src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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_ElecCoul_VdwLJ_GeomW3W3_VF_avx_128_fma_double
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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     int              vdwjidx1A,vdwjidx1B;
  88     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 107     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 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
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 132     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 133     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 134     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 135
 136     jq0              = _mm_set1_pd(charge[inr+0]);
 137     jq1              = _mm_set1_pd(charge[inr+1]);
 138     jq2              = _mm_set1_pd(charge[inr+2]);
 139     vdwjidx0A        = 2*vdwtype[inr+0];
 140     qq00             = _mm_mul_pd(iq0,jq0);
 141     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 142     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 143     qq01             = _mm_mul_pd(iq0,jq1);
 144     qq02             = _mm_mul_pd(iq0,jq2);
 145     qq10             = _mm_mul_pd(iq1,jq0);
 146     qq11             = _mm_mul_pd(iq1,jq1);
 147     qq12             = _mm_mul_pd(iq1,jq2);
 148     qq20             = _mm_mul_pd(iq2,jq0);
 149     qq21             = _mm_mul_pd(iq2,jq1);
 150     qq22             = _mm_mul_pd(iq2,jq2);
 151
 152     /* Avoid stupid compiler warnings */
 153     jnrA = jnrB = 0;
 154     j_coord_offsetA = 0;
 155     j_coord_offsetB = 0;
 156
 157     outeriter        = 0;
 158     inneriter        = 0;
 159
 160     /* Start outer loop over neighborlists */
 161     for(iidx=0; iidx<nri; iidx++)
 162     {
 163         /* Load shift vector for this list */
 164         i_shift_offset   = DIM*shiftidx[iidx];
 165
 166         /* Load limits for loop over neighbors */
 167         j_index_start    = jindex[iidx];
 168         j_index_end      = jindex[iidx+1];
 169
 170         /* Get outer coordinate index */
 171         inr              = iinr[iidx];
 172         i_coord_offset   = DIM*inr;
 173
 174         /* Load i particle coords and add shift vector */
 175         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 176                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 177
 178         fix0             = _mm_setzero_pd();
 179         fiy0             = _mm_setzero_pd();
 180         fiz0             = _mm_setzero_pd();
 181         fix1             = _mm_setzero_pd();
 182         fiy1             = _mm_setzero_pd();
 183         fiz1             = _mm_setzero_pd();
 184         fix2             = _mm_setzero_pd();
 185         fiy2             = _mm_setzero_pd();
 186         fiz2             = _mm_setzero_pd();
 187
 188         /* Reset potential sums */
 189         velecsum         = _mm_setzero_pd();
 190         vvdwsum          = _mm_setzero_pd();
 191
 192         /* Start inner kernel loop */
 193         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 194         {
 195
 196             /* Get j neighbor index, and coordinate index */
 197             jnrA             = jjnr[jidx];
 198             jnrB             = jjnr[jidx+1];
 199             j_coord_offsetA  = DIM*jnrA;
 200             j_coord_offsetB  = DIM*jnrB;
 201
 202             /* load j atom coordinates */
 203             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 204                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 205
 206             /* Calculate displacement vector */
 207             dx00             = _mm_sub_pd(ix0,jx0);
 208             dy00             = _mm_sub_pd(iy0,jy0);
 209             dz00             = _mm_sub_pd(iz0,jz0);
 210             dx01             = _mm_sub_pd(ix0,jx1);
 211             dy01             = _mm_sub_pd(iy0,jy1);
 212             dz01             = _mm_sub_pd(iz0,jz1);
 213             dx02             = _mm_sub_pd(ix0,jx2);
 214             dy02             = _mm_sub_pd(iy0,jy2);
 215             dz02             = _mm_sub_pd(iz0,jz2);
 216             dx10             = _mm_sub_pd(ix1,jx0);
 217             dy10             = _mm_sub_pd(iy1,jy0);
 218             dz10             = _mm_sub_pd(iz1,jz0);
 219             dx11             = _mm_sub_pd(ix1,jx1);
 220             dy11             = _mm_sub_pd(iy1,jy1);
 221             dz11             = _mm_sub_pd(iz1,jz1);
 222             dx12             = _mm_sub_pd(ix1,jx2);
 223             dy12             = _mm_sub_pd(iy1,jy2);
 224             dz12             = _mm_sub_pd(iz1,jz2);
 225             dx20             = _mm_sub_pd(ix2,jx0);
 226             dy20             = _mm_sub_pd(iy2,jy0);
 227             dz20             = _mm_sub_pd(iz2,jz0);
 228             dx21             = _mm_sub_pd(ix2,jx1);
 229             dy21             = _mm_sub_pd(iy2,jy1);
 230             dz21             = _mm_sub_pd(iz2,jz1);
 231             dx22             = _mm_sub_pd(ix2,jx2);
 232             dy22             = _mm_sub_pd(iy2,jy2);
 233             dz22             = _mm_sub_pd(iz2,jz2);
 234
 235             /* Calculate squared distance and things based on it */
 236             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 237             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 238             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 239             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 240             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 241             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 242             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 243             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 244             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 245
 246             rinv00           = avx128fma_invsqrt_d(rsq00);
 247             rinv01           = avx128fma_invsqrt_d(rsq01);
 248             rinv02           = avx128fma_invsqrt_d(rsq02);
 249             rinv10           = avx128fma_invsqrt_d(rsq10);
 250             rinv11           = avx128fma_invsqrt_d(rsq11);
 251             rinv12           = avx128fma_invsqrt_d(rsq12);
 252             rinv20           = avx128fma_invsqrt_d(rsq20);
 253             rinv21           = avx128fma_invsqrt_d(rsq21);
 254             rinv22           = avx128fma_invsqrt_d(rsq22);
 255
 256             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 257             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 258             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 259             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 260             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 261             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 262             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 263             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 264             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 265
 266             fjx0             = _mm_setzero_pd();
 267             fjy0             = _mm_setzero_pd();
 268             fjz0             = _mm_setzero_pd();
 269             fjx1             = _mm_setzero_pd();
 270             fjy1             = _mm_setzero_pd();
 271             fjz1             = _mm_setzero_pd();
 272             fjx2             = _mm_setzero_pd();
 273             fjy2             = _mm_setzero_pd();
 274             fjz2             = _mm_setzero_pd();
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             /* COULOMB ELECTROSTATICS */
 281             velec            = _mm_mul_pd(qq00,rinv00);
 282             felec            = _mm_mul_pd(velec,rinvsq00);
 283
 284             /* LENNARD-JONES DISPERSION/REPULSION */
 285
 286             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 287             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 288             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 289             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 290             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             velecsum         = _mm_add_pd(velecsum,velec);
 294             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 295
 296             fscal            = _mm_add_pd(felec,fvdw);
 297
 298             /* Update vectorial force */
 299             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 300             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 301             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 302
 303             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 304             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 305             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             /* COULOMB ELECTROSTATICS */
 312             velec            = _mm_mul_pd(qq01,rinv01);
 313             felec            = _mm_mul_pd(velec,rinvsq01);
 314
 315             /* Update potential sum for this i atom from the interaction with this j atom. */
 316             velecsum         = _mm_add_pd(velecsum,velec);
 317
 318             fscal            = felec;
 319
 320             /* Update vectorial force */
 321             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 322             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 323             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 324
 325             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 326             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 327             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 328
 329             /**************************
 330              * CALCULATE INTERACTIONS *
 331              **************************/
 332
 333             /* COULOMB ELECTROSTATICS */
 334             velec            = _mm_mul_pd(qq02,rinv02);
 335             felec            = _mm_mul_pd(velec,rinvsq02);
 336
 337             /* Update potential sum for this i atom from the interaction with this j atom. */
 338             velecsum         = _mm_add_pd(velecsum,velec);
 339
 340             fscal            = felec;
 341
 342             /* Update vectorial force */
 343             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 344             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 345             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 346
 347             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 348             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 349             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             /* COULOMB ELECTROSTATICS */
 356             velec            = _mm_mul_pd(qq10,rinv10);
 357             felec            = _mm_mul_pd(velec,rinvsq10);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velecsum         = _mm_add_pd(velecsum,velec);
 361
 362             fscal            = felec;
 363
 364             /* Update vectorial force */
 365             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 366             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 367             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 368
 369             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 370             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 371             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             /* COULOMB ELECTROSTATICS */
 378             velec            = _mm_mul_pd(qq11,rinv11);
 379             felec            = _mm_mul_pd(velec,rinvsq11);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velecsum         = _mm_add_pd(velecsum,velec);
 383
 384             fscal            = felec;
 385
 386             /* Update vectorial force */
 387             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 388             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 389             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 390
 391             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 392             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 393             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* COULOMB ELECTROSTATICS */
 400             velec            = _mm_mul_pd(qq12,rinv12);
 401             felec            = _mm_mul_pd(velec,rinvsq12);
 402
 403             /* Update potential sum for this i atom from the interaction with this j atom. */
 404             velecsum         = _mm_add_pd(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             /* Update vectorial force */
 409             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 410             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 411             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 412
 413             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 414             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 415             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             /* COULOMB ELECTROSTATICS */
 422             velec            = _mm_mul_pd(qq20,rinv20);
 423             felec            = _mm_mul_pd(velec,rinvsq20);
 424
 425             /* Update potential sum for this i atom from the interaction with this j atom. */
 426             velecsum         = _mm_add_pd(velecsum,velec);
 427
 428             fscal            = felec;
 429
 430             /* Update vectorial force */
 431             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 432             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 433             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 434
 435             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 436             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 437             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 438
 439             /**************************
 440              * CALCULATE INTERACTIONS *
 441              **************************/
 442
 443             /* COULOMB ELECTROSTATICS */
 444             velec            = _mm_mul_pd(qq21,rinv21);
 445             felec            = _mm_mul_pd(velec,rinvsq21);
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velecsum         = _mm_add_pd(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             /* Update vectorial force */
 453             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 454             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 455             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 456
 457             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 458             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 459             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             /* COULOMB ELECTROSTATICS */
 466             velec            = _mm_mul_pd(qq22,rinv22);
 467             felec            = _mm_mul_pd(velec,rinvsq22);
 468
 469             /* Update potential sum for this i atom from the interaction with this j atom. */
 470             velecsum         = _mm_add_pd(velecsum,velec);
 471
 472             fscal            = felec;
 473
 474             /* Update vectorial force */
 475             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 476             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 477             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 478
 479             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 480             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 481             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 482
 483             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 484
 485             /* Inner loop uses 291 flops */
 486         }
 487
 488         if(jidx<j_index_end)
 489         {
 490
 491             jnrA             = jjnr[jidx];
 492             j_coord_offsetA  = DIM*jnrA;
 493
 494             /* load j atom coordinates */
 495             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 496                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 497
 498             /* Calculate displacement vector */
 499             dx00             = _mm_sub_pd(ix0,jx0);
 500             dy00             = _mm_sub_pd(iy0,jy0);
 501             dz00             = _mm_sub_pd(iz0,jz0);
 502             dx01             = _mm_sub_pd(ix0,jx1);
 503             dy01             = _mm_sub_pd(iy0,jy1);
 504             dz01             = _mm_sub_pd(iz0,jz1);
 505             dx02             = _mm_sub_pd(ix0,jx2);
 506             dy02             = _mm_sub_pd(iy0,jy2);
 507             dz02             = _mm_sub_pd(iz0,jz2);
 508             dx10             = _mm_sub_pd(ix1,jx0);
 509             dy10             = _mm_sub_pd(iy1,jy0);
 510             dz10             = _mm_sub_pd(iz1,jz0);
 511             dx11             = _mm_sub_pd(ix1,jx1);
 512             dy11             = _mm_sub_pd(iy1,jy1);
 513             dz11             = _mm_sub_pd(iz1,jz1);
 514             dx12             = _mm_sub_pd(ix1,jx2);
 515             dy12             = _mm_sub_pd(iy1,jy2);
 516             dz12             = _mm_sub_pd(iz1,jz2);
 517             dx20             = _mm_sub_pd(ix2,jx0);
 518             dy20             = _mm_sub_pd(iy2,jy0);
 519             dz20             = _mm_sub_pd(iz2,jz0);
 520             dx21             = _mm_sub_pd(ix2,jx1);
 521             dy21             = _mm_sub_pd(iy2,jy1);
 522             dz21             = _mm_sub_pd(iz2,jz1);
 523             dx22             = _mm_sub_pd(ix2,jx2);
 524             dy22             = _mm_sub_pd(iy2,jy2);
 525             dz22             = _mm_sub_pd(iz2,jz2);
 526
 527             /* Calculate squared distance and things based on it */
 528             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 529             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 530             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 531             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 532             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 533             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 534             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 535             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 536             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 537
 538             rinv00           = avx128fma_invsqrt_d(rsq00);
 539             rinv01           = avx128fma_invsqrt_d(rsq01);
 540             rinv02           = avx128fma_invsqrt_d(rsq02);
 541             rinv10           = avx128fma_invsqrt_d(rsq10);
 542             rinv11           = avx128fma_invsqrt_d(rsq11);
 543             rinv12           = avx128fma_invsqrt_d(rsq12);
 544             rinv20           = avx128fma_invsqrt_d(rsq20);
 545             rinv21           = avx128fma_invsqrt_d(rsq21);
 546             rinv22           = avx128fma_invsqrt_d(rsq22);
 547
 548             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 549             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 550             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 551             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 552             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 553             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 554             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 555             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 556             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 557
 558             fjx0             = _mm_setzero_pd();
 559             fjy0             = _mm_setzero_pd();
 560             fjz0             = _mm_setzero_pd();
 561             fjx1             = _mm_setzero_pd();
 562             fjy1             = _mm_setzero_pd();
 563             fjz1             = _mm_setzero_pd();
 564             fjx2             = _mm_setzero_pd();
 565             fjy2             = _mm_setzero_pd();
 566             fjz2             = _mm_setzero_pd();
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             /* COULOMB ELECTROSTATICS */
 573             velec            = _mm_mul_pd(qq00,rinv00);
 574             felec            = _mm_mul_pd(velec,rinvsq00);
 575
 576             /* LENNARD-JONES DISPERSION/REPULSION */
 577
 578             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 579             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 580             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 581             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 582             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 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             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 588             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 589
 590             fscal            = _mm_add_pd(felec,fvdw);
 591
 592             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 593
 594             /* Update vectorial force */
 595             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 596             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 597             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 598
 599             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 600             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 601             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 602
 603             /**************************
 604              * CALCULATE INTERACTIONS *
 605              **************************/
 606
 607             /* COULOMB ELECTROSTATICS */
 608             velec            = _mm_mul_pd(qq01,rinv01);
 609             felec            = _mm_mul_pd(velec,rinvsq01);
 610
 611             /* Update potential sum for this i atom from the interaction with this j atom. */
 612             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 613             velecsum         = _mm_add_pd(velecsum,velec);
 614
 615             fscal            = felec;
 616
 617             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 618
 619             /* Update vectorial force */
 620             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 621             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 622             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 623
 624             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 625             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 626             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 627
 628             /**************************
 629              * CALCULATE INTERACTIONS *
 630              **************************/
 631
 632             /* COULOMB ELECTROSTATICS */
 633             velec            = _mm_mul_pd(qq02,rinv02);
 634             felec            = _mm_mul_pd(velec,rinvsq02);
 635
 636             /* Update potential sum for this i atom from the interaction with this j atom. */
 637             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 638             velecsum         = _mm_add_pd(velecsum,velec);
 639
 640             fscal            = felec;
 641
 642             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 643
 644             /* Update vectorial force */
 645             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 646             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 647             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 648
 649             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 650             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 651             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 652
 653             /**************************
 654              * CALCULATE INTERACTIONS *
 655              **************************/
 656
 657             /* COULOMB ELECTROSTATICS */
 658             velec            = _mm_mul_pd(qq10,rinv10);
 659             felec            = _mm_mul_pd(velec,rinvsq10);
 660
 661             /* Update potential sum for this i atom from the interaction with this j atom. */
 662             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 663             velecsum         = _mm_add_pd(velecsum,velec);
 664
 665             fscal            = felec;
 666
 667             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 668
 669             /* Update vectorial force */
 670             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 671             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 672             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 673
 674             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 675             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 676             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             /* COULOMB ELECTROSTATICS */
 683             velec            = _mm_mul_pd(qq11,rinv11);
 684             felec            = _mm_mul_pd(velec,rinvsq11);
 685
 686             /* Update potential sum for this i atom from the interaction with this j atom. */
 687             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 688             velecsum         = _mm_add_pd(velecsum,velec);
 689
 690             fscal            = felec;
 691
 692             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 693
 694             /* Update vectorial force */
 695             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 696             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 697             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 698
 699             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 700             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 701             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             /* COULOMB ELECTROSTATICS */
 708             velec            = _mm_mul_pd(qq12,rinv12);
 709             felec            = _mm_mul_pd(velec,rinvsq12);
 710
 711             /* Update potential sum for this i atom from the interaction with this j atom. */
 712             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 713             velecsum         = _mm_add_pd(velecsum,velec);
 714
 715             fscal            = felec;
 716
 717             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 718
 719             /* Update vectorial force */
 720             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 721             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 722             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 723
 724             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 725             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 726             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 727
 728             /**************************
 729              * CALCULATE INTERACTIONS *
 730              **************************/
 731
 732             /* COULOMB ELECTROSTATICS */
 733             velec            = _mm_mul_pd(qq20,rinv20);
 734             felec            = _mm_mul_pd(velec,rinvsq20);
 735
 736             /* Update potential sum for this i atom from the interaction with this j atom. */
 737             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 738             velecsum         = _mm_add_pd(velecsum,velec);
 739
 740             fscal            = felec;
 741
 742             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 743
 744             /* Update vectorial force */
 745             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 746             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 747             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 748
 749             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 750             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 751             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 752
 753             /**************************
 754              * CALCULATE INTERACTIONS *
 755              **************************/
 756
 757             /* COULOMB ELECTROSTATICS */
 758             velec            = _mm_mul_pd(qq21,rinv21);
 759             felec            = _mm_mul_pd(velec,rinvsq21);
 760
 761             /* Update potential sum for this i atom from the interaction with this j atom. */
 762             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 763             velecsum         = _mm_add_pd(velecsum,velec);
 764
 765             fscal            = felec;
 766
 767             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 768
 769             /* Update vectorial force */
 770             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 771             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 772             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 773
 774             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 775             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 776             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 777
 778             /**************************
 779              * CALCULATE INTERACTIONS *
 780              **************************/
 781
 782             /* COULOMB ELECTROSTATICS */
 783             velec            = _mm_mul_pd(qq22,rinv22);
 784             felec            = _mm_mul_pd(velec,rinvsq22);
 785
 786             /* Update potential sum for this i atom from the interaction with this j atom. */
 787             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 788             velecsum         = _mm_add_pd(velecsum,velec);
 789
 790             fscal            = felec;
 791
 792             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 793
 794             /* Update vectorial force */
 795             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 796             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 797             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 798
 799             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 800             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 801             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 802
 803             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 804
 805             /* Inner loop uses 291 flops */
 806         }
 807
 808         /* End of innermost loop */
 809
 810         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 811                                               f+i_coord_offset,fshift+i_shift_offset);
 812
 813         ggid                        = gid[iidx];
 814         /* Update potential energies */
 815         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 816         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 817
 818         /* Increment number of inner iterations */
 819         inneriter                  += j_index_end - j_index_start;
 820
 821         /* Outer loop uses 20 flops */
 822     }
 823
 824     /* Increment number of outer iterations */
 825     outeriter        += nri;
 826
 827     /* Update outer/inner flops */
 828
 829     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*291);
 830 }
 831 /*
 832  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_avx_128_fma_double
 833  * Electrostatics interaction: Coulomb
 834  * VdW interaction:            LennardJones
 835  * Geometry:                   Water3-Water3
 836  * Calculate force/pot:        Force
 837  */
 838 void
 839 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_avx_128_fma_double
 840                     (t_nblist                    * gmx_restrict       nlist,
 841                      rvec                        * gmx_restrict          xx,
 842                      rvec                        * gmx_restrict          ff,
 843                      struct t_forcerec           * gmx_restrict          fr,
 844                      t_mdatoms                   * gmx_restrict     mdatoms,
 845                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 846                      t_nrnb                      * gmx_restrict        nrnb)
 847 {
 848     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 849      * just 0 for non-waters.
 850      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 851      * jnr indices corresponding to data put in the four positions in the SIMD register.
 852      */
 853     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 854     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 855     int              jnrA,jnrB;
 856     int              j_coord_offsetA,j_coord_offsetB;
 857     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 858     real             rcutoff_scalar;
 859     real             *shiftvec,*fshift,*x,*f;
 860     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 861     int              vdwioffset0;
 862     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 863     int              vdwioffset1;
 864     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 865     int              vdwioffset2;
 866     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 867     int              vdwjidx0A,vdwjidx0B;
 868     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 869     int              vdwjidx1A,vdwjidx1B;
 870     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 871     int              vdwjidx2A,vdwjidx2B;
 872     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 873     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 874     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 875     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 876     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 877     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 878     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 879     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 880     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 881     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 882     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 883     real             *charge;
 884     int              nvdwtype;
 885     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 886     int              *vdwtype;
 887     real             *vdwparam;
 888     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 889     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 890     __m128d          dummy_mask,cutoff_mask;
 891     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 892     __m128d          one     = _mm_set1_pd(1.0);
 893     __m128d          two     = _mm_set1_pd(2.0);
 894     x                = xx[0];
 895     f                = ff[0];
 896
 897     nri              = nlist->nri;
 898     iinr             = nlist->iinr;
 899     jindex           = nlist->jindex;
 900     jjnr             = nlist->jjnr;
 901     shiftidx         = nlist->shift;
 902     gid              = nlist->gid;
 903     shiftvec         = fr->shift_vec[0];
 904     fshift           = fr->fshift[0];
 905     facel            = _mm_set1_pd(fr->ic->epsfac);
 906     charge           = mdatoms->chargeA;
 907     nvdwtype         = fr->ntype;
 908     vdwparam         = fr->nbfp;
 909     vdwtype          = mdatoms->typeA;
 910
 911     /* Setup water-specific parameters */
 912     inr              = nlist->iinr[0];
 913     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 914     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 915     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 916     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 917
 918     jq0              = _mm_set1_pd(charge[inr+0]);
 919     jq1              = _mm_set1_pd(charge[inr+1]);
 920     jq2              = _mm_set1_pd(charge[inr+2]);
 921     vdwjidx0A        = 2*vdwtype[inr+0];
 922     qq00             = _mm_mul_pd(iq0,jq0);
 923     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 924     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 925     qq01             = _mm_mul_pd(iq0,jq1);
 926     qq02             = _mm_mul_pd(iq0,jq2);
 927     qq10             = _mm_mul_pd(iq1,jq0);
 928     qq11             = _mm_mul_pd(iq1,jq1);
 929     qq12             = _mm_mul_pd(iq1,jq2);
 930     qq20             = _mm_mul_pd(iq2,jq0);
 931     qq21             = _mm_mul_pd(iq2,jq1);
 932     qq22             = _mm_mul_pd(iq2,jq2);
 933
 934     /* Avoid stupid compiler warnings */
 935     jnrA = jnrB = 0;
 936     j_coord_offsetA = 0;
 937     j_coord_offsetB = 0;
 938
 939     outeriter        = 0;
 940     inneriter        = 0;
 941
 942     /* Start outer loop over neighborlists */
 943     for(iidx=0; iidx<nri; iidx++)
 944     {
 945         /* Load shift vector for this list */
 946         i_shift_offset   = DIM*shiftidx[iidx];
 947
 948         /* Load limits for loop over neighbors */
 949         j_index_start    = jindex[iidx];
 950         j_index_end      = jindex[iidx+1];
 951
 952         /* Get outer coordinate index */
 953         inr              = iinr[iidx];
 954         i_coord_offset   = DIM*inr;
 955
 956         /* Load i particle coords and add shift vector */
 957         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 958                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 959
 960         fix0             = _mm_setzero_pd();
 961         fiy0             = _mm_setzero_pd();
 962         fiz0             = _mm_setzero_pd();
 963         fix1             = _mm_setzero_pd();
 964         fiy1             = _mm_setzero_pd();
 965         fiz1             = _mm_setzero_pd();
 966         fix2             = _mm_setzero_pd();
 967         fiy2             = _mm_setzero_pd();
 968         fiz2             = _mm_setzero_pd();
 969
 970         /* Start inner kernel loop */
 971         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 972         {
 973
 974             /* Get j neighbor index, and coordinate index */
 975             jnrA             = jjnr[jidx];
 976             jnrB             = jjnr[jidx+1];
 977             j_coord_offsetA  = DIM*jnrA;
 978             j_coord_offsetB  = DIM*jnrB;
 979
 980             /* load j atom coordinates */
 981             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 982                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 983
 984             /* Calculate displacement vector */
 985             dx00             = _mm_sub_pd(ix0,jx0);
 986             dy00             = _mm_sub_pd(iy0,jy0);
 987             dz00             = _mm_sub_pd(iz0,jz0);
 988             dx01             = _mm_sub_pd(ix0,jx1);
 989             dy01             = _mm_sub_pd(iy0,jy1);
 990             dz01             = _mm_sub_pd(iz0,jz1);
 991             dx02             = _mm_sub_pd(ix0,jx2);
 992             dy02             = _mm_sub_pd(iy0,jy2);
 993             dz02             = _mm_sub_pd(iz0,jz2);
 994             dx10             = _mm_sub_pd(ix1,jx0);
 995             dy10             = _mm_sub_pd(iy1,jy0);
 996             dz10             = _mm_sub_pd(iz1,jz0);
 997             dx11             = _mm_sub_pd(ix1,jx1);
 998             dy11             = _mm_sub_pd(iy1,jy1);
 999             dz11             = _mm_sub_pd(iz1,jz1);
1000             dx12             = _mm_sub_pd(ix1,jx2);
1001             dy12             = _mm_sub_pd(iy1,jy2);
1002             dz12             = _mm_sub_pd(iz1,jz2);
1003             dx20             = _mm_sub_pd(ix2,jx0);
1004             dy20             = _mm_sub_pd(iy2,jy0);
1005             dz20             = _mm_sub_pd(iz2,jz0);
1006             dx21             = _mm_sub_pd(ix2,jx1);
1007             dy21             = _mm_sub_pd(iy2,jy1);
1008             dz21             = _mm_sub_pd(iz2,jz1);
1009             dx22             = _mm_sub_pd(ix2,jx2);
1010             dy22             = _mm_sub_pd(iy2,jy2);
1011             dz22             = _mm_sub_pd(iz2,jz2);
1012
1013             /* Calculate squared distance and things based on it */
1014             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1015             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1016             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1017             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1018             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1019             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1020             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1021             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1022             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1023
1024             rinv00           = avx128fma_invsqrt_d(rsq00);
1025             rinv01           = avx128fma_invsqrt_d(rsq01);
1026             rinv02           = avx128fma_invsqrt_d(rsq02);
1027             rinv10           = avx128fma_invsqrt_d(rsq10);
1028             rinv11           = avx128fma_invsqrt_d(rsq11);
1029             rinv12           = avx128fma_invsqrt_d(rsq12);
1030             rinv20           = avx128fma_invsqrt_d(rsq20);
1031             rinv21           = avx128fma_invsqrt_d(rsq21);
1032             rinv22           = avx128fma_invsqrt_d(rsq22);
1033
1034             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1035             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1036             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1037             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1038             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1039             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1040             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1041             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1042             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1043
1044             fjx0             = _mm_setzero_pd();
1045             fjy0             = _mm_setzero_pd();
1046             fjz0             = _mm_setzero_pd();
1047             fjx1             = _mm_setzero_pd();
1048             fjy1             = _mm_setzero_pd();
1049             fjz1             = _mm_setzero_pd();
1050             fjx2             = _mm_setzero_pd();
1051             fjy2             = _mm_setzero_pd();
1052             fjz2             = _mm_setzero_pd();
1053
1054             /**************************
1055              * CALCULATE INTERACTIONS *
1056              **************************/
1057
1058             /* COULOMB ELECTROSTATICS */
1059             velec            = _mm_mul_pd(qq00,rinv00);
1060             felec            = _mm_mul_pd(velec,rinvsq00);
1061
1062             /* LENNARD-JONES DISPERSION/REPULSION */
1063
1064             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1065             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1066
1067             fscal            = _mm_add_pd(felec,fvdw);
1068
1069             /* Update vectorial force */
1070             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1071             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1072             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1073
1074             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1075             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1076             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1077
1078             /**************************
1079              * CALCULATE INTERACTIONS *
1080              **************************/
1081
1082             /* COULOMB ELECTROSTATICS */
1083             velec            = _mm_mul_pd(qq01,rinv01);
1084             felec            = _mm_mul_pd(velec,rinvsq01);
1085
1086             fscal            = felec;
1087
1088             /* Update vectorial force */
1089             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1090             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1091             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1092
1093             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1094             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1095             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1096
1097             /**************************
1098              * CALCULATE INTERACTIONS *
1099              **************************/
1100
1101             /* COULOMB ELECTROSTATICS */
1102             velec            = _mm_mul_pd(qq02,rinv02);
1103             felec            = _mm_mul_pd(velec,rinvsq02);
1104
1105             fscal            = felec;
1106
1107             /* Update vectorial force */
1108             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1109             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1110             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1111
1112             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1113             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1114             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1115
1116             /**************************
1117              * CALCULATE INTERACTIONS *
1118              **************************/
1119
1120             /* COULOMB ELECTROSTATICS */
1121             velec            = _mm_mul_pd(qq10,rinv10);
1122             felec            = _mm_mul_pd(velec,rinvsq10);
1123
1124             fscal            = felec;
1125
1126             /* Update vectorial force */
1127             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1128             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1129             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1130
1131             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1132             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1133             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1134
1135             /**************************
1136              * CALCULATE INTERACTIONS *
1137              **************************/
1138
1139             /* COULOMB ELECTROSTATICS */
1140             velec            = _mm_mul_pd(qq11,rinv11);
1141             felec            = _mm_mul_pd(velec,rinvsq11);
1142
1143             fscal            = felec;
1144
1145             /* Update vectorial force */
1146             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1147             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1148             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1149
1150             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1151             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1152             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1153
1154             /**************************
1155              * CALCULATE INTERACTIONS *
1156              **************************/
1157
1158             /* COULOMB ELECTROSTATICS */
1159             velec            = _mm_mul_pd(qq12,rinv12);
1160             felec            = _mm_mul_pd(velec,rinvsq12);
1161
1162             fscal            = felec;
1163
1164             /* Update vectorial force */
1165             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1166             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1167             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1168
1169             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1170             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1171             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1172
1173             /**************************
1174              * CALCULATE INTERACTIONS *
1175              **************************/
1176
1177             /* COULOMB ELECTROSTATICS */
1178             velec            = _mm_mul_pd(qq20,rinv20);
1179             felec            = _mm_mul_pd(velec,rinvsq20);
1180
1181             fscal            = felec;
1182
1183             /* Update vectorial force */
1184             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1185             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1186             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1187
1188             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1189             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1190             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1191
1192             /**************************
1193              * CALCULATE INTERACTIONS *
1194              **************************/
1195
1196             /* COULOMB ELECTROSTATICS */
1197             velec            = _mm_mul_pd(qq21,rinv21);
1198             felec            = _mm_mul_pd(velec,rinvsq21);
1199
1200             fscal            = felec;
1201
1202             /* Update vectorial force */
1203             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1204             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1205             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1206
1207             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1208             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1209             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1210
1211             /**************************
1212              * CALCULATE INTERACTIONS *
1213              **************************/
1214
1215             /* COULOMB ELECTROSTATICS */
1216             velec            = _mm_mul_pd(qq22,rinv22);
1217             felec            = _mm_mul_pd(velec,rinvsq22);
1218
1219             fscal            = felec;
1220
1221             /* Update vectorial force */
1222             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1223             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1224             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1225
1226             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1227             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1228             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1229
1230             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1231
1232             /* Inner loop uses 277 flops */
1233         }
1234
1235         if(jidx<j_index_end)
1236         {
1237
1238             jnrA             = jjnr[jidx];
1239             j_coord_offsetA  = DIM*jnrA;
1240
1241             /* load j atom coordinates */
1242             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1243                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1244
1245             /* Calculate displacement vector */
1246             dx00             = _mm_sub_pd(ix0,jx0);
1247             dy00             = _mm_sub_pd(iy0,jy0);
1248             dz00             = _mm_sub_pd(iz0,jz0);
1249             dx01             = _mm_sub_pd(ix0,jx1);
1250             dy01             = _mm_sub_pd(iy0,jy1);
1251             dz01             = _mm_sub_pd(iz0,jz1);
1252             dx02             = _mm_sub_pd(ix0,jx2);
1253             dy02             = _mm_sub_pd(iy0,jy2);
1254             dz02             = _mm_sub_pd(iz0,jz2);
1255             dx10             = _mm_sub_pd(ix1,jx0);
1256             dy10             = _mm_sub_pd(iy1,jy0);
1257             dz10             = _mm_sub_pd(iz1,jz0);
1258             dx11             = _mm_sub_pd(ix1,jx1);
1259             dy11             = _mm_sub_pd(iy1,jy1);
1260             dz11             = _mm_sub_pd(iz1,jz1);
1261             dx12             = _mm_sub_pd(ix1,jx2);
1262             dy12             = _mm_sub_pd(iy1,jy2);
1263             dz12             = _mm_sub_pd(iz1,jz2);
1264             dx20             = _mm_sub_pd(ix2,jx0);
1265             dy20             = _mm_sub_pd(iy2,jy0);
1266             dz20             = _mm_sub_pd(iz2,jz0);
1267             dx21             = _mm_sub_pd(ix2,jx1);
1268             dy21             = _mm_sub_pd(iy2,jy1);
1269             dz21             = _mm_sub_pd(iz2,jz1);
1270             dx22             = _mm_sub_pd(ix2,jx2);
1271             dy22             = _mm_sub_pd(iy2,jy2);
1272             dz22             = _mm_sub_pd(iz2,jz2);
1273
1274             /* Calculate squared distance and things based on it */
1275             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1276             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1277             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1278             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1279             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1280             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1281             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1282             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1283             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1284
1285             rinv00           = avx128fma_invsqrt_d(rsq00);
1286             rinv01           = avx128fma_invsqrt_d(rsq01);
1287             rinv02           = avx128fma_invsqrt_d(rsq02);
1288             rinv10           = avx128fma_invsqrt_d(rsq10);
1289             rinv11           = avx128fma_invsqrt_d(rsq11);
1290             rinv12           = avx128fma_invsqrt_d(rsq12);
1291             rinv20           = avx128fma_invsqrt_d(rsq20);
1292             rinv21           = avx128fma_invsqrt_d(rsq21);
1293             rinv22           = avx128fma_invsqrt_d(rsq22);
1294
1295             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1296             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1297             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1298             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1299             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1300             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1301             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1302             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1303             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1304
1305             fjx0             = _mm_setzero_pd();
1306             fjy0             = _mm_setzero_pd();
1307             fjz0             = _mm_setzero_pd();
1308             fjx1             = _mm_setzero_pd();
1309             fjy1             = _mm_setzero_pd();
1310             fjz1             = _mm_setzero_pd();
1311             fjx2             = _mm_setzero_pd();
1312             fjy2             = _mm_setzero_pd();
1313             fjz2             = _mm_setzero_pd();
1314
1315             /**************************
1316              * CALCULATE INTERACTIONS *
1317              **************************/
1318
1319             /* COULOMB ELECTROSTATICS */
1320             velec            = _mm_mul_pd(qq00,rinv00);
1321             felec            = _mm_mul_pd(velec,rinvsq00);
1322
1323             /* LENNARD-JONES DISPERSION/REPULSION */
1324
1325             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1326             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1327
1328             fscal            = _mm_add_pd(felec,fvdw);
1329
1330             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1331
1332             /* Update vectorial force */
1333             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1334             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1335             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1336
1337             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1338             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1339             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1340
1341             /**************************
1342              * CALCULATE INTERACTIONS *
1343              **************************/
1344
1345             /* COULOMB ELECTROSTATICS */
1346             velec            = _mm_mul_pd(qq01,rinv01);
1347             felec            = _mm_mul_pd(velec,rinvsq01);
1348
1349             fscal            = felec;
1350
1351             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1352
1353             /* Update vectorial force */
1354             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1355             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1356             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1357
1358             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1359             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1360             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1361
1362             /**************************
1363              * CALCULATE INTERACTIONS *
1364              **************************/
1365
1366             /* COULOMB ELECTROSTATICS */
1367             velec            = _mm_mul_pd(qq02,rinv02);
1368             felec            = _mm_mul_pd(velec,rinvsq02);
1369
1370             fscal            = felec;
1371
1372             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1373
1374             /* Update vectorial force */
1375             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1376             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1377             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1378
1379             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1380             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1381             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1382
1383             /**************************
1384              * CALCULATE INTERACTIONS *
1385              **************************/
1386
1387             /* COULOMB ELECTROSTATICS */
1388             velec            = _mm_mul_pd(qq10,rinv10);
1389             felec            = _mm_mul_pd(velec,rinvsq10);
1390
1391             fscal            = felec;
1392
1393             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1394
1395             /* Update vectorial force */
1396             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1397             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1398             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1399
1400             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1401             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1402             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1403
1404             /**************************
1405              * CALCULATE INTERACTIONS *
1406              **************************/
1407
1408             /* COULOMB ELECTROSTATICS */
1409             velec            = _mm_mul_pd(qq11,rinv11);
1410             felec            = _mm_mul_pd(velec,rinvsq11);
1411
1412             fscal            = felec;
1413
1414             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1415
1416             /* Update vectorial force */
1417             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1418             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1419             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1420
1421             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1422             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1423             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1424
1425             /**************************
1426              * CALCULATE INTERACTIONS *
1427              **************************/
1428
1429             /* COULOMB ELECTROSTATICS */
1430             velec            = _mm_mul_pd(qq12,rinv12);
1431             felec            = _mm_mul_pd(velec,rinvsq12);
1432
1433             fscal            = felec;
1434
1435             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1436
1437             /* Update vectorial force */
1438             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1439             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1440             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1441
1442             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1443             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1444             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1445
1446             /**************************
1447              * CALCULATE INTERACTIONS *
1448              **************************/
1449
1450             /* COULOMB ELECTROSTATICS */
1451             velec            = _mm_mul_pd(qq20,rinv20);
1452             felec            = _mm_mul_pd(velec,rinvsq20);
1453
1454             fscal            = felec;
1455
1456             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1457
1458             /* Update vectorial force */
1459             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1460             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1461             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1462
1463             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1464             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1465             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1466
1467             /**************************
1468              * CALCULATE INTERACTIONS *
1469              **************************/
1470
1471             /* COULOMB ELECTROSTATICS */
1472             velec            = _mm_mul_pd(qq21,rinv21);
1473             felec            = _mm_mul_pd(velec,rinvsq21);
1474
1475             fscal            = felec;
1476
1477             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1478
1479             /* Update vectorial force */
1480             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1481             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1482             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1483
1484             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1485             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1486             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1487
1488             /**************************
1489              * CALCULATE INTERACTIONS *
1490              **************************/
1491
1492             /* COULOMB ELECTROSTATICS */
1493             velec            = _mm_mul_pd(qq22,rinv22);
1494             felec            = _mm_mul_pd(velec,rinvsq22);
1495
1496             fscal            = felec;
1497
1498             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1499
1500             /* Update vectorial force */
1501             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1502             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1503             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1504
1505             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1506             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1507             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1508
1509             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1510
1511             /* Inner loop uses 277 flops */
1512         }
1513
1514         /* End of innermost loop */
1515
1516         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1517                                               f+i_coord_offset,fshift+i_shift_offset);
1518
1519         /* Increment number of inner iterations */
1520         inneriter                  += j_index_end - j_index_start;
1521
1522         /* Outer loop uses 18 flops */
1523     }
1524
1525     /* Increment number of outer iterations */
1526     outeriter        += nri;
1527
1528     /* Update outer/inner flops */
1529
1530     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*277);
1531 }