src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEw_VdwLJEw_GeomW3P1_avx_128_fma_double.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_double 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_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomW3P1_VF_avx_128_fma_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Water3-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEw_VdwLJEw_GeomW3P1_VF_avx_128_fma_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  89     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  90     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  97     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  98     __m128d           c6grid_00;
  99     __m128d           c6grid_10;
 100     __m128d           c6grid_20;
 101     real             *vdwgridparam;
 102     __m128d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 103     __m128d           one_half  = _mm_set1_pd(0.5);
 104     __m128d           minus_one = _mm_set1_pd(-1.0);
 105     __m128i          ewitab;
 106     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 107     real             *ewtab;
 108     __m128d          dummy_mask,cutoff_mask;
 109     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 110     __m128d          one     = _mm_set1_pd(1.0);
 111     __m128d          two     = _mm_set1_pd(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_pd(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     nvdwtype         = fr->ntype;
 126     vdwparam         = fr->nbfp;
 127     vdwtype          = mdatoms->typeA;
 128     vdwgridparam     = fr->ljpme_c6grid;
 129     sh_lj_ewald      = _mm_set1_pd(fr->ic->sh_lj_ewald);
 130     ewclj            = _mm_set1_pd(fr->ic->ewaldcoeff_lj);
 131     ewclj2           = _mm_mul_pd(minus_one,_mm_mul_pd(ewclj,ewclj));
 132
 133     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 141     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 142     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     /* Avoid stupid compiler warnings */
 146     jnrA = jnrB = 0;
 147     j_coord_offsetA = 0;
 148     j_coord_offsetB = 0;
 149
 150     outeriter        = 0;
 151     inneriter        = 0;
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 169                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 170
 171         fix0             = _mm_setzero_pd();
 172         fiy0             = _mm_setzero_pd();
 173         fiz0             = _mm_setzero_pd();
 174         fix1             = _mm_setzero_pd();
 175         fiy1             = _mm_setzero_pd();
 176         fiz1             = _mm_setzero_pd();
 177         fix2             = _mm_setzero_pd();
 178         fiy2             = _mm_setzero_pd();
 179         fiz2             = _mm_setzero_pd();
 180
 181         /* Reset potential sums */
 182         velecsum         = _mm_setzero_pd();
 183         vvdwsum          = _mm_setzero_pd();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194
 195             /* load j atom coordinates */
 196             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                               &jx0,&jy0,&jz0);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _mm_sub_pd(ix0,jx0);
 201             dy00             = _mm_sub_pd(iy0,jy0);
 202             dz00             = _mm_sub_pd(iz0,jz0);
 203             dx10             = _mm_sub_pd(ix1,jx0);
 204             dy10             = _mm_sub_pd(iy1,jy0);
 205             dz10             = _mm_sub_pd(iz1,jz0);
 206             dx20             = _mm_sub_pd(ix2,jx0);
 207             dy20             = _mm_sub_pd(iy2,jy0);
 208             dz20             = _mm_sub_pd(iz2,jz0);
 209
 210             /* Calculate squared distance and things based on it */
 211             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 212             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 213             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 214
 215             rinv00           = avx128fma_invsqrt_d(rsq00);
 216             rinv10           = avx128fma_invsqrt_d(rsq10);
 217             rinv20           = avx128fma_invsqrt_d(rsq20);
 218
 219             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 220             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 221             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 222
 223             /* Load parameters for j particles */
 224             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 225             vdwjidx0A        = 2*vdwtype[jnrA+0];
 226             vdwjidx0B        = 2*vdwtype[jnrB+0];
 227
 228             fjx0             = _mm_setzero_pd();
 229             fjy0             = _mm_setzero_pd();
 230             fjz0             = _mm_setzero_pd();
 231
 232             /**************************
 233              * CALCULATE INTERACTIONS *
 234              **************************/
 235
 236             r00              = _mm_mul_pd(rsq00,rinv00);
 237
 238             /* Compute parameters for interactions between i and j atoms */
 239             qq00             = _mm_mul_pd(iq0,jq0);
 240             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 241                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 242             c6grid_00       = gmx_mm_load_2real_swizzle_pd(vdwgridparam+vdwioffset0+vdwjidx0A,
 243                                                                vdwgridparam+vdwioffset0+vdwjidx0B);
 244
 245             /* EWALD ELECTROSTATICS */
 246
 247             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 248             ewrt             = _mm_mul_pd(r00,ewtabscale);
 249             ewitab           = _mm_cvttpd_epi32(ewrt);
 250 #ifdef __XOP__
 251             eweps            = _mm_frcz_pd(ewrt);
 252 #else
 253             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 254 #endif
 255             twoeweps         = _mm_add_pd(eweps,eweps);
 256             ewitab           = _mm_slli_epi32(ewitab,2);
 257             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 258             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 259             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 260             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 261             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 262             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 263             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 264             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 265             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 266             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 267
 268             /* Analytical LJ-PME */
 269             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 270             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 271             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 272             exponent         = avx128fma_exp_d(ewcljrsq);
 273             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 274             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 275             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 276             vvdw6            = _mm_mul_pd(_mm_macc_pd(-c6grid_00,_mm_sub_pd(one,poly),c6_00),rinvsix);
 277             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 278             vvdw             = _mm_msub_pd(vvdw12,one_twelfth,_mm_mul_pd(vvdw6,one_sixth));
 279             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 280             fvdw             = _mm_mul_pd(_mm_add_pd(vvdw12,_mm_msub_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6),vvdw6)),rinvsq00);
 281
 282             /* Update potential sum for this i atom from the interaction with this j atom. */
 283             velecsum         = _mm_add_pd(velecsum,velec);
 284             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 285
 286             fscal            = _mm_add_pd(felec,fvdw);
 287
 288             /* Update vectorial force */
 289             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 290             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 291             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 292
 293             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 294             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 295             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             r10              = _mm_mul_pd(rsq10,rinv10);
 302
 303             /* Compute parameters for interactions between i and j atoms */
 304             qq10             = _mm_mul_pd(iq1,jq0);
 305
 306             /* EWALD ELECTROSTATICS */
 307
 308             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 309             ewrt             = _mm_mul_pd(r10,ewtabscale);
 310             ewitab           = _mm_cvttpd_epi32(ewrt);
 311 #ifdef __XOP__
 312             eweps            = _mm_frcz_pd(ewrt);
 313 #else
 314             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 315 #endif
 316             twoeweps         = _mm_add_pd(eweps,eweps);
 317             ewitab           = _mm_slli_epi32(ewitab,2);
 318             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 319             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 320             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 321             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 322             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 323             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 324             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 325             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 326             velec            = _mm_mul_pd(qq10,_mm_sub_pd(rinv10,velec));
 327             felec            = _mm_mul_pd(_mm_mul_pd(qq10,rinv10),_mm_sub_pd(rinvsq10,felec));
 328
 329             /* Update potential sum for this i atom from the interaction with this j atom. */
 330             velecsum         = _mm_add_pd(velecsum,velec);
 331
 332             fscal            = felec;
 333
 334             /* Update vectorial force */
 335             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 336             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 337             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 338
 339             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 340             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 341             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 342
 343             /**************************
 344              * CALCULATE INTERACTIONS *
 345              **************************/
 346
 347             r20              = _mm_mul_pd(rsq20,rinv20);
 348
 349             /* Compute parameters for interactions between i and j atoms */
 350             qq20             = _mm_mul_pd(iq2,jq0);
 351
 352             /* EWALD ELECTROSTATICS */
 353
 354             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 355             ewrt             = _mm_mul_pd(r20,ewtabscale);
 356             ewitab           = _mm_cvttpd_epi32(ewrt);
 357 #ifdef __XOP__
 358             eweps            = _mm_frcz_pd(ewrt);
 359 #else
 360             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 361 #endif
 362             twoeweps         = _mm_add_pd(eweps,eweps);
 363             ewitab           = _mm_slli_epi32(ewitab,2);
 364             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 365             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 366             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 367             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 368             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 369             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 370             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 371             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 372             velec            = _mm_mul_pd(qq20,_mm_sub_pd(rinv20,velec));
 373             felec            = _mm_mul_pd(_mm_mul_pd(qq20,rinv20),_mm_sub_pd(rinvsq20,felec));
 374
 375             /* Update potential sum for this i atom from the interaction with this j atom. */
 376             velecsum         = _mm_add_pd(velecsum,velec);
 377
 378             fscal            = felec;
 379
 380             /* Update vectorial force */
 381             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 382             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 383             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 384
 385             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 386             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 387             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 388
 389             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 390
 391             /* Inner loop uses 159 flops */
 392         }
 393
 394         if(jidx<j_index_end)
 395         {
 396
 397             jnrA             = jjnr[jidx];
 398             j_coord_offsetA  = DIM*jnrA;
 399
 400             /* load j atom coordinates */
 401             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 402                                               &jx0,&jy0,&jz0);
 403
 404             /* Calculate displacement vector */
 405             dx00             = _mm_sub_pd(ix0,jx0);
 406             dy00             = _mm_sub_pd(iy0,jy0);
 407             dz00             = _mm_sub_pd(iz0,jz0);
 408             dx10             = _mm_sub_pd(ix1,jx0);
 409             dy10             = _mm_sub_pd(iy1,jy0);
 410             dz10             = _mm_sub_pd(iz1,jz0);
 411             dx20             = _mm_sub_pd(ix2,jx0);
 412             dy20             = _mm_sub_pd(iy2,jy0);
 413             dz20             = _mm_sub_pd(iz2,jz0);
 414
 415             /* Calculate squared distance and things based on it */
 416             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 417             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 418             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 419
 420             rinv00           = avx128fma_invsqrt_d(rsq00);
 421             rinv10           = avx128fma_invsqrt_d(rsq10);
 422             rinv20           = avx128fma_invsqrt_d(rsq20);
 423
 424             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 425             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 426             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 427
 428             /* Load parameters for j particles */
 429             jq0              = _mm_load_sd(charge+jnrA+0);
 430             vdwjidx0A        = 2*vdwtype[jnrA+0];
 431
 432             fjx0             = _mm_setzero_pd();
 433             fjy0             = _mm_setzero_pd();
 434             fjz0             = _mm_setzero_pd();
 435
 436             /**************************
 437              * CALCULATE INTERACTIONS *
 438              **************************/
 439
 440             r00              = _mm_mul_pd(rsq00,rinv00);
 441
 442             /* Compute parameters for interactions between i and j atoms */
 443             qq00             = _mm_mul_pd(iq0,jq0);
 444             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 445             c6grid_00       = gmx_mm_load_1real_pd(vdwgridparam+vdwioffset0+vdwjidx0A);
 446
 447             /* EWALD ELECTROSTATICS */
 448
 449             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 450             ewrt             = _mm_mul_pd(r00,ewtabscale);
 451             ewitab           = _mm_cvttpd_epi32(ewrt);
 452 #ifdef __XOP__
 453             eweps            = _mm_frcz_pd(ewrt);
 454 #else
 455             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 456 #endif
 457             twoeweps         = _mm_add_pd(eweps,eweps);
 458             ewitab           = _mm_slli_epi32(ewitab,2);
 459             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 460             ewtabD           = _mm_setzero_pd();
 461             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 462             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 463             ewtabFn          = _mm_setzero_pd();
 464             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 465             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 466             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 467             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 468             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 469
 470             /* Analytical LJ-PME */
 471             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 472             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 473             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 474             exponent         = avx128fma_exp_d(ewcljrsq);
 475             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 476             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 477             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 478             vvdw6            = _mm_mul_pd(_mm_macc_pd(-c6grid_00,_mm_sub_pd(one,poly),c6_00),rinvsix);
 479             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 480             vvdw             = _mm_msub_pd(vvdw12,one_twelfth,_mm_mul_pd(vvdw6,one_sixth));
 481             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 482             fvdw             = _mm_mul_pd(_mm_add_pd(vvdw12,_mm_msub_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6),vvdw6)),rinvsq00);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 486             velecsum         = _mm_add_pd(velecsum,velec);
 487             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 488             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 489
 490             fscal            = _mm_add_pd(felec,fvdw);
 491
 492             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 493
 494             /* Update vectorial force */
 495             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 496             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 497             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 498
 499             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 500             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 501             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 502
 503             /**************************
 504              * CALCULATE INTERACTIONS *
 505              **************************/
 506
 507             r10              = _mm_mul_pd(rsq10,rinv10);
 508
 509             /* Compute parameters for interactions between i and j atoms */
 510             qq10             = _mm_mul_pd(iq1,jq0);
 511
 512             /* EWALD ELECTROSTATICS */
 513
 514             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 515             ewrt             = _mm_mul_pd(r10,ewtabscale);
 516             ewitab           = _mm_cvttpd_epi32(ewrt);
 517 #ifdef __XOP__
 518             eweps            = _mm_frcz_pd(ewrt);
 519 #else
 520             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 521 #endif
 522             twoeweps         = _mm_add_pd(eweps,eweps);
 523             ewitab           = _mm_slli_epi32(ewitab,2);
 524             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 525             ewtabD           = _mm_setzero_pd();
 526             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 527             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 528             ewtabFn          = _mm_setzero_pd();
 529             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 530             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 531             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 532             velec            = _mm_mul_pd(qq10,_mm_sub_pd(rinv10,velec));
 533             felec            = _mm_mul_pd(_mm_mul_pd(qq10,rinv10),_mm_sub_pd(rinvsq10,felec));
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 537             velecsum         = _mm_add_pd(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 542
 543             /* Update vectorial force */
 544             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 545             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 546             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 547
 548             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 549             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 550             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 551
 552             /**************************
 553              * CALCULATE INTERACTIONS *
 554              **************************/
 555
 556             r20              = _mm_mul_pd(rsq20,rinv20);
 557
 558             /* Compute parameters for interactions between i and j atoms */
 559             qq20             = _mm_mul_pd(iq2,jq0);
 560
 561             /* EWALD ELECTROSTATICS */
 562
 563             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 564             ewrt             = _mm_mul_pd(r20,ewtabscale);
 565             ewitab           = _mm_cvttpd_epi32(ewrt);
 566 #ifdef __XOP__
 567             eweps            = _mm_frcz_pd(ewrt);
 568 #else
 569             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 570 #endif
 571             twoeweps         = _mm_add_pd(eweps,eweps);
 572             ewitab           = _mm_slli_epi32(ewitab,2);
 573             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 574             ewtabD           = _mm_setzero_pd();
 575             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 576             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 577             ewtabFn          = _mm_setzero_pd();
 578             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 579             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 580             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 581             velec            = _mm_mul_pd(qq20,_mm_sub_pd(rinv20,velec));
 582             felec            = _mm_mul_pd(_mm_mul_pd(qq20,rinv20),_mm_sub_pd(rinvsq20,felec));
 583
 584             /* Update potential sum for this i atom from the interaction with this j atom. */
 585             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 586             velecsum         = _mm_add_pd(velecsum,velec);
 587
 588             fscal            = felec;
 589
 590             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 591
 592             /* Update vectorial force */
 593             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 594             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 595             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 596
 597             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 598             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 599             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 600
 601             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 602
 603             /* Inner loop uses 159 flops */
 604         }
 605
 606         /* End of innermost loop */
 607
 608         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 609                                               f+i_coord_offset,fshift+i_shift_offset);
 610
 611         ggid                        = gid[iidx];
 612         /* Update potential energies */
 613         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 614         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 615
 616         /* Increment number of inner iterations */
 617         inneriter                  += j_index_end - j_index_start;
 618
 619         /* Outer loop uses 20 flops */
 620     }
 621
 622     /* Increment number of outer iterations */
 623     outeriter        += nri;
 624
 625     /* Update outer/inner flops */
 626
 627     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*159);
 628 }
 629 /*
 630  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJEw_GeomW3P1_F_avx_128_fma_double
 631  * Electrostatics interaction: Ewald
 632  * VdW interaction:            LJEwald
 633  * Geometry:                   Water3-Particle
 634  * Calculate force/pot:        Force
 635  */
 636 void
 637 nb_kernel_ElecEw_VdwLJEw_GeomW3P1_F_avx_128_fma_double
 638                     (t_nblist                    * gmx_restrict       nlist,
 639                      rvec                        * gmx_restrict          xx,
 640                      rvec                        * gmx_restrict          ff,
 641                      struct t_forcerec           * gmx_restrict          fr,
 642                      t_mdatoms                   * gmx_restrict     mdatoms,
 643                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 644                      t_nrnb                      * gmx_restrict        nrnb)
 645 {
 646     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 647      * just 0 for non-waters.
 648      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 649      * jnr indices corresponding to data put in the four positions in the SIMD register.
 650      */
 651     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 652     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 653     int              jnrA,jnrB;
 654     int              j_coord_offsetA,j_coord_offsetB;
 655     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 656     real             rcutoff_scalar;
 657     real             *shiftvec,*fshift,*x,*f;
 658     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 659     int              vdwioffset0;
 660     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 661     int              vdwioffset1;
 662     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 663     int              vdwioffset2;
 664     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 665     int              vdwjidx0A,vdwjidx0B;
 666     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 667     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 668     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 669     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 670     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 671     real             *charge;
 672     int              nvdwtype;
 673     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 674     int              *vdwtype;
 675     real             *vdwparam;
 676     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 677     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 678     __m128d           c6grid_00;
 679     __m128d           c6grid_10;
 680     __m128d           c6grid_20;
 681     real             *vdwgridparam;
 682     __m128d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 683     __m128d           one_half  = _mm_set1_pd(0.5);
 684     __m128d           minus_one = _mm_set1_pd(-1.0);
 685     __m128i          ewitab;
 686     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 687     real             *ewtab;
 688     __m128d          dummy_mask,cutoff_mask;
 689     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 690     __m128d          one     = _mm_set1_pd(1.0);
 691     __m128d          two     = _mm_set1_pd(2.0);
 692     x                = xx[0];
 693     f                = ff[0];
 694
 695     nri              = nlist->nri;
 696     iinr             = nlist->iinr;
 697     jindex           = nlist->jindex;
 698     jjnr             = nlist->jjnr;
 699     shiftidx         = nlist->shift;
 700     gid              = nlist->gid;
 701     shiftvec         = fr->shift_vec[0];
 702     fshift           = fr->fshift[0];
 703     facel            = _mm_set1_pd(fr->ic->epsfac);
 704     charge           = mdatoms->chargeA;
 705     nvdwtype         = fr->ntype;
 706     vdwparam         = fr->nbfp;
 707     vdwtype          = mdatoms->typeA;
 708     vdwgridparam     = fr->ljpme_c6grid;
 709     sh_lj_ewald      = _mm_set1_pd(fr->ic->sh_lj_ewald);
 710     ewclj            = _mm_set1_pd(fr->ic->ewaldcoeff_lj);
 711     ewclj2           = _mm_mul_pd(minus_one,_mm_mul_pd(ewclj,ewclj));
 712
 713     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 714     ewtab            = fr->ic->tabq_coul_F;
 715     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 716     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 717
 718     /* Setup water-specific parameters */
 719     inr              = nlist->iinr[0];
 720     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 721     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 722     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 723     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 724
 725     /* Avoid stupid compiler warnings */
 726     jnrA = jnrB = 0;
 727     j_coord_offsetA = 0;
 728     j_coord_offsetB = 0;
 729
 730     outeriter        = 0;
 731     inneriter        = 0;
 732
 733     /* Start outer loop over neighborlists */
 734     for(iidx=0; iidx<nri; iidx++)
 735     {
 736         /* Load shift vector for this list */
 737         i_shift_offset   = DIM*shiftidx[iidx];
 738
 739         /* Load limits for loop over neighbors */
 740         j_index_start    = jindex[iidx];
 741         j_index_end      = jindex[iidx+1];
 742
 743         /* Get outer coordinate index */
 744         inr              = iinr[iidx];
 745         i_coord_offset   = DIM*inr;
 746
 747         /* Load i particle coords and add shift vector */
 748         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 749                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 750
 751         fix0             = _mm_setzero_pd();
 752         fiy0             = _mm_setzero_pd();
 753         fiz0             = _mm_setzero_pd();
 754         fix1             = _mm_setzero_pd();
 755         fiy1             = _mm_setzero_pd();
 756         fiz1             = _mm_setzero_pd();
 757         fix2             = _mm_setzero_pd();
 758         fiy2             = _mm_setzero_pd();
 759         fiz2             = _mm_setzero_pd();
 760
 761         /* Start inner kernel loop */
 762         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 763         {
 764
 765             /* Get j neighbor index, and coordinate index */
 766             jnrA             = jjnr[jidx];
 767             jnrB             = jjnr[jidx+1];
 768             j_coord_offsetA  = DIM*jnrA;
 769             j_coord_offsetB  = DIM*jnrB;
 770
 771             /* load j atom coordinates */
 772             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 773                                               &jx0,&jy0,&jz0);
 774
 775             /* Calculate displacement vector */
 776             dx00             = _mm_sub_pd(ix0,jx0);
 777             dy00             = _mm_sub_pd(iy0,jy0);
 778             dz00             = _mm_sub_pd(iz0,jz0);
 779             dx10             = _mm_sub_pd(ix1,jx0);
 780             dy10             = _mm_sub_pd(iy1,jy0);
 781             dz10             = _mm_sub_pd(iz1,jz0);
 782             dx20             = _mm_sub_pd(ix2,jx0);
 783             dy20             = _mm_sub_pd(iy2,jy0);
 784             dz20             = _mm_sub_pd(iz2,jz0);
 785
 786             /* Calculate squared distance and things based on it */
 787             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 788             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 789             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 790
 791             rinv00           = avx128fma_invsqrt_d(rsq00);
 792             rinv10           = avx128fma_invsqrt_d(rsq10);
 793             rinv20           = avx128fma_invsqrt_d(rsq20);
 794
 795             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 796             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 797             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 798
 799             /* Load parameters for j particles */
 800             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 801             vdwjidx0A        = 2*vdwtype[jnrA+0];
 802             vdwjidx0B        = 2*vdwtype[jnrB+0];
 803
 804             fjx0             = _mm_setzero_pd();
 805             fjy0             = _mm_setzero_pd();
 806             fjz0             = _mm_setzero_pd();
 807
 808             /**************************
 809              * CALCULATE INTERACTIONS *
 810              **************************/
 811
 812             r00              = _mm_mul_pd(rsq00,rinv00);
 813
 814             /* Compute parameters for interactions between i and j atoms */
 815             qq00             = _mm_mul_pd(iq0,jq0);
 816             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 817                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 818             c6grid_00       = gmx_mm_load_2real_swizzle_pd(vdwgridparam+vdwioffset0+vdwjidx0A,
 819                                                                vdwgridparam+vdwioffset0+vdwjidx0B);
 820
 821             /* EWALD ELECTROSTATICS */
 822
 823             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 824             ewrt             = _mm_mul_pd(r00,ewtabscale);
 825             ewitab           = _mm_cvttpd_epi32(ewrt);
 826 #ifdef __XOP__
 827             eweps            = _mm_frcz_pd(ewrt);
 828 #else
 829             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 830 #endif
 831             twoeweps         = _mm_add_pd(eweps,eweps);
 832             gmx_mm_load_2pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),ewtab+_mm_extract_epi32(ewitab,1),
 833                                          &ewtabF,&ewtabFn);
 834             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 835             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 836
 837             /* Analytical LJ-PME */
 838             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 839             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 840             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 841             exponent         = avx128fma_exp_d(ewcljrsq);
 842             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 843             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 844             /* f6A = 6 * C6grid * (1 - poly) */
 845             f6A              = _mm_mul_pd(c6grid_00,_mm_sub_pd(one,poly));
 846             /* f6B = C6grid * exponent * beta^6 */
 847             f6B              = _mm_mul_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6));
 848             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 849             fvdw              = _mm_mul_pd(_mm_macc_pd(_mm_msub_pd(c12_00,rinvsix,_mm_sub_pd(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 850
 851             fscal            = _mm_add_pd(felec,fvdw);
 852
 853             /* Update vectorial force */
 854             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 855             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 856             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 857
 858             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 859             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 860             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 861
 862             /**************************
 863              * CALCULATE INTERACTIONS *
 864              **************************/
 865
 866             r10              = _mm_mul_pd(rsq10,rinv10);
 867
 868             /* Compute parameters for interactions between i and j atoms */
 869             qq10             = _mm_mul_pd(iq1,jq0);
 870
 871             /* EWALD ELECTROSTATICS */
 872
 873             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 874             ewrt             = _mm_mul_pd(r10,ewtabscale);
 875             ewitab           = _mm_cvttpd_epi32(ewrt);
 876 #ifdef __XOP__
 877             eweps            = _mm_frcz_pd(ewrt);
 878 #else
 879             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 880 #endif
 881             twoeweps         = _mm_add_pd(eweps,eweps);
 882             gmx_mm_load_2pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),ewtab+_mm_extract_epi32(ewitab,1),
 883                                          &ewtabF,&ewtabFn);
 884             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 885             felec            = _mm_mul_pd(_mm_mul_pd(qq10,rinv10),_mm_sub_pd(rinvsq10,felec));
 886
 887             fscal            = felec;
 888
 889             /* Update vectorial force */
 890             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 891             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 892             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 893
 894             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 895             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 896             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 897
 898             /**************************
 899              * CALCULATE INTERACTIONS *
 900              **************************/
 901
 902             r20              = _mm_mul_pd(rsq20,rinv20);
 903
 904             /* Compute parameters for interactions between i and j atoms */
 905             qq20             = _mm_mul_pd(iq2,jq0);
 906
 907             /* EWALD ELECTROSTATICS */
 908
 909             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 910             ewrt             = _mm_mul_pd(r20,ewtabscale);
 911             ewitab           = _mm_cvttpd_epi32(ewrt);
 912 #ifdef __XOP__
 913             eweps            = _mm_frcz_pd(ewrt);
 914 #else
 915             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 916 #endif
 917             twoeweps         = _mm_add_pd(eweps,eweps);
 918             gmx_mm_load_2pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),ewtab+_mm_extract_epi32(ewitab,1),
 919                                          &ewtabF,&ewtabFn);
 920             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 921             felec            = _mm_mul_pd(_mm_mul_pd(qq20,rinv20),_mm_sub_pd(rinvsq20,felec));
 922
 923             fscal            = felec;
 924
 925             /* Update vectorial force */
 926             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 927             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 928             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 929
 930             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 931             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 932             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 933
 934             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 935
 936             /* Inner loop uses 141 flops */
 937         }
 938
 939         if(jidx<j_index_end)
 940         {
 941
 942             jnrA             = jjnr[jidx];
 943             j_coord_offsetA  = DIM*jnrA;
 944
 945             /* load j atom coordinates */
 946             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 947                                               &jx0,&jy0,&jz0);
 948
 949             /* Calculate displacement vector */
 950             dx00             = _mm_sub_pd(ix0,jx0);
 951             dy00             = _mm_sub_pd(iy0,jy0);
 952             dz00             = _mm_sub_pd(iz0,jz0);
 953             dx10             = _mm_sub_pd(ix1,jx0);
 954             dy10             = _mm_sub_pd(iy1,jy0);
 955             dz10             = _mm_sub_pd(iz1,jz0);
 956             dx20             = _mm_sub_pd(ix2,jx0);
 957             dy20             = _mm_sub_pd(iy2,jy0);
 958             dz20             = _mm_sub_pd(iz2,jz0);
 959
 960             /* Calculate squared distance and things based on it */
 961             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 962             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 963             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 964
 965             rinv00           = avx128fma_invsqrt_d(rsq00);
 966             rinv10           = avx128fma_invsqrt_d(rsq10);
 967             rinv20           = avx128fma_invsqrt_d(rsq20);
 968
 969             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 970             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 971             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 972
 973             /* Load parameters for j particles */
 974             jq0              = _mm_load_sd(charge+jnrA+0);
 975             vdwjidx0A        = 2*vdwtype[jnrA+0];
 976
 977             fjx0             = _mm_setzero_pd();
 978             fjy0             = _mm_setzero_pd();
 979             fjz0             = _mm_setzero_pd();
 980
 981             /**************************
 982              * CALCULATE INTERACTIONS *
 983              **************************/
 984
 985             r00              = _mm_mul_pd(rsq00,rinv00);
 986
 987             /* Compute parameters for interactions between i and j atoms */
 988             qq00             = _mm_mul_pd(iq0,jq0);
 989             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 990             c6grid_00       = gmx_mm_load_1real_pd(vdwgridparam+vdwioffset0+vdwjidx0A);
 991
 992             /* EWALD ELECTROSTATICS */
 993
 994             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 995             ewrt             = _mm_mul_pd(r00,ewtabscale);
 996             ewitab           = _mm_cvttpd_epi32(ewrt);
 997 #ifdef __XOP__
 998             eweps            = _mm_frcz_pd(ewrt);
 999 #else
1000             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1001 #endif
1002             twoeweps         = _mm_add_pd(eweps,eweps);
1003             gmx_mm_load_1pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
1004             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
1005             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
1006
1007             /* Analytical LJ-PME */
1008             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1009             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
1010             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
1011             exponent         = avx128fma_exp_d(ewcljrsq);
1012             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
1013             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
1014             /* f6A = 6 * C6grid * (1 - poly) */
1015             f6A              = _mm_mul_pd(c6grid_00,_mm_sub_pd(one,poly));
1016             /* f6B = C6grid * exponent * beta^6 */
1017             f6B              = _mm_mul_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6));
1018             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
1019             fvdw              = _mm_mul_pd(_mm_macc_pd(_mm_msub_pd(c12_00,rinvsix,_mm_sub_pd(c6_00,f6A)),rinvsix,f6B),rinvsq00);
1020
1021             fscal            = _mm_add_pd(felec,fvdw);
1022
1023             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1024
1025             /* Update vectorial force */
1026             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1027             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1028             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1029
1030             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1031             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1032             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             r10              = _mm_mul_pd(rsq10,rinv10);
1039
1040             /* Compute parameters for interactions between i and j atoms */
1041             qq10             = _mm_mul_pd(iq1,jq0);
1042
1043             /* EWALD ELECTROSTATICS */
1044
1045             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1046             ewrt             = _mm_mul_pd(r10,ewtabscale);
1047             ewitab           = _mm_cvttpd_epi32(ewrt);
1048 #ifdef __XOP__
1049             eweps            = _mm_frcz_pd(ewrt);
1050 #else
1051             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1052 #endif
1053             twoeweps         = _mm_add_pd(eweps,eweps);
1054             gmx_mm_load_1pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
1055             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
1056             felec            = _mm_mul_pd(_mm_mul_pd(qq10,rinv10),_mm_sub_pd(rinvsq10,felec));
1057
1058             fscal            = felec;
1059
1060             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1061
1062             /* Update vectorial force */
1063             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1064             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1065             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1066
1067             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1068             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1069             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1070
1071             /**************************
1072              * CALCULATE INTERACTIONS *
1073              **************************/
1074
1075             r20              = _mm_mul_pd(rsq20,rinv20);
1076
1077             /* Compute parameters for interactions between i and j atoms */
1078             qq20             = _mm_mul_pd(iq2,jq0);
1079
1080             /* EWALD ELECTROSTATICS */
1081
1082             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1083             ewrt             = _mm_mul_pd(r20,ewtabscale);
1084             ewitab           = _mm_cvttpd_epi32(ewrt);
1085 #ifdef __XOP__
1086             eweps            = _mm_frcz_pd(ewrt);
1087 #else
1088             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1089 #endif
1090             twoeweps         = _mm_add_pd(eweps,eweps);
1091             gmx_mm_load_1pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
1092             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
1093             felec            = _mm_mul_pd(_mm_mul_pd(qq20,rinv20),_mm_sub_pd(rinvsq20,felec));
1094
1095             fscal            = felec;
1096
1097             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1098
1099             /* Update vectorial force */
1100             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1101             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1102             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1103
1104             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1105             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1106             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1107
1108             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1109
1110             /* Inner loop uses 141 flops */
1111         }
1112
1113         /* End of innermost loop */
1114
1115         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1116                                               f+i_coord_offset,fshift+i_shift_offset);
1117
1118         /* Increment number of inner iterations */
1119         inneriter                  += j_index_end - j_index_start;
1120
1121         /* Outer loop uses 18 flops */
1122     }
1123
1124     /* Increment number of outer iterations */
1125     outeriter        += nri;
1126
1127     /* Update outer/inner flops */
1128
1129     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*141);
1130 }