src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwNone_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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_ElecRFCut_VdwNone_GeomW3W3_VF_avx_128_fma_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            None
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwNone_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     __m128d          dummy_mask,cutoff_mask;
 103     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 104     __m128d          one     = _mm_set1_pd(1.0);
 105     __m128d          two     = _mm_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_pd(fr->ic->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = _mm_set1_pd(fr->ic->k_rf);
 120     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 121     crf              = _mm_set1_pd(fr->ic->c_rf);
 122
 123     /* Setup water-specific parameters */
 124     inr              = nlist->iinr[0];
 125     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128
 129     jq0              = _mm_set1_pd(charge[inr+0]);
 130     jq1              = _mm_set1_pd(charge[inr+1]);
 131     jq2              = _mm_set1_pd(charge[inr+2]);
 132     qq00             = _mm_mul_pd(iq0,jq0);
 133     qq01             = _mm_mul_pd(iq0,jq1);
 134     qq02             = _mm_mul_pd(iq0,jq2);
 135     qq10             = _mm_mul_pd(iq1,jq0);
 136     qq11             = _mm_mul_pd(iq1,jq1);
 137     qq12             = _mm_mul_pd(iq1,jq2);
 138     qq20             = _mm_mul_pd(iq2,jq0);
 139     qq21             = _mm_mul_pd(iq2,jq1);
 140     qq22             = _mm_mul_pd(iq2,jq2);
 141
 142     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 143     rcutoff_scalar   = fr->ic->rcoulomb;
 144     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 145     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 146
 147     /* Avoid stupid compiler warnings */
 148     jnrA = jnrB = 0;
 149     j_coord_offsetA = 0;
 150     j_coord_offsetB = 0;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     /* Start outer loop over neighborlists */
 156     for(iidx=0; iidx<nri; iidx++)
 157     {
 158         /* Load shift vector for this list */
 159         i_shift_offset   = DIM*shiftidx[iidx];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 171                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 172
 173         fix0             = _mm_setzero_pd();
 174         fiy0             = _mm_setzero_pd();
 175         fiz0             = _mm_setzero_pd();
 176         fix1             = _mm_setzero_pd();
 177         fiy1             = _mm_setzero_pd();
 178         fiz1             = _mm_setzero_pd();
 179         fix2             = _mm_setzero_pd();
 180         fiy2             = _mm_setzero_pd();
 181         fiz2             = _mm_setzero_pd();
 182
 183         /* Reset potential sums */
 184         velecsum         = _mm_setzero_pd();
 185
 186         /* Start inner kernel loop */
 187         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 188         {
 189
 190             /* Get j neighbor index, and coordinate index */
 191             jnrA             = jjnr[jidx];
 192             jnrB             = jjnr[jidx+1];
 193             j_coord_offsetA  = DIM*jnrA;
 194             j_coord_offsetB  = DIM*jnrB;
 195
 196             /* load j atom coordinates */
 197             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 198                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm_sub_pd(ix0,jx0);
 202             dy00             = _mm_sub_pd(iy0,jy0);
 203             dz00             = _mm_sub_pd(iz0,jz0);
 204             dx01             = _mm_sub_pd(ix0,jx1);
 205             dy01             = _mm_sub_pd(iy0,jy1);
 206             dz01             = _mm_sub_pd(iz0,jz1);
 207             dx02             = _mm_sub_pd(ix0,jx2);
 208             dy02             = _mm_sub_pd(iy0,jy2);
 209             dz02             = _mm_sub_pd(iz0,jz2);
 210             dx10             = _mm_sub_pd(ix1,jx0);
 211             dy10             = _mm_sub_pd(iy1,jy0);
 212             dz10             = _mm_sub_pd(iz1,jz0);
 213             dx11             = _mm_sub_pd(ix1,jx1);
 214             dy11             = _mm_sub_pd(iy1,jy1);
 215             dz11             = _mm_sub_pd(iz1,jz1);
 216             dx12             = _mm_sub_pd(ix1,jx2);
 217             dy12             = _mm_sub_pd(iy1,jy2);
 218             dz12             = _mm_sub_pd(iz1,jz2);
 219             dx20             = _mm_sub_pd(ix2,jx0);
 220             dy20             = _mm_sub_pd(iy2,jy0);
 221             dz20             = _mm_sub_pd(iz2,jz0);
 222             dx21             = _mm_sub_pd(ix2,jx1);
 223             dy21             = _mm_sub_pd(iy2,jy1);
 224             dz21             = _mm_sub_pd(iz2,jz1);
 225             dx22             = _mm_sub_pd(ix2,jx2);
 226             dy22             = _mm_sub_pd(iy2,jy2);
 227             dz22             = _mm_sub_pd(iz2,jz2);
 228
 229             /* Calculate squared distance and things based on it */
 230             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 231             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 232             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 233             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 234             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 235             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 236             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 237             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 238             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 239
 240             rinv00           = avx128fma_invsqrt_d(rsq00);
 241             rinv01           = avx128fma_invsqrt_d(rsq01);
 242             rinv02           = avx128fma_invsqrt_d(rsq02);
 243             rinv10           = avx128fma_invsqrt_d(rsq10);
 244             rinv11           = avx128fma_invsqrt_d(rsq11);
 245             rinv12           = avx128fma_invsqrt_d(rsq12);
 246             rinv20           = avx128fma_invsqrt_d(rsq20);
 247             rinv21           = avx128fma_invsqrt_d(rsq21);
 248             rinv22           = avx128fma_invsqrt_d(rsq22);
 249
 250             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 251             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 252             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 253             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 254             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 255             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 256             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 257             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 258             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 259
 260             fjx0             = _mm_setzero_pd();
 261             fjy0             = _mm_setzero_pd();
 262             fjz0             = _mm_setzero_pd();
 263             fjx1             = _mm_setzero_pd();
 264             fjy1             = _mm_setzero_pd();
 265             fjz1             = _mm_setzero_pd();
 266             fjx2             = _mm_setzero_pd();
 267             fjy2             = _mm_setzero_pd();
 268             fjz2             = _mm_setzero_pd();
 269
 270             /**************************
 271              * CALCULATE INTERACTIONS *
 272              **************************/
 273
 274             if (gmx_mm_any_lt(rsq00,rcutoff2))
 275             {
 276
 277             /* REACTION-FIELD ELECTROSTATICS */
 278             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 279             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 280
 281             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 282
 283             /* Update potential sum for this i atom from the interaction with this j atom. */
 284             velec            = _mm_and_pd(velec,cutoff_mask);
 285             velecsum         = _mm_add_pd(velecsum,velec);
 286
 287             fscal            = felec;
 288
 289             fscal            = _mm_and_pd(fscal,cutoff_mask);
 290
 291             /* Update vectorial force */
 292             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 293             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 294             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 295
 296             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 297             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 298             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 299
 300             }
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             if (gmx_mm_any_lt(rsq01,rcutoff2))
 307             {
 308
 309             /* REACTION-FIELD ELECTROSTATICS */
 310             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 311             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 312
 313             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 314
 315             /* Update potential sum for this i atom from the interaction with this j atom. */
 316             velec            = _mm_and_pd(velec,cutoff_mask);
 317             velecsum         = _mm_add_pd(velecsum,velec);
 318
 319             fscal            = felec;
 320
 321             fscal            = _mm_and_pd(fscal,cutoff_mask);
 322
 323             /* Update vectorial force */
 324             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 325             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 326             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 327
 328             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 329             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 330             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 331
 332             }
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             if (gmx_mm_any_lt(rsq02,rcutoff2))
 339             {
 340
 341             /* REACTION-FIELD ELECTROSTATICS */
 342             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 343             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 344
 345             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velec            = _mm_and_pd(velec,cutoff_mask);
 349             velecsum         = _mm_add_pd(velecsum,velec);
 350
 351             fscal            = felec;
 352
 353             fscal            = _mm_and_pd(fscal,cutoff_mask);
 354
 355             /* Update vectorial force */
 356             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 357             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 358             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 359
 360             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 361             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 362             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 363
 364             }
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             if (gmx_mm_any_lt(rsq10,rcutoff2))
 371             {
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 375             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 376
 377             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm_and_pd(velec,cutoff_mask);
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             fscal            = _mm_and_pd(fscal,cutoff_mask);
 386
 387             /* Update vectorial force */
 388             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 389             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 390             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 391
 392             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 393             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 394             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 395
 396             }
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             if (gmx_mm_any_lt(rsq11,rcutoff2))
 403             {
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 407             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 408
 409             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velec            = _mm_and_pd(velec,cutoff_mask);
 413             velecsum         = _mm_add_pd(velecsum,velec);
 414
 415             fscal            = felec;
 416
 417             fscal            = _mm_and_pd(fscal,cutoff_mask);
 418
 419             /* Update vectorial force */
 420             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 421             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 422             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 423
 424             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 425             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 426             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 427
 428             }
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             if (gmx_mm_any_lt(rsq12,rcutoff2))
 435             {
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 439             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 440
 441             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velec            = _mm_and_pd(velec,cutoff_mask);
 445             velecsum         = _mm_add_pd(velecsum,velec);
 446
 447             fscal            = felec;
 448
 449             fscal            = _mm_and_pd(fscal,cutoff_mask);
 450
 451             /* Update vectorial force */
 452             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 453             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 454             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 455
 456             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 457             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 458             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 459
 460             }
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             if (gmx_mm_any_lt(rsq20,rcutoff2))
 467             {
 468
 469             /* REACTION-FIELD ELECTROSTATICS */
 470             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 471             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 472
 473             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velec            = _mm_and_pd(velec,cutoff_mask);
 477             velecsum         = _mm_add_pd(velecsum,velec);
 478
 479             fscal            = felec;
 480
 481             fscal            = _mm_and_pd(fscal,cutoff_mask);
 482
 483             /* Update vectorial force */
 484             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 485             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 486             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 487
 488             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 489             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 490             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 491
 492             }
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             if (gmx_mm_any_lt(rsq21,rcutoff2))
 499             {
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 503             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 504
 505             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velec            = _mm_and_pd(velec,cutoff_mask);
 509             velecsum         = _mm_add_pd(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             fscal            = _mm_and_pd(fscal,cutoff_mask);
 514
 515             /* Update vectorial force */
 516             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 517             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 518             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 519
 520             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 521             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 522             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 523
 524             }
 525
 526             /**************************
 527              * CALCULATE INTERACTIONS *
 528              **************************/
 529
 530             if (gmx_mm_any_lt(rsq22,rcutoff2))
 531             {
 532
 533             /* REACTION-FIELD ELECTROSTATICS */
 534             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 535             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 536
 537             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 538
 539             /* Update potential sum for this i atom from the interaction with this j atom. */
 540             velec            = _mm_and_pd(velec,cutoff_mask);
 541             velecsum         = _mm_add_pd(velecsum,velec);
 542
 543             fscal            = felec;
 544
 545             fscal            = _mm_and_pd(fscal,cutoff_mask);
 546
 547             /* Update vectorial force */
 548             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 549             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 550             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 551
 552             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 553             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 554             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 555
 556             }
 557
 558             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 559
 560             /* Inner loop uses 351 flops */
 561         }
 562
 563         if(jidx<j_index_end)
 564         {
 565
 566             jnrA             = jjnr[jidx];
 567             j_coord_offsetA  = DIM*jnrA;
 568
 569             /* load j atom coordinates */
 570             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 571                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 572
 573             /* Calculate displacement vector */
 574             dx00             = _mm_sub_pd(ix0,jx0);
 575             dy00             = _mm_sub_pd(iy0,jy0);
 576             dz00             = _mm_sub_pd(iz0,jz0);
 577             dx01             = _mm_sub_pd(ix0,jx1);
 578             dy01             = _mm_sub_pd(iy0,jy1);
 579             dz01             = _mm_sub_pd(iz0,jz1);
 580             dx02             = _mm_sub_pd(ix0,jx2);
 581             dy02             = _mm_sub_pd(iy0,jy2);
 582             dz02             = _mm_sub_pd(iz0,jz2);
 583             dx10             = _mm_sub_pd(ix1,jx0);
 584             dy10             = _mm_sub_pd(iy1,jy0);
 585             dz10             = _mm_sub_pd(iz1,jz0);
 586             dx11             = _mm_sub_pd(ix1,jx1);
 587             dy11             = _mm_sub_pd(iy1,jy1);
 588             dz11             = _mm_sub_pd(iz1,jz1);
 589             dx12             = _mm_sub_pd(ix1,jx2);
 590             dy12             = _mm_sub_pd(iy1,jy2);
 591             dz12             = _mm_sub_pd(iz1,jz2);
 592             dx20             = _mm_sub_pd(ix2,jx0);
 593             dy20             = _mm_sub_pd(iy2,jy0);
 594             dz20             = _mm_sub_pd(iz2,jz0);
 595             dx21             = _mm_sub_pd(ix2,jx1);
 596             dy21             = _mm_sub_pd(iy2,jy1);
 597             dz21             = _mm_sub_pd(iz2,jz1);
 598             dx22             = _mm_sub_pd(ix2,jx2);
 599             dy22             = _mm_sub_pd(iy2,jy2);
 600             dz22             = _mm_sub_pd(iz2,jz2);
 601
 602             /* Calculate squared distance and things based on it */
 603             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 604             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 605             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 606             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 607             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 608             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 609             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 610             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 611             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 612
 613             rinv00           = avx128fma_invsqrt_d(rsq00);
 614             rinv01           = avx128fma_invsqrt_d(rsq01);
 615             rinv02           = avx128fma_invsqrt_d(rsq02);
 616             rinv10           = avx128fma_invsqrt_d(rsq10);
 617             rinv11           = avx128fma_invsqrt_d(rsq11);
 618             rinv12           = avx128fma_invsqrt_d(rsq12);
 619             rinv20           = avx128fma_invsqrt_d(rsq20);
 620             rinv21           = avx128fma_invsqrt_d(rsq21);
 621             rinv22           = avx128fma_invsqrt_d(rsq22);
 622
 623             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 624             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 625             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 626             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 627             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 628             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 629             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 630             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 631             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 632
 633             fjx0             = _mm_setzero_pd();
 634             fjy0             = _mm_setzero_pd();
 635             fjz0             = _mm_setzero_pd();
 636             fjx1             = _mm_setzero_pd();
 637             fjy1             = _mm_setzero_pd();
 638             fjz1             = _mm_setzero_pd();
 639             fjx2             = _mm_setzero_pd();
 640             fjy2             = _mm_setzero_pd();
 641             fjz2             = _mm_setzero_pd();
 642
 643             /**************************
 644              * CALCULATE INTERACTIONS *
 645              **************************/
 646
 647             if (gmx_mm_any_lt(rsq00,rcutoff2))
 648             {
 649
 650             /* REACTION-FIELD ELECTROSTATICS */
 651             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 652             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 653
 654             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 655
 656             /* Update potential sum for this i atom from the interaction with this j atom. */
 657             velec            = _mm_and_pd(velec,cutoff_mask);
 658             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 659             velecsum         = _mm_add_pd(velecsum,velec);
 660
 661             fscal            = felec;
 662
 663             fscal            = _mm_and_pd(fscal,cutoff_mask);
 664
 665             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 666
 667             /* Update vectorial force */
 668             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 669             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 670             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 671
 672             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 673             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 674             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 675
 676             }
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             if (gmx_mm_any_lt(rsq01,rcutoff2))
 683             {
 684
 685             /* REACTION-FIELD ELECTROSTATICS */
 686             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 687             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 688
 689             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 690
 691             /* Update potential sum for this i atom from the interaction with this j atom. */
 692             velec            = _mm_and_pd(velec,cutoff_mask);
 693             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 694             velecsum         = _mm_add_pd(velecsum,velec);
 695
 696             fscal            = felec;
 697
 698             fscal            = _mm_and_pd(fscal,cutoff_mask);
 699
 700             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 701
 702             /* Update vectorial force */
 703             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 704             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 705             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 706
 707             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 708             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 709             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 710
 711             }
 712
 713             /**************************
 714              * CALCULATE INTERACTIONS *
 715              **************************/
 716
 717             if (gmx_mm_any_lt(rsq02,rcutoff2))
 718             {
 719
 720             /* REACTION-FIELD ELECTROSTATICS */
 721             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 722             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 723
 724             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 725
 726             /* Update potential sum for this i atom from the interaction with this j atom. */
 727             velec            = _mm_and_pd(velec,cutoff_mask);
 728             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 729             velecsum         = _mm_add_pd(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_and_pd(fscal,cutoff_mask);
 734
 735             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 736
 737             /* Update vectorial force */
 738             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 739             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 740             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 741
 742             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 743             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 744             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 745
 746             }
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             if (gmx_mm_any_lt(rsq10,rcutoff2))
 753             {
 754
 755             /* REACTION-FIELD ELECTROSTATICS */
 756             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 757             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 758
 759             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 760
 761             /* Update potential sum for this i atom from the interaction with this j atom. */
 762             velec            = _mm_and_pd(velec,cutoff_mask);
 763             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 764             velecsum         = _mm_add_pd(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_and_pd(fscal,cutoff_mask);
 769
 770             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 771
 772             /* Update vectorial force */
 773             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 774             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 775             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 776
 777             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 778             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 779             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 780
 781             }
 782
 783             /**************************
 784              * CALCULATE INTERACTIONS *
 785              **************************/
 786
 787             if (gmx_mm_any_lt(rsq11,rcutoff2))
 788             {
 789
 790             /* REACTION-FIELD ELECTROSTATICS */
 791             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 792             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 793
 794             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 795
 796             /* Update potential sum for this i atom from the interaction with this j atom. */
 797             velec            = _mm_and_pd(velec,cutoff_mask);
 798             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 799             velecsum         = _mm_add_pd(velecsum,velec);
 800
 801             fscal            = felec;
 802
 803             fscal            = _mm_and_pd(fscal,cutoff_mask);
 804
 805             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 806
 807             /* Update vectorial force */
 808             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 809             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 810             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 811
 812             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 813             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 814             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 815
 816             }
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             if (gmx_mm_any_lt(rsq12,rcutoff2))
 823             {
 824
 825             /* REACTION-FIELD ELECTROSTATICS */
 826             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 827             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 828
 829             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 830
 831             /* Update potential sum for this i atom from the interaction with this j atom. */
 832             velec            = _mm_and_pd(velec,cutoff_mask);
 833             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 834             velecsum         = _mm_add_pd(velecsum,velec);
 835
 836             fscal            = felec;
 837
 838             fscal            = _mm_and_pd(fscal,cutoff_mask);
 839
 840             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 841
 842             /* Update vectorial force */
 843             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 844             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 845             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 846
 847             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 848             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 849             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 850
 851             }
 852
 853             /**************************
 854              * CALCULATE INTERACTIONS *
 855              **************************/
 856
 857             if (gmx_mm_any_lt(rsq20,rcutoff2))
 858             {
 859
 860             /* REACTION-FIELD ELECTROSTATICS */
 861             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 862             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 863
 864             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 865
 866             /* Update potential sum for this i atom from the interaction with this j atom. */
 867             velec            = _mm_and_pd(velec,cutoff_mask);
 868             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 869             velecsum         = _mm_add_pd(velecsum,velec);
 870
 871             fscal            = felec;
 872
 873             fscal            = _mm_and_pd(fscal,cutoff_mask);
 874
 875             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 876
 877             /* Update vectorial force */
 878             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 879             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 880             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 881
 882             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 883             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 884             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 885
 886             }
 887
 888             /**************************
 889              * CALCULATE INTERACTIONS *
 890              **************************/
 891
 892             if (gmx_mm_any_lt(rsq21,rcutoff2))
 893             {
 894
 895             /* REACTION-FIELD ELECTROSTATICS */
 896             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 897             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 898
 899             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 900
 901             /* Update potential sum for this i atom from the interaction with this j atom. */
 902             velec            = _mm_and_pd(velec,cutoff_mask);
 903             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 904             velecsum         = _mm_add_pd(velecsum,velec);
 905
 906             fscal            = felec;
 907
 908             fscal            = _mm_and_pd(fscal,cutoff_mask);
 909
 910             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 911
 912             /* Update vectorial force */
 913             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 914             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 915             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 916
 917             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 918             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 919             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 920
 921             }
 922
 923             /**************************
 924              * CALCULATE INTERACTIONS *
 925              **************************/
 926
 927             if (gmx_mm_any_lt(rsq22,rcutoff2))
 928             {
 929
 930             /* REACTION-FIELD ELECTROSTATICS */
 931             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 932             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 933
 934             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 935
 936             /* Update potential sum for this i atom from the interaction with this j atom. */
 937             velec            = _mm_and_pd(velec,cutoff_mask);
 938             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 939             velecsum         = _mm_add_pd(velecsum,velec);
 940
 941             fscal            = felec;
 942
 943             fscal            = _mm_and_pd(fscal,cutoff_mask);
 944
 945             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 946
 947             /* Update vectorial force */
 948             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 949             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 950             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 951
 952             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 953             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 954             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 955
 956             }
 957
 958             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 959
 960             /* Inner loop uses 351 flops */
 961         }
 962
 963         /* End of innermost loop */
 964
 965         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 966                                               f+i_coord_offset,fshift+i_shift_offset);
 967
 968         ggid                        = gid[iidx];
 969         /* Update potential energies */
 970         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 971
 972         /* Increment number of inner iterations */
 973         inneriter                  += j_index_end - j_index_start;
 974
 975         /* Outer loop uses 19 flops */
 976     }
 977
 978     /* Increment number of outer iterations */
 979     outeriter        += nri;
 980
 981     /* Update outer/inner flops */
 982
 983     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*351);
 984 }
 985 /*
 986  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_avx_128_fma_double
 987  * Electrostatics interaction: ReactionField
 988  * VdW interaction:            None
 989  * Geometry:                   Water3-Water3
 990  * Calculate force/pot:        Force
 991  */
 992 void
 993 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_avx_128_fma_double
 994                     (t_nblist                    * gmx_restrict       nlist,
 995                      rvec                        * gmx_restrict          xx,
 996                      rvec                        * gmx_restrict          ff,
 997                      struct t_forcerec           * gmx_restrict          fr,
 998                      t_mdatoms                   * gmx_restrict     mdatoms,
 999                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1000                      t_nrnb                      * gmx_restrict        nrnb)
1001 {
1002     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1003      * just 0 for non-waters.
1004      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1005      * jnr indices corresponding to data put in the four positions in the SIMD register.
1006      */
1007     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1008     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1009     int              jnrA,jnrB;
1010     int              j_coord_offsetA,j_coord_offsetB;
1011     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1012     real             rcutoff_scalar;
1013     real             *shiftvec,*fshift,*x,*f;
1014     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1015     int              vdwioffset0;
1016     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1017     int              vdwioffset1;
1018     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1019     int              vdwioffset2;
1020     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1021     int              vdwjidx0A,vdwjidx0B;
1022     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1023     int              vdwjidx1A,vdwjidx1B;
1024     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1025     int              vdwjidx2A,vdwjidx2B;
1026     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1027     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1028     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1029     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1030     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1031     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1032     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1033     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1034     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1035     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1036     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1037     real             *charge;
1038     __m128d          dummy_mask,cutoff_mask;
1039     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1040     __m128d          one     = _mm_set1_pd(1.0);
1041     __m128d          two     = _mm_set1_pd(2.0);
1042     x                = xx[0];
1043     f                = ff[0];
1044
1045     nri              = nlist->nri;
1046     iinr             = nlist->iinr;
1047     jindex           = nlist->jindex;
1048     jjnr             = nlist->jjnr;
1049     shiftidx         = nlist->shift;
1050     gid              = nlist->gid;
1051     shiftvec         = fr->shift_vec[0];
1052     fshift           = fr->fshift[0];
1053     facel            = _mm_set1_pd(fr->ic->epsfac);
1054     charge           = mdatoms->chargeA;
1055     krf              = _mm_set1_pd(fr->ic->k_rf);
1056     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1057     crf              = _mm_set1_pd(fr->ic->c_rf);
1058
1059     /* Setup water-specific parameters */
1060     inr              = nlist->iinr[0];
1061     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1062     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1063     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1064
1065     jq0              = _mm_set1_pd(charge[inr+0]);
1066     jq1              = _mm_set1_pd(charge[inr+1]);
1067     jq2              = _mm_set1_pd(charge[inr+2]);
1068     qq00             = _mm_mul_pd(iq0,jq0);
1069     qq01             = _mm_mul_pd(iq0,jq1);
1070     qq02             = _mm_mul_pd(iq0,jq2);
1071     qq10             = _mm_mul_pd(iq1,jq0);
1072     qq11             = _mm_mul_pd(iq1,jq1);
1073     qq12             = _mm_mul_pd(iq1,jq2);
1074     qq20             = _mm_mul_pd(iq2,jq0);
1075     qq21             = _mm_mul_pd(iq2,jq1);
1076     qq22             = _mm_mul_pd(iq2,jq2);
1077
1078     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1079     rcutoff_scalar   = fr->ic->rcoulomb;
1080     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1081     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1082
1083     /* Avoid stupid compiler warnings */
1084     jnrA = jnrB = 0;
1085     j_coord_offsetA = 0;
1086     j_coord_offsetB = 0;
1087
1088     outeriter        = 0;
1089     inneriter        = 0;
1090
1091     /* Start outer loop over neighborlists */
1092     for(iidx=0; iidx<nri; iidx++)
1093     {
1094         /* Load shift vector for this list */
1095         i_shift_offset   = DIM*shiftidx[iidx];
1096
1097         /* Load limits for loop over neighbors */
1098         j_index_start    = jindex[iidx];
1099         j_index_end      = jindex[iidx+1];
1100
1101         /* Get outer coordinate index */
1102         inr              = iinr[iidx];
1103         i_coord_offset   = DIM*inr;
1104
1105         /* Load i particle coords and add shift vector */
1106         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1107                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1108
1109         fix0             = _mm_setzero_pd();
1110         fiy0             = _mm_setzero_pd();
1111         fiz0             = _mm_setzero_pd();
1112         fix1             = _mm_setzero_pd();
1113         fiy1             = _mm_setzero_pd();
1114         fiz1             = _mm_setzero_pd();
1115         fix2             = _mm_setzero_pd();
1116         fiy2             = _mm_setzero_pd();
1117         fiz2             = _mm_setzero_pd();
1118
1119         /* Start inner kernel loop */
1120         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1121         {
1122
1123             /* Get j neighbor index, and coordinate index */
1124             jnrA             = jjnr[jidx];
1125             jnrB             = jjnr[jidx+1];
1126             j_coord_offsetA  = DIM*jnrA;
1127             j_coord_offsetB  = DIM*jnrB;
1128
1129             /* load j atom coordinates */
1130             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1131                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1132
1133             /* Calculate displacement vector */
1134             dx00             = _mm_sub_pd(ix0,jx0);
1135             dy00             = _mm_sub_pd(iy0,jy0);
1136             dz00             = _mm_sub_pd(iz0,jz0);
1137             dx01             = _mm_sub_pd(ix0,jx1);
1138             dy01             = _mm_sub_pd(iy0,jy1);
1139             dz01             = _mm_sub_pd(iz0,jz1);
1140             dx02             = _mm_sub_pd(ix0,jx2);
1141             dy02             = _mm_sub_pd(iy0,jy2);
1142             dz02             = _mm_sub_pd(iz0,jz2);
1143             dx10             = _mm_sub_pd(ix1,jx0);
1144             dy10             = _mm_sub_pd(iy1,jy0);
1145             dz10             = _mm_sub_pd(iz1,jz0);
1146             dx11             = _mm_sub_pd(ix1,jx1);
1147             dy11             = _mm_sub_pd(iy1,jy1);
1148             dz11             = _mm_sub_pd(iz1,jz1);
1149             dx12             = _mm_sub_pd(ix1,jx2);
1150             dy12             = _mm_sub_pd(iy1,jy2);
1151             dz12             = _mm_sub_pd(iz1,jz2);
1152             dx20             = _mm_sub_pd(ix2,jx0);
1153             dy20             = _mm_sub_pd(iy2,jy0);
1154             dz20             = _mm_sub_pd(iz2,jz0);
1155             dx21             = _mm_sub_pd(ix2,jx1);
1156             dy21             = _mm_sub_pd(iy2,jy1);
1157             dz21             = _mm_sub_pd(iz2,jz1);
1158             dx22             = _mm_sub_pd(ix2,jx2);
1159             dy22             = _mm_sub_pd(iy2,jy2);
1160             dz22             = _mm_sub_pd(iz2,jz2);
1161
1162             /* Calculate squared distance and things based on it */
1163             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1164             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1165             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1166             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1167             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1168             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1169             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1170             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1171             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1172
1173             rinv00           = avx128fma_invsqrt_d(rsq00);
1174             rinv01           = avx128fma_invsqrt_d(rsq01);
1175             rinv02           = avx128fma_invsqrt_d(rsq02);
1176             rinv10           = avx128fma_invsqrt_d(rsq10);
1177             rinv11           = avx128fma_invsqrt_d(rsq11);
1178             rinv12           = avx128fma_invsqrt_d(rsq12);
1179             rinv20           = avx128fma_invsqrt_d(rsq20);
1180             rinv21           = avx128fma_invsqrt_d(rsq21);
1181             rinv22           = avx128fma_invsqrt_d(rsq22);
1182
1183             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1184             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1185             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1186             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1187             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1188             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1189             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1190             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1191             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1192
1193             fjx0             = _mm_setzero_pd();
1194             fjy0             = _mm_setzero_pd();
1195             fjz0             = _mm_setzero_pd();
1196             fjx1             = _mm_setzero_pd();
1197             fjy1             = _mm_setzero_pd();
1198             fjz1             = _mm_setzero_pd();
1199             fjx2             = _mm_setzero_pd();
1200             fjy2             = _mm_setzero_pd();
1201             fjz2             = _mm_setzero_pd();
1202
1203             /**************************
1204              * CALCULATE INTERACTIONS *
1205              **************************/
1206
1207             if (gmx_mm_any_lt(rsq00,rcutoff2))
1208             {
1209
1210             /* REACTION-FIELD ELECTROSTATICS */
1211             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1212
1213             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1214
1215             fscal            = felec;
1216
1217             fscal            = _mm_and_pd(fscal,cutoff_mask);
1218
1219             /* Update vectorial force */
1220             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1221             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1222             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1223
1224             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1225             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1226             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1227
1228             }
1229
1230             /**************************
1231              * CALCULATE INTERACTIONS *
1232              **************************/
1233
1234             if (gmx_mm_any_lt(rsq01,rcutoff2))
1235             {
1236
1237             /* REACTION-FIELD ELECTROSTATICS */
1238             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1239
1240             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1241
1242             fscal            = felec;
1243
1244             fscal            = _mm_and_pd(fscal,cutoff_mask);
1245
1246             /* Update vectorial force */
1247             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1248             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1249             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1250
1251             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1252             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1253             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1254
1255             }
1256
1257             /**************************
1258              * CALCULATE INTERACTIONS *
1259              **************************/
1260
1261             if (gmx_mm_any_lt(rsq02,rcutoff2))
1262             {
1263
1264             /* REACTION-FIELD ELECTROSTATICS */
1265             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1266
1267             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1268
1269             fscal            = felec;
1270
1271             fscal            = _mm_and_pd(fscal,cutoff_mask);
1272
1273             /* Update vectorial force */
1274             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1275             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1276             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1277
1278             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1279             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1280             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1281
1282             }
1283
1284             /**************************
1285              * CALCULATE INTERACTIONS *
1286              **************************/
1287
1288             if (gmx_mm_any_lt(rsq10,rcutoff2))
1289             {
1290
1291             /* REACTION-FIELD ELECTROSTATICS */
1292             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1293
1294             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1295
1296             fscal            = felec;
1297
1298             fscal            = _mm_and_pd(fscal,cutoff_mask);
1299
1300             /* Update vectorial force */
1301             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1302             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1303             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1304
1305             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1306             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1307             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1308
1309             }
1310
1311             /**************************
1312              * CALCULATE INTERACTIONS *
1313              **************************/
1314
1315             if (gmx_mm_any_lt(rsq11,rcutoff2))
1316             {
1317
1318             /* REACTION-FIELD ELECTROSTATICS */
1319             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1320
1321             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1322
1323             fscal            = felec;
1324
1325             fscal            = _mm_and_pd(fscal,cutoff_mask);
1326
1327             /* Update vectorial force */
1328             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1329             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1330             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1331
1332             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1333             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1334             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1335
1336             }
1337
1338             /**************************
1339              * CALCULATE INTERACTIONS *
1340              **************************/
1341
1342             if (gmx_mm_any_lt(rsq12,rcutoff2))
1343             {
1344
1345             /* REACTION-FIELD ELECTROSTATICS */
1346             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1347
1348             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1349
1350             fscal            = felec;
1351
1352             fscal            = _mm_and_pd(fscal,cutoff_mask);
1353
1354             /* Update vectorial force */
1355             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1356             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1357             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1358
1359             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1360             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1361             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1362
1363             }
1364
1365             /**************************
1366              * CALCULATE INTERACTIONS *
1367              **************************/
1368
1369             if (gmx_mm_any_lt(rsq20,rcutoff2))
1370             {
1371
1372             /* REACTION-FIELD ELECTROSTATICS */
1373             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1374
1375             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1376
1377             fscal            = felec;
1378
1379             fscal            = _mm_and_pd(fscal,cutoff_mask);
1380
1381             /* Update vectorial force */
1382             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1383             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1384             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1385
1386             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1387             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1388             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1389
1390             }
1391
1392             /**************************
1393              * CALCULATE INTERACTIONS *
1394              **************************/
1395
1396             if (gmx_mm_any_lt(rsq21,rcutoff2))
1397             {
1398
1399             /* REACTION-FIELD ELECTROSTATICS */
1400             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1401
1402             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1403
1404             fscal            = felec;
1405
1406             fscal            = _mm_and_pd(fscal,cutoff_mask);
1407
1408             /* Update vectorial force */
1409             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1410             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1411             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1412
1413             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1414             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1415             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1416
1417             }
1418
1419             /**************************
1420              * CALCULATE INTERACTIONS *
1421              **************************/
1422
1423             if (gmx_mm_any_lt(rsq22,rcutoff2))
1424             {
1425
1426             /* REACTION-FIELD ELECTROSTATICS */
1427             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1428
1429             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1430
1431             fscal            = felec;
1432
1433             fscal            = _mm_and_pd(fscal,cutoff_mask);
1434
1435             /* Update vectorial force */
1436             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1437             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1438             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1439
1440             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1441             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1442             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1443
1444             }
1445
1446             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1447
1448             /* Inner loop uses 297 flops */
1449         }
1450
1451         if(jidx<j_index_end)
1452         {
1453
1454             jnrA             = jjnr[jidx];
1455             j_coord_offsetA  = DIM*jnrA;
1456
1457             /* load j atom coordinates */
1458             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1459                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1460
1461             /* Calculate displacement vector */
1462             dx00             = _mm_sub_pd(ix0,jx0);
1463             dy00             = _mm_sub_pd(iy0,jy0);
1464             dz00             = _mm_sub_pd(iz0,jz0);
1465             dx01             = _mm_sub_pd(ix0,jx1);
1466             dy01             = _mm_sub_pd(iy0,jy1);
1467             dz01             = _mm_sub_pd(iz0,jz1);
1468             dx02             = _mm_sub_pd(ix0,jx2);
1469             dy02             = _mm_sub_pd(iy0,jy2);
1470             dz02             = _mm_sub_pd(iz0,jz2);
1471             dx10             = _mm_sub_pd(ix1,jx0);
1472             dy10             = _mm_sub_pd(iy1,jy0);
1473             dz10             = _mm_sub_pd(iz1,jz0);
1474             dx11             = _mm_sub_pd(ix1,jx1);
1475             dy11             = _mm_sub_pd(iy1,jy1);
1476             dz11             = _mm_sub_pd(iz1,jz1);
1477             dx12             = _mm_sub_pd(ix1,jx2);
1478             dy12             = _mm_sub_pd(iy1,jy2);
1479             dz12             = _mm_sub_pd(iz1,jz2);
1480             dx20             = _mm_sub_pd(ix2,jx0);
1481             dy20             = _mm_sub_pd(iy2,jy0);
1482             dz20             = _mm_sub_pd(iz2,jz0);
1483             dx21             = _mm_sub_pd(ix2,jx1);
1484             dy21             = _mm_sub_pd(iy2,jy1);
1485             dz21             = _mm_sub_pd(iz2,jz1);
1486             dx22             = _mm_sub_pd(ix2,jx2);
1487             dy22             = _mm_sub_pd(iy2,jy2);
1488             dz22             = _mm_sub_pd(iz2,jz2);
1489
1490             /* Calculate squared distance and things based on it */
1491             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1492             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1493             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1494             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1495             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1496             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1497             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1498             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1499             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1500
1501             rinv00           = avx128fma_invsqrt_d(rsq00);
1502             rinv01           = avx128fma_invsqrt_d(rsq01);
1503             rinv02           = avx128fma_invsqrt_d(rsq02);
1504             rinv10           = avx128fma_invsqrt_d(rsq10);
1505             rinv11           = avx128fma_invsqrt_d(rsq11);
1506             rinv12           = avx128fma_invsqrt_d(rsq12);
1507             rinv20           = avx128fma_invsqrt_d(rsq20);
1508             rinv21           = avx128fma_invsqrt_d(rsq21);
1509             rinv22           = avx128fma_invsqrt_d(rsq22);
1510
1511             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1512             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1513             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1514             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1515             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1516             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1517             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1518             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1519             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1520
1521             fjx0             = _mm_setzero_pd();
1522             fjy0             = _mm_setzero_pd();
1523             fjz0             = _mm_setzero_pd();
1524             fjx1             = _mm_setzero_pd();
1525             fjy1             = _mm_setzero_pd();
1526             fjz1             = _mm_setzero_pd();
1527             fjx2             = _mm_setzero_pd();
1528             fjy2             = _mm_setzero_pd();
1529             fjz2             = _mm_setzero_pd();
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             if (gmx_mm_any_lt(rsq00,rcutoff2))
1536             {
1537
1538             /* REACTION-FIELD ELECTROSTATICS */
1539             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1540
1541             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_and_pd(fscal,cutoff_mask);
1546
1547             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1548
1549             /* Update vectorial force */
1550             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1551             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1552             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1553
1554             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1555             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1556             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1557
1558             }
1559
1560             /**************************
1561              * CALCULATE INTERACTIONS *
1562              **************************/
1563
1564             if (gmx_mm_any_lt(rsq01,rcutoff2))
1565             {
1566
1567             /* REACTION-FIELD ELECTROSTATICS */
1568             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1569
1570             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1571
1572             fscal            = felec;
1573
1574             fscal            = _mm_and_pd(fscal,cutoff_mask);
1575
1576             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1577
1578             /* Update vectorial force */
1579             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1580             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1581             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1582
1583             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1584             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1585             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1586
1587             }
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             if (gmx_mm_any_lt(rsq02,rcutoff2))
1594             {
1595
1596             /* REACTION-FIELD ELECTROSTATICS */
1597             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1598
1599             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1600
1601             fscal            = felec;
1602
1603             fscal            = _mm_and_pd(fscal,cutoff_mask);
1604
1605             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1606
1607             /* Update vectorial force */
1608             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1609             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1610             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1611
1612             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1613             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1614             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1615
1616             }
1617
1618             /**************************
1619              * CALCULATE INTERACTIONS *
1620              **************************/
1621
1622             if (gmx_mm_any_lt(rsq10,rcutoff2))
1623             {
1624
1625             /* REACTION-FIELD ELECTROSTATICS */
1626             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1627
1628             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1629
1630             fscal            = felec;
1631
1632             fscal            = _mm_and_pd(fscal,cutoff_mask);
1633
1634             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1635
1636             /* Update vectorial force */
1637             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1638             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1639             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1640
1641             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1642             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1643             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1644
1645             }
1646
1647             /**************************
1648              * CALCULATE INTERACTIONS *
1649              **************************/
1650
1651             if (gmx_mm_any_lt(rsq11,rcutoff2))
1652             {
1653
1654             /* REACTION-FIELD ELECTROSTATICS */
1655             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1656
1657             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1658
1659             fscal            = felec;
1660
1661             fscal            = _mm_and_pd(fscal,cutoff_mask);
1662
1663             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1664
1665             /* Update vectorial force */
1666             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1667             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1668             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1669
1670             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1671             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1672             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1673
1674             }
1675
1676             /**************************
1677              * CALCULATE INTERACTIONS *
1678              **************************/
1679
1680             if (gmx_mm_any_lt(rsq12,rcutoff2))
1681             {
1682
1683             /* REACTION-FIELD ELECTROSTATICS */
1684             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1685
1686             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1687
1688             fscal            = felec;
1689
1690             fscal            = _mm_and_pd(fscal,cutoff_mask);
1691
1692             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1693
1694             /* Update vectorial force */
1695             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1696             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1697             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1698
1699             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1700             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1701             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1702
1703             }
1704
1705             /**************************
1706              * CALCULATE INTERACTIONS *
1707              **************************/
1708
1709             if (gmx_mm_any_lt(rsq20,rcutoff2))
1710             {
1711
1712             /* REACTION-FIELD ELECTROSTATICS */
1713             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1714
1715             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1716
1717             fscal            = felec;
1718
1719             fscal            = _mm_and_pd(fscal,cutoff_mask);
1720
1721             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1722
1723             /* Update vectorial force */
1724             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1725             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1726             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1727
1728             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1729             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1730             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1731
1732             }
1733
1734             /**************************
1735              * CALCULATE INTERACTIONS *
1736              **************************/
1737
1738             if (gmx_mm_any_lt(rsq21,rcutoff2))
1739             {
1740
1741             /* REACTION-FIELD ELECTROSTATICS */
1742             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1743
1744             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1745
1746             fscal            = felec;
1747
1748             fscal            = _mm_and_pd(fscal,cutoff_mask);
1749
1750             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1751
1752             /* Update vectorial force */
1753             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1754             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1755             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1756
1757             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1758             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1759             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1760
1761             }
1762
1763             /**************************
1764              * CALCULATE INTERACTIONS *
1765              **************************/
1766
1767             if (gmx_mm_any_lt(rsq22,rcutoff2))
1768             {
1769
1770             /* REACTION-FIELD ELECTROSTATICS */
1771             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1772
1773             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1774
1775             fscal            = felec;
1776
1777             fscal            = _mm_and_pd(fscal,cutoff_mask);
1778
1779             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1780
1781             /* Update vectorial force */
1782             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1783             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1784             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1785
1786             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1787             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1788             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1789
1790             }
1791
1792             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1793
1794             /* Inner loop uses 297 flops */
1795         }
1796
1797         /* End of innermost loop */
1798
1799         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1800                                               f+i_coord_offset,fshift+i_shift_offset);
1801
1802         /* Increment number of inner iterations */
1803         inneriter                  += j_index_end - j_index_start;
1804
1805         /* Outer loop uses 18 flops */
1806     }
1807
1808     /* Increment number of outer iterations */
1809     outeriter        += nri;
1810
1811     /* Update outer/inner flops */
1812
1813     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*297);
1814 }