(c12_00,FF);
 534             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 535             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 536
 537             /* Update potential sum for this i atom from the interaction with this j atom. */
 538             velec            = _mm_andnot_ps(dummy_mask,velec);
 539             velecsum         = _mm_add_ps(velecsum,velec);
 540             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 541             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 542
 543             fscal            = _mm_add_ps(felec,fvdw);
 544
 545             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 546
 547              /* Update vectorial force */
 548             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 549             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 550             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 551
 552             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 553             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 554             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 555
 556             /**************************
 557              * CALCULATE INTERACTIONS *
 558              **************************/
 559
 560             r10              = _mm_mul_ps(rsq10,rinv10);
 561             r10              = _mm_andnot_ps(dummy_mask,r10);
 562
 563             /* Compute parameters for interactions between i and j atoms */
 564             qq10             = _mm_mul_ps(iq1,jq0);
 565
 566             /* EWALD ELECTROSTATICS */
 567
 568             /* Analytical PME correction */
 569             zeta2            = _mm_mul_ps(beta2,rsq10);
 570             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 571             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 572             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 573             felec            = _mm_mul_ps(qq10,felec);
 574             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 575             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv10);
 576             velec            = _mm_mul_ps(qq10,velec);
 577
 578             /* Update potential sum for this i atom from the interaction with this j atom. */
 579             velec            = _mm_andnot_ps(dummy_mask,velec);
 580             velecsum         = _mm_add_ps(velecsum,velec);
 581
 582             fscal            = felec;
 583
 584             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 585
 586              /* Update vectorial force */
 587             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 588             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 589             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 590
 591             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 592             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 593             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             r20              = _mm_mul_ps(rsq20,rinv20);
 600             r20              = _mm_andnot_ps(dummy_mask,r20);
 601
 602             /* Compute parameters for interactions between i and j atoms */
 603             qq20             = _mm_mul_ps(iq2,jq0);
 604
 605             /* EWALD ELECTROSTATICS */
 606
 607             /* Analytical PME correction */
 608             zeta2            = _mm_mul_ps(beta2,rsq20);
 609             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 610             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 611             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 612             felec            = _mm_mul_ps(qq20,felec);
 613             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 614             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv20);
 615             velec            = _mm_mul_ps(qq20,velec);
 616
 617             /* Update potential sum for this i atom from the interaction with this j atom. */
 618             velec            = _mm_andnot_ps(dummy_mask,velec);
 619             velecsum         = _mm_add_ps(velecsum,velec);
 620
 621             fscal            = felec;
 622
 623             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 624
 625              /* Update vectorial force */
 626             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 627             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 628             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 629
 630             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 631             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 632             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 633
 634             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 635             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 636             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 637             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 638
 639             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 640
 641             /* Inner loop uses 124 flops */
 642         }
 643
 644         /* End of innermost loop */
 645
 646         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 647                                               f+i_coord_offset,fshift+i_shift_offset);
 648
 649         ggid                        = gid[iidx];
 650         /* Update potential energies */
 651         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 652         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 653
 654         /* Increment number of inner iterations */
 655         inneriter                  += j_index_end - j_index_start;
 656
 657         /* Outer loop uses 20 flops */
 658     }
 659
 660     /* Increment number of outer iterations */
 661     outeriter        += nri;
 662
 663     /* Update outer/inner flops */
 664
 665     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*124);
 666 }
 667 /*
 668  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwCSTab_GeomW3P1_F_avx_128_fma_single
 669  * Electrostatics interaction: Ewald
 670  * VdW interaction:            CubicSplineTable
 671  * Geometry:                   Water3-Particle
 672  * Calculate force/pot:        Force
 673  */
 674 void
 675 nb_kernel_ElecEw_VdwCSTab_GeomW3P1_F_avx_128_fma_single
 676                     (t_nblist                    * gmx_restrict       nlist,
 677                      rvec                        * gmx_restrict          xx,
 678                      rvec                        * gmx_restrict          ff,
 679                      struct t_forcerec           * gmx_restrict          fr,
 680                      t_mdatoms                   * gmx_restrict     mdatoms,
 681                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 682                      t_nrnb                      * gmx_restrict        nrnb)
 683 {
 684     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 685      * just 0 for non-waters.
 686      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 687      * jnr indices corresponding to data put in the four positions in the SIMD register.
 688      */
 689     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 690     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 691     int              jnrA,jnrB,jnrC,jnrD;
 692     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 693     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 694     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 695     real             rcutoff_scalar;
 696     real             *shiftvec,*fshift,*x,*f;
 697     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 698     real             scratch[4*DIM];
 699     __m128           fscal,rcutoff,rcutoff2,jidxall;
 700     int              vdwioffset0;
 701     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 702     int              vdwioffset1;
 703     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 704     int              vdwioffset2;
 705     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 706     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 707     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 708     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 709     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 710     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 711     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 712     real             *charge;
 713     int              nvdwtype;
 714     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 715     int              *vdwtype;
 716     real             *vdwparam;
 717     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 718     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 719     __m128i          vfitab;
 720     __m128i          ifour       = _mm_set1_epi32(4);
 721     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 722     real             *vftab;
 723     __m128i          ewitab;
 724     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 725     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 726     real             *ewtab;
 727     __m128           dummy_mask,cutoff_mask;
 728     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 729     __m128           one     = _mm_set1_ps(1.0);
 730     __m128           two     = _mm_set1_ps(2.0);
 731     x                = xx[0];
 732     f                = ff[0];
 733
 734     nri              = nlist->nri;
 735     iinr             = nlist->iinr;
 736     jindex           = nlist->jindex;
 737     jjnr             = nlist->jjnr;
 738     shiftidx         = nlist->shift;
 739     gid              = nlist->gid;
 740     shiftvec         = fr->shift_vec[0];
 741     fshift           = fr->fshift[0];
 742     facel            = _mm_set1_ps(fr->ic->epsfac);
 743     charge           = mdatoms->chargeA;
 744     nvdwtype         = fr->ntype;
 745     vdwparam         = fr->nbfp;
 746     vdwtype          = mdatoms->typeA;
 747
 748     vftab            = kernel_data->table_vdw->data;
 749     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 750
 751     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 752     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 753     beta2            = _mm_mul_ps(beta,beta);
 754     beta3            = _mm_mul_ps(beta,beta2);
 755     ewtab            = fr->ic->tabq_coul_F;
 756     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 757     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 758
 759     /* Setup water-specific parameters */
 760     inr              = nlist->iinr[0];
 761     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 762     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 763     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 764     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 765
 766     /* Avoid stupid compiler warnings */
 767     jnrA = jnrB = jnrC = jnrD = 0;
 768     j_coord_offsetA = 0;
 769     j_coord_offsetB = 0;
 770     j_coord_offsetC = 0;
 771     j_coord_offsetD = 0;
 772
 773     outeriter        = 0;
 774     inneriter        = 0;
 775
 776     for(iidx=0;iidx<4*DIM;iidx++)
 777     {
 778         scratch[iidx] = 0.0;
 779     }
 780
 781     /* Start outer loop over neighborlists */
 782     for(iidx=0; iidx<nri; iidx++)
 783     {
 784         /* Load shift vector for this list */
 785         i_shift_offset   = DIM*shiftidx[iidx];
 786
 787         /* Load limits for loop over neighbors */
 788         j_index_start    = jindex[iidx];
 789         j_index_end      = jindex[iidx+1];
 790
 791         /* Get outer coordinate index */
 792         inr              = iinr[iidx];
 793         i_coord_offset   = DIM*inr;
 794
 795         /* Load i particle coords and add shift vector */
 796         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 797                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 798
 799         fix0             = _mm_setzero_ps();
 800         fiy0             = _mm_setzero_ps();
 801         fiz0             = _mm_setzero_ps();
 802         fix1             = _mm_setzero_ps();
 803         fiy1             = _mm_setzero_ps();
 804         fiz1             = _mm_setzero_ps();
 805         fix2             = _mm_setzero_ps();
 806         fiy2             = _mm_setzero_ps();
 807         fiz2             = _mm_setzero_ps();
 808
 809         /* Start inner kernel loop */
 810         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 811         {
 812
 813             /* Get j neighbor index, and coordinate index */
 814             jnrA             = jjnr[jidx];
 815             jnrB             = jjnr[jidx+1];
 816             jnrC             = jjnr[jidx+2];
 817             jnrD             = jjnr[jidx+3];
 818             j_coord_offsetA  = DIM*jnrA;
 819             j_coord_offsetB  = DIM*jnrB;
 820             j_coord_offsetC  = DIM*jnrC;
 821             j_coord_offsetD  = DIM*jnrD;
 822
 823             /* load j atom coordinates */
 824             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 825                                               x+j_coord_offsetC,x+j_coord_offsetD,
 826                                               &jx0,&jy0,&jz0);
 827
 828             /* Calculate displacement vector */
 829             dx00             = _mm_sub_ps(ix0,jx0);
 830             dy00             = _mm_sub_ps(iy0,jy0);
 831             dz00             = _mm_sub_ps(iz0,jz0);
 832             dx10             = _mm_sub_ps(ix1,jx0);
 833             dy10             = _mm_sub_ps(iy1,jy0);
 834             dz10             = _mm_sub_ps(iz1,jz0);
 835             dx20             = _mm_sub_ps(ix2,jx0);
 836             dy20             = _mm_sub_ps(iy2,jy0);
 837             dz20             = _mm_sub_ps(iz2,jz0);
 838
 839             /* Calculate squared distance and things based on it */
 840             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 841             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 842             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 843
 844             rinv00           = avx128fma_invsqrt_f(rsq00);
 845             rinv10           = avx128fma_invsqrt_f(rsq10);
 846             rinv20           = avx128fma_invsqrt_f(rsq20);
 847
 848             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 849             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 850             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 851
 852             /* Load parameters for j particles */
 853             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 854                                                               charge+jnrC+0,charge+jnrD+0);
 855             vdwjidx0A        = 2*vdwtype[jnrA+0];
 856             vdwjidx0B        = 2*vdwtype[jnrB+0];
 857             vdwjidx0C        = 2*vdwtype[jnrC+0];
 858             vdwjidx0D        = 2*vdwtype[jnrD+0];
 859
 860             fjx0             = _mm_setzero_ps();
 861             fjy0             = _mm_setzero_ps();
 862             fjz0             = _mm_setzero_ps();
 863
 864             /**************************
 865              * CALCULATE INTERACTIONS *
 866              **************************/
 867
 868             r00              = _mm_mul_ps(rsq00,rinv00);
 869
 870             /* Compute parameters for interactions between i and j atoms */
 871             qq00             = _mm_mul_ps(iq0,jq0);
 872             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 873                                          vdwparam+vdwioffset0+vdwjidx0B,
 874                                          vdwparam+vdwioffset0+vdwjidx0C,
 875                                          vdwparam+vdwioffset0+vdwjidx0D,
 876                                          &c6_00,&c12_00);
 877
 878             /* Calculate table index by multiplying r with table scale and truncate to integer */
 879             rt               = _mm_mul_ps(r00,vftabscale);
 880             vfitab           = _mm_cvttps_epi32(rt);
 881 #ifdef __XOP__
 882             vfeps            = _mm_frcz_ps(rt);
 883 #else
 884             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 885 #endif
 886             twovfeps         = _mm_add_ps(vfeps,vfeps);
 887             vfitab           = _mm_slli_epi32(vfitab,3);
 888
 889             /* EWALD ELECTROSTATICS */
 890
 891             /* Analytical PME correction */
 892             zeta2            = _mm_mul_ps(beta2,rsq00);
 893             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 894             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 895             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 896             felec            = _mm_mul_ps(qq00,felec);
 897
 898             /* CUBIC SPLINE TABLE DISPERSION */
 899             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 900             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 901             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 902             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 903             _MM_TRANSPOSE4_PS(Y,F,G,H);
 904             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 905             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 906             fvdw6            = _mm_mul_ps(c6_00,FF);
 907
 908             /* CUBIC SPLINE TABLE REPULSION */
 909             vfitab           = _mm_add_epi32(vfitab,ifour);
 910             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 911             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 912             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 913             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 914             _MM_TRANSPOSE4_PS(Y,F,G,H);
 915             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 916             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 917             fvdw12           = _mm_mul_ps(c12_00,FF);
 918             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 919
 920             fscal            = _mm_add_ps(felec,fvdw);
 921
 922              /* Update vectorial force */
 923             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 924             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 925             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 926
 927             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 928             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 929             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 930
 931             /**************************
 932              * CALCULATE INTERACTIONS *
 933              **************************/
 934
 935             r10              = _mm_mul_ps(rsq10,rinv10);
 936
 937             /* Compute parameters for interactions between i and j atoms */
 938             qq10             = _mm_mul_ps(iq1,jq0);
 939
 940             /* EWALD ELECTROSTATICS */
 941
 942             /* Analytical PME correction */
 943             zeta2            = _mm_mul_ps(beta2,rsq10);
 944             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 945             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 946             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 947             felec            = _mm_mul_ps(qq10,felec);
 948
 949             fscal            = felec;
 950
 951              /* Update vectorial force */
 952             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 953             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 954             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 955
 956             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 957             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 958             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 959
 960             /**************************
 961              * CALCULATE INTERACTIONS *
 962              **************************/
 963
 964             r20              = _mm_mul_ps(rsq20,rinv20);
 965
 966             /* Compute parameters for interactions between i and j atoms */
 967             qq20             = _mm_mul_ps(iq2,jq0);
 968
 969             /* EWALD ELECTROSTATICS */
 970
 971             /* Analytical PME correction */
 972             zeta2            = _mm_mul_ps(beta2,rsq20);
 973             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 974             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 975             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 976             felec            = _mm_mul_ps(qq20,felec);
 977
 978             fscal            = felec;
 979
 980              /* Update vectorial force */
 981             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 982             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 983             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 984
 985             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 986             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 987             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 988
 989             fjptrA             = f+j_coord_offsetA;
 990             fjptrB             = f+j_coord_offsetB;
 991             fjptrC             = f+j_coord_offsetC;
 992             fjptrD             = f+j_coord_offsetD;
 993
 994             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 995
 996             /* Inner loop uses 110 flops */
 997         }
 998
 999         if(jidx<j_index_end)
1000         {
1001
1002             /* Get j neighbor index, and coordinate index */
1003             jnrlistA         = jjnr[jidx];
1004             jnrlistB         = jjnr[jidx+1];
1005             jnrlistC         = jjnr[jidx+2];
1006             jnrlistD         = jjnr[jidx+3];
1007             /* Sign of each element will be negative for non-real atoms.
1008              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1009              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1010              */
1011             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1012             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1013             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1014             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1015             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1016             j_coord_offsetA  = DIM*jnrA;
1017             j_coord_offsetB  = DIM*jnrB;
1018             j_coord_offsetC  = DIM*jnrC;
1019             j_coord_offsetD  = DIM*jnrD;
1020
1021             /* load j atom coordinates */
1022             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1023                                               x+j_coord_offsetC,x+j_coord_offsetD,
1024                                               &jx0,&jy0,&jz0);
1025
1026             /* Calculate displacement vector */
1027             dx00             = _mm_sub_ps(ix0,jx0);
1028             dy00             = _mm_sub_ps(iy0,jy0);
1029             dz00             = _mm_sub_ps(iz0,jz0);
1030             dx10             = _mm_sub_ps(ix1,jx0);
1031             dy10             = _mm_sub_ps(iy1,jy0);
1032             dz10             = _mm_sub_ps(iz1,jz0);
1033             dx20             = _mm_sub_ps(ix2,jx0);
1034             dy20             = _mm_sub_ps(iy2,jy0);
1035             dz20             = _mm_sub_ps(iz2,jz0);
1036
1037             /* Calculate squared distance and things based on it */
1038             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1039             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1040             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1041
1042             rinv00           = avx128fma_invsqrt_f(rsq00);
1043             rinv10           = avx128fma_invsqrt_f(rsq10);
1044             rinv20           = avx128fma_invsqrt_f(rsq20);
1045
1046             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1047             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1048             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1049
1050             /* Load parameters for j particles */
1051             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1052                                                               charge+jnrC+0,charge+jnrD+0);
1053             vdwjidx0A        = 2*vdwtype[jnrA+0];
1054             vdwjidx0B        = 2*vdwtype[jnrB+0];
1055             vdwjidx0C        = 2*vdwtype[jnrC+0];
1056             vdwjidx0D        = 2*vdwtype[jnrD+0];
1057
1058             fjx0             = _mm_setzero_ps();
1059             fjy0             = _mm_setzero_ps();
1060             fjz0             = _mm_setzero_ps();
1061
1062             /**************************
1063              * CALCULATE INTERACTIONS *
1064              **************************/
1065
1066             r00              = _mm_mul_ps(rsq00,rinv00);
1067             r00              = _mm_andnot_ps(dummy_mask,r00);
1068
1069             /* Compute parameters for interactions between i and j atoms */
1070             qq00             = _mm_mul_ps(iq0,jq0);
1071             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1072                                          vdwparam+vdwioffset0+vdwjidx0B,
1073                                          vdwparam+vdwioffset0+vdwjidx0C,
1074                                          vdwparam+vdwioffset0+vdwjidx0D,
1075                                          &c6_00,&c12_00);
1076
1077             /* Calculate table index by multiplying r with table scale and truncate to integer */
1078             rt               = _mm_mul_ps(r00,vftabscale);
1079             vfitab           = _mm_cvttps_epi32(rt);
1080 #ifdef __XOP__
1081             vfeps            = _mm_frcz_ps(rt);
1082 #else
1083             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1084 #endif
1085             twovfeps         = _mm_add_ps(vfeps,vfeps);
1086             vfitab           = _mm_slli_epi32(vfitab,3);
1087
1088             /* EWALD ELECTROSTATICS */
1089
1090             /* Analytical PME correction */
1091             zeta2            = _mm_mul_ps(beta2,rsq00);
1092             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1093             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1094             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1095             felec            = _mm_mul_ps(qq00,felec);
1096
1097             /* CUBIC SPLINE TABLE DISPERSION */
1098             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1099             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1100             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1101             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1102             _MM_TRANSPOSE4_PS(Y,F,G,H);
1103             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1104             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1105             fvdw6            = _mm_mul_ps(c6_00,FF);
1106
1107             /* CUBIC SPLINE TABLE REPULSION */
1108             vfitab           = _mm_add_epi32(vfitab,ifour);
1109             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1110             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1111             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1112             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1113             _MM_TRANSPOSE4_PS(Y,F,G,H);
1114             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1115             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1116             fvdw12           = _mm_mul_ps(c12_00,FF);
1117             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1118
1119             fscal            = _mm_add_ps(felec,fvdw);
1120
1121             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1122
1123              /* Update vectorial force */
1124             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1125             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1126             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1127
1128             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1129             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1130             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1131
1132             /**************************
1133              * CALCULATE INTERACTIONS *
1134              **************************/
1135
1136             r10              = _mm_mul_ps(rsq10,rinv10);
1137             r10              = _mm_andnot_ps(dummy_mask,r10);
1138
1139             /* Compute parameters for interactions between i and j atoms */
1140             qq10             = _mm_mul_ps(iq1,jq0);
1141
1142             /* EWALD ELECTROSTATICS */
1143
1144             /* Analytical PME correction */
1145             zeta2            = _mm_mul_ps(beta2,rsq10);
1146             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1147             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1148             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1149             felec            = _mm_mul_ps(qq10,felec);
1150
1151             fscal            = felec;
1152
1153             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1154
1155              /* Update vectorial force */
1156             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1157             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1158             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1159
1160             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1161             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1162             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1163
1164             /**************************
1165              * CALCULATE INTERACTIONS *
1166              **************************/
1167
1168             r20              = _mm_mul_ps(rsq20,rinv20);
1169             r20              = _mm_andnot_ps(dummy_mask,r20);
1170
1171             /* Compute parameters for interactions between i and j atoms */
1172             qq20             = _mm_mul_ps(iq2,jq0);
1173
1174             /* EWALD ELECTROSTATICS */
1175
1176             /* Analytical PME correction */
1177             zeta2            = _mm_mul_ps(beta2,rsq20);
1178             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1179             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
1180             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1181             felec            = _mm_mul_ps(qq20,felec);
1182
1183             fscal            = felec;
1184
1185             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1186
1187              /* Update vectorial force */
1188             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1189             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1190             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1191
1192             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1193             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1194             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1195
1196             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1197             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1198             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1199             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1200
1201             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1202
1203             /* Inner loop uses 113 flops */
1204         }
1205
1206         /* End of innermost loop */
1207
1208         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1209                                               f+i_coord_offset,fshift+i_shift_offset);
1210
1211         /* Increment number of inner iterations */
1212         inneriter                  += j_index_end - j_index_start;
1213
1214         /* Outer loop uses 18 flops */
1215     }
1216
1217     /* Increment number of outer iterations */
1218     outeriter        += nri;
1219
1220     /* Update outer/inner flops */
1221
1222     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*113);
1223 }