src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwNone_GeomW4P1_avx_128_fma_single.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_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwNone_GeomW4P1_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset1;
  83     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  91     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  92     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  93     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     __m128           dummy_mask,cutoff_mask;
  96     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  97     __m128           one     = _mm_set1_ps(1.0);
  98     __m128           two     = _mm_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_ps(fr->ic->epsfac);
 111     charge           = mdatoms->chargeA;
 112
 113     /* Setup water-specific parameters */
 114     inr              = nlist->iinr[0];
 115     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 116     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 117     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 118
 119     /* Avoid stupid compiler warnings */
 120     jnrA = jnrB = jnrC = jnrD = 0;
 121     j_coord_offsetA = 0;
 122     j_coord_offsetB = 0;
 123     j_coord_offsetC = 0;
 124     j_coord_offsetD = 0;
 125
 126     outeriter        = 0;
 127     inneriter        = 0;
 128
 129     for(iidx=0;iidx<4*DIM;iidx++)
 130     {
 131         scratch[iidx] = 0.0;
 132     }
 133
 134     /* Start outer loop over neighborlists */
 135     for(iidx=0; iidx<nri; iidx++)
 136     {
 137         /* Load shift vector for this list */
 138         i_shift_offset   = DIM*shiftidx[iidx];
 139
 140         /* Load limits for loop over neighbors */
 141         j_index_start    = jindex[iidx];
 142         j_index_end      = jindex[iidx+1];
 143
 144         /* Get outer coordinate index */
 145         inr              = iinr[iidx];
 146         i_coord_offset   = DIM*inr;
 147
 148         /* Load i particle coords and add shift vector */
 149         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 150                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 151
 152         fix1             = _mm_setzero_ps();
 153         fiy1             = _mm_setzero_ps();
 154         fiz1             = _mm_setzero_ps();
 155         fix2             = _mm_setzero_ps();
 156         fiy2             = _mm_setzero_ps();
 157         fiz2             = _mm_setzero_ps();
 158         fix3             = _mm_setzero_ps();
 159         fiy3             = _mm_setzero_ps();
 160         fiz3             = _mm_setzero_ps();
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_ps();
 164
 165         /* Start inner kernel loop */
 166         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 167         {
 168
 169             /* Get j neighbor index, and coordinate index */
 170             jnrA             = jjnr[jidx];
 171             jnrB             = jjnr[jidx+1];
 172             jnrC             = jjnr[jidx+2];
 173             jnrD             = jjnr[jidx+3];
 174             j_coord_offsetA  = DIM*jnrA;
 175             j_coord_offsetB  = DIM*jnrB;
 176             j_coord_offsetC  = DIM*jnrC;
 177             j_coord_offsetD  = DIM*jnrD;
 178
 179             /* load j atom coordinates */
 180             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 181                                               x+j_coord_offsetC,x+j_coord_offsetD,
 182                                               &jx0,&jy0,&jz0);
 183
 184             /* Calculate displacement vector */
 185             dx10             = _mm_sub_ps(ix1,jx0);
 186             dy10             = _mm_sub_ps(iy1,jy0);
 187             dz10             = _mm_sub_ps(iz1,jz0);
 188             dx20             = _mm_sub_ps(ix2,jx0);
 189             dy20             = _mm_sub_ps(iy2,jy0);
 190             dz20             = _mm_sub_ps(iz2,jz0);
 191             dx30             = _mm_sub_ps(ix3,jx0);
 192             dy30             = _mm_sub_ps(iy3,jy0);
 193             dz30             = _mm_sub_ps(iz3,jz0);
 194
 195             /* Calculate squared distance and things based on it */
 196             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 197             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 198             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 199
 200             rinv10           = avx128fma_invsqrt_f(rsq10);
 201             rinv20           = avx128fma_invsqrt_f(rsq20);
 202             rinv30           = avx128fma_invsqrt_f(rsq30);
 203
 204             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 205             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 206             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 207
 208             /* Load parameters for j particles */
 209             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 210                                                               charge+jnrC+0,charge+jnrD+0);
 211
 212             fjx0             = _mm_setzero_ps();
 213             fjy0             = _mm_setzero_ps();
 214             fjz0             = _mm_setzero_ps();
 215
 216             /**************************
 217              * CALCULATE INTERACTIONS *
 218              **************************/
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq10             = _mm_mul_ps(iq1,jq0);
 222
 223             /* COULOMB ELECTROSTATICS */
 224             velec            = _mm_mul_ps(qq10,rinv10);
 225             felec            = _mm_mul_ps(velec,rinvsq10);
 226
 227             /* Update potential sum for this i atom from the interaction with this j atom. */
 228             velecsum         = _mm_add_ps(velecsum,velec);
 229
 230             fscal            = felec;
 231
 232              /* Update vectorial force */
 233             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 234             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 235             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 236
 237             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 238             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 239             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 240
 241             /**************************
 242              * CALCULATE INTERACTIONS *
 243              **************************/
 244
 245             /* Compute parameters for interactions between i and j atoms */
 246             qq20             = _mm_mul_ps(iq2,jq0);
 247
 248             /* COULOMB ELECTROSTATICS */
 249             velec            = _mm_mul_ps(qq20,rinv20);
 250             felec            = _mm_mul_ps(velec,rinvsq20);
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velecsum         = _mm_add_ps(velecsum,velec);
 254
 255             fscal            = felec;
 256
 257              /* Update vectorial force */
 258             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 259             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 260             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 261
 262             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 263             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 264             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 265
 266             /**************************
 267              * CALCULATE INTERACTIONS *
 268              **************************/
 269
 270             /* Compute parameters for interactions between i and j atoms */
 271             qq30             = _mm_mul_ps(iq3,jq0);
 272
 273             /* COULOMB ELECTROSTATICS */
 274             velec            = _mm_mul_ps(qq30,rinv30);
 275             felec            = _mm_mul_ps(velec,rinvsq30);
 276
 277             /* Update potential sum for this i atom from the interaction with this j atom. */
 278             velecsum         = _mm_add_ps(velecsum,velec);
 279
 280             fscal            = felec;
 281
 282              /* Update vectorial force */
 283             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 284             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 285             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 286
 287             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 288             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 289             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 290
 291             fjptrA             = f+j_coord_offsetA;
 292             fjptrB             = f+j_coord_offsetB;
 293             fjptrC             = f+j_coord_offsetC;
 294             fjptrD             = f+j_coord_offsetD;
 295
 296             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 297
 298             /* Inner loop uses 93 flops */
 299         }
 300
 301         if(jidx<j_index_end)
 302         {
 303
 304             /* Get j neighbor index, and coordinate index */
 305             jnrlistA         = jjnr[jidx];
 306             jnrlistB         = jjnr[jidx+1];
 307             jnrlistC         = jjnr[jidx+2];
 308             jnrlistD         = jjnr[jidx+3];
 309             /* Sign of each element will be negative for non-real atoms.
 310              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 311              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 312              */
 313             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 314             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 315             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 316             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 317             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 318             j_coord_offsetA  = DIM*jnrA;
 319             j_coord_offsetB  = DIM*jnrB;
 320             j_coord_offsetC  = DIM*jnrC;
 321             j_coord_offsetD  = DIM*jnrD;
 322
 323             /* load j atom coordinates */
 324             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 325                                               x+j_coord_offsetC,x+j_coord_offsetD,
 326                                               &jx0,&jy0,&jz0);
 327
 328             /* Calculate displacement vector */
 329             dx10             = _mm_sub_ps(ix1,jx0);
 330             dy10             = _mm_sub_ps(iy1,jy0);
 331             dz10             = _mm_sub_ps(iz1,jz0);
 332             dx20             = _mm_sub_ps(ix2,jx0);
 333             dy20             = _mm_sub_ps(iy2,jy0);
 334             dz20             = _mm_sub_ps(iz2,jz0);
 335             dx30             = _mm_sub_ps(ix3,jx0);
 336             dy30             = _mm_sub_ps(iy3,jy0);
 337             dz30             = _mm_sub_ps(iz3,jz0);
 338
 339             /* Calculate squared distance and things based on it */
 340             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 341             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 342             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 343
 344             rinv10           = avx128fma_invsqrt_f(rsq10);
 345             rinv20           = avx128fma_invsqrt_f(rsq20);
 346             rinv30           = avx128fma_invsqrt_f(rsq30);
 347
 348             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 349             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 350             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 351
 352             /* Load parameters for j particles */
 353             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 354                                                               charge+jnrC+0,charge+jnrD+0);
 355
 356             fjx0             = _mm_setzero_ps();
 357             fjy0             = _mm_setzero_ps();
 358             fjz0             = _mm_setzero_ps();
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             /* Compute parameters for interactions between i and j atoms */
 365             qq10             = _mm_mul_ps(iq1,jq0);
 366
 367             /* COULOMB ELECTROSTATICS */
 368             velec            = _mm_mul_ps(qq10,rinv10);
 369             felec            = _mm_mul_ps(velec,rinvsq10);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm_andnot_ps(dummy_mask,velec);
 373             velecsum         = _mm_add_ps(velecsum,velec);
 374
 375             fscal            = felec;
 376
 377             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 378
 379              /* Update vectorial force */
 380             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 381             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 382             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 383
 384             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 385             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 386             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* Compute parameters for interactions between i and j atoms */
 393             qq20             = _mm_mul_ps(iq2,jq0);
 394
 395             /* COULOMB ELECTROSTATICS */
 396             velec            = _mm_mul_ps(qq20,rinv20);
 397             felec            = _mm_mul_ps(velec,rinvsq20);
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velec            = _mm_andnot_ps(dummy_mask,velec);
 401             velecsum         = _mm_add_ps(velecsum,velec);
 402
 403             fscal            = felec;
 404
 405             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 406
 407              /* Update vectorial force */
 408             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 409             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 410             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 411
 412             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 413             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 414             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* Compute parameters for interactions between i and j atoms */
 421             qq30             = _mm_mul_ps(iq3,jq0);
 422
 423             /* COULOMB ELECTROSTATICS */
 424             velec            = _mm_mul_ps(qq30,rinv30);
 425             felec            = _mm_mul_ps(velec,rinvsq30);
 426
 427             /* Update potential sum for this i atom from the interaction with this j atom. */
 428             velec            = _mm_andnot_ps(dummy_mask,velec);
 429             velecsum         = _mm_add_ps(velecsum,velec);
 430
 431             fscal            = felec;
 432
 433             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 434
 435              /* Update vectorial force */
 436             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 437             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 438             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 439
 440             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 441             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 442             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 443
 444             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 445             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 446             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 447             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 448
 449             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 450
 451             /* Inner loop uses 93 flops */
 452         }
 453
 454         /* End of innermost loop */
 455
 456         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 457                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 458
 459         ggid                        = gid[iidx];
 460         /* Update potential energies */
 461         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 462
 463         /* Increment number of inner iterations */
 464         inneriter                  += j_index_end - j_index_start;
 465
 466         /* Outer loop uses 19 flops */
 467     }
 468
 469     /* Increment number of outer iterations */
 470     outeriter        += nri;
 471
 472     /* Update outer/inner flops */
 473
 474     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4_VF,outeriter*19 + inneriter*93);
 475 }
 476 /*
 477  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4P1_F_avx_128_fma_single
 478  * Electrostatics interaction: Coulomb
 479  * VdW interaction:            None
 480  * Geometry:                   Water4-Particle
 481  * Calculate force/pot:        Force
 482  */
 483 void
 484 nb_kernel_ElecCoul_VdwNone_GeomW4P1_F_avx_128_fma_single
 485                     (t_nblist                    * gmx_restrict       nlist,
 486                      rvec                        * gmx_restrict          xx,
 487                      rvec                        * gmx_restrict          ff,
 488                      struct t_forcerec           * gmx_restrict          fr,
 489                      t_mdatoms                   * gmx_restrict     mdatoms,
 490                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 491                      t_nrnb                      * gmx_restrict        nrnb)
 492 {
 493     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 494      * just 0 for non-waters.
 495      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 496      * jnr indices corresponding to data put in the four positions in the SIMD register.
 497      */
 498     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 499     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 500     int              jnrA,jnrB,jnrC,jnrD;
 501     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 502     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 503     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 504     real             rcutoff_scalar;
 505     real             *shiftvec,*fshift,*x,*f;
 506     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 507     real             scratch[4*DIM];
 508     __m128           fscal,rcutoff,rcutoff2,jidxall;
 509     int              vdwioffset1;
 510     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 511     int              vdwioffset2;
 512     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 513     int              vdwioffset3;
 514     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 515     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 516     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 517     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 518     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 519     __m128           dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 520     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 521     real             *charge;
 522     __m128           dummy_mask,cutoff_mask;
 523     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 524     __m128           one     = _mm_set1_ps(1.0);
 525     __m128           two     = _mm_set1_ps(2.0);
 526     x                = xx[0];
 527     f                = ff[0];
 528
 529     nri              = nlist->nri;
 530     iinr             = nlist->iinr;
 531     jindex           = nlist->jindex;
 532     jjnr             = nlist->jjnr;
 533     shiftidx         = nlist->shift;
 534     gid              = nlist->gid;
 535     shiftvec         = fr->shift_vec[0];
 536     fshift           = fr->fshift[0];
 537     facel            = _mm_set1_ps(fr->ic->epsfac);
 538     charge           = mdatoms->chargeA;
 539
 540     /* Setup water-specific parameters */
 541     inr              = nlist->iinr[0];
 542     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 543     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 544     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 545
 546     /* Avoid stupid compiler warnings */
 547     jnrA = jnrB = jnrC = jnrD = 0;
 548     j_coord_offsetA = 0;
 549     j_coord_offsetB = 0;
 550     j_coord_offsetC = 0;
 551     j_coord_offsetD = 0;
 552
 553     outeriter        = 0;
 554     inneriter        = 0;
 555
 556     for(iidx=0;iidx<4*DIM;iidx++)
 557     {
 558         scratch[iidx] = 0.0;
 559     }
 560
 561     /* Start outer loop over neighborlists */
 562     for(iidx=0; iidx<nri; iidx++)
 563     {
 564         /* Load shift vector for this list */
 565         i_shift_offset   = DIM*shiftidx[iidx];
 566
 567         /* Load limits for loop over neighbors */
 568         j_index_start    = jindex[iidx];
 569         j_index_end      = jindex[iidx+1];
 570
 571         /* Get outer coordinate index */
 572         inr              = iinr[iidx];
 573         i_coord_offset   = DIM*inr;
 574
 575         /* Load i particle coords and add shift vector */
 576         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 577                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 578
 579         fix1             = _mm_setzero_ps();
 580         fiy1             = _mm_setzero_ps();
 581         fiz1             = _mm_setzero_ps();
 582         fix2             = _mm_setzero_ps();
 583         fiy2             = _mm_setzero_ps();
 584         fiz2             = _mm_setzero_ps();
 585         fix3             = _mm_setzero_ps();
 586         fiy3             = _mm_setzero_ps();
 587         fiz3             = _mm_setzero_ps();
 588
 589         /* Start inner kernel loop */
 590         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 591         {
 592
 593             /* Get j neighbor index, and coordinate index */
 594             jnrA             = jjnr[jidx];
 595             jnrB             = jjnr[jidx+1];
 596             jnrC             = jjnr[jidx+2];
 597             jnrD             = jjnr[jidx+3];
 598             j_coord_offsetA  = DIM*jnrA;
 599             j_coord_offsetB  = DIM*jnrB;
 600             j_coord_offsetC  = DIM*jnrC;
 601             j_coord_offsetD  = DIM*jnrD;
 602
 603             /* load j atom coordinates */
 604             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 605                                               x+j_coord_offsetC,x+j_coord_offsetD,
 606                                               &jx0,&jy0,&jz0);
 607
 608             /* Calculate displacement vector */
 609             dx10             = _mm_sub_ps(ix1,jx0);
 610             dy10             = _mm_sub_ps(iy1,jy0);
 611             dz10             = _mm_sub_ps(iz1,jz0);
 612             dx20             = _mm_sub_ps(ix2,jx0);
 613             dy20             = _mm_sub_ps(iy2,jy0);
 614             dz20             = _mm_sub_ps(iz2,jz0);
 615             dx30             = _mm_sub_ps(ix3,jx0);
 616             dy30             = _mm_sub_ps(iy3,jy0);
 617             dz30             = _mm_sub_ps(iz3,jz0);
 618
 619             /* Calculate squared distance and things based on it */
 620             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 621             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 622             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 623
 624             rinv10           = avx128fma_invsqrt_f(rsq10);
 625             rinv20           = avx128fma_invsqrt_f(rsq20);
 626             rinv30           = avx128fma_invsqrt_f(rsq30);
 627
 628             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 629             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 630             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 631
 632             /* Load parameters for j particles */
 633             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 634                                                               charge+jnrC+0,charge+jnrD+0);
 635
 636             fjx0             = _mm_setzero_ps();
 637             fjy0             = _mm_setzero_ps();
 638             fjz0             = _mm_setzero_ps();
 639
 640             /**************************
 641              * CALCULATE INTERACTIONS *
 642              **************************/
 643
 644             /* Compute parameters for interactions between i and j atoms */
 645             qq10             = _mm_mul_ps(iq1,jq0);
 646
 647             /* COULOMB ELECTROSTATICS */
 648             velec            = _mm_mul_ps(qq10,rinv10);
 649             felec            = _mm_mul_ps(velec,rinvsq10);
 650
 651             fscal            = felec;
 652
 653              /* Update vectorial force */
 654             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 655             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 656             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 657
 658             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 659             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 660             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 661
 662             /**************************
 663              * CALCULATE INTERACTIONS *
 664              **************************/
 665
 666             /* Compute parameters for interactions between i and j atoms */
 667             qq20             = _mm_mul_ps(iq2,jq0);
 668
 669             /* COULOMB ELECTROSTATICS */
 670             velec            = _mm_mul_ps(qq20,rinv20);
 671             felec            = _mm_mul_ps(velec,rinvsq20);
 672
 673             fscal            = felec;
 674
 675              /* Update vectorial force */
 676             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 677             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 678             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 679
 680             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 681             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 682             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 683
 684             /**************************
 685              * CALCULATE INTERACTIONS *
 686              **************************/
 687
 688             /* Compute parameters for interactions between i and j atoms */
 689             qq30             = _mm_mul_ps(iq3,jq0);
 690
 691             /* COULOMB ELECTROSTATICS */
 692             velec            = _mm_mul_ps(qq30,rinv30);
 693             felec            = _mm_mul_ps(velec,rinvsq30);
 694
 695             fscal            = felec;
 696
 697              /* Update vectorial force */
 698             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 699             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 700             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 701
 702             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 703             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 704             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 705
 706             fjptrA             = f+j_coord_offsetA;
 707             fjptrB             = f+j_coord_offsetB;
 708             fjptrC             = f+j_coord_offsetC;
 709             fjptrD             = f+j_coord_offsetD;
 710
 711             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 712
 713             /* Inner loop uses 90 flops */
 714         }
 715
 716         if(jidx<j_index_end)
 717         {
 718
 719             /* Get j neighbor index, and coordinate index */
 720             jnrlistA         = jjnr[jidx];
 721             jnrlistB         = jjnr[jidx+1];
 722             jnrlistC         = jjnr[jidx+2];
 723             jnrlistD         = jjnr[jidx+3];
 724             /* Sign of each element will be negative for non-real atoms.
 725              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 726              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 727              */
 728             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 729             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 730             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 731             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 732             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 733             j_coord_offsetA  = DIM*jnrA;
 734             j_coord_offsetB  = DIM*jnrB;
 735             j_coord_offsetC  = DIM*jnrC;
 736             j_coord_offsetD  = DIM*jnrD;
 737
 738             /* load j atom coordinates */
 739             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 740                                               x+j_coord_offsetC,x+j_coord_offsetD,
 741                                               &jx0,&jy0,&jz0);
 742
 743             /* Calculate displacement vector */
 744             dx10             = _mm_sub_ps(ix1,jx0);
 745             dy10             = _mm_sub_ps(iy1,jy0);
 746             dz10             = _mm_sub_ps(iz1,jz0);
 747             dx20             = _mm_sub_ps(ix2,jx0);
 748             dy20             = _mm_sub_ps(iy2,jy0);
 749             dz20             = _mm_sub_ps(iz2,jz0);
 750             dx30             = _mm_sub_ps(ix3,jx0);
 751             dy30             = _mm_sub_ps(iy3,jy0);
 752             dz30             = _mm_sub_ps(iz3,jz0);
 753
 754             /* Calculate squared distance and things based on it */
 755             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 756             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 757             rsq30            = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
 758
 759             rinv10           = avx128fma_invsqrt_f(rsq10);
 760             rinv20           = avx128fma_invsqrt_f(rsq20);
 761             rinv30           = avx128fma_invsqrt_f(rsq30);
 762
 763             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 764             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 765             rinvsq30         = _mm_mul_ps(rinv30,rinv30);
 766
 767             /* Load parameters for j particles */
 768             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 769                                                               charge+jnrC+0,charge+jnrD+0);
 770
 771             fjx0             = _mm_setzero_ps();
 772             fjy0             = _mm_setzero_ps();
 773             fjz0             = _mm_setzero_ps();
 774
 775             /**************************
 776              * CALCULATE INTERACTIONS *
 777              **************************/
 778
 779             /* Compute parameters for interactions between i and j atoms */
 780             qq10             = _mm_mul_ps(iq1,jq0);
 781
 782             /* COULOMB ELECTROSTATICS */
 783             velec            = _mm_mul_ps(qq10,rinv10);
 784             felec            = _mm_mul_ps(velec,rinvsq10);
 785
 786             fscal            = felec;
 787
 788             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 789
 790              /* Update vectorial force */
 791             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 792             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 793             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 794
 795             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 796             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 797             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 798
 799             /**************************
 800              * CALCULATE INTERACTIONS *
 801              **************************/
 802
 803             /* Compute parameters for interactions between i and j atoms */
 804             qq20             = _mm_mul_ps(iq2,jq0);
 805
 806             /* COULOMB ELECTROSTATICS */
 807             velec            = _mm_mul_ps(qq20,rinv20);
 808             felec            = _mm_mul_ps(velec,rinvsq20);
 809
 810             fscal            = felec;
 811
 812             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 813
 814              /* Update vectorial force */
 815             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 816             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 817             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 818
 819             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 820             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 821             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 822
 823             /**************************
 824              * CALCULATE INTERACTIONS *
 825              **************************/
 826
 827             /* Compute parameters for interactions between i and j atoms */
 828             qq30             = _mm_mul_ps(iq3,jq0);
 829
 830             /* COULOMB ELECTROSTATICS */
 831             velec            = _mm_mul_ps(qq30,rinv30);
 832             felec            = _mm_mul_ps(velec,rinvsq30);
 833
 834             fscal            = felec;
 835
 836             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 837
 838              /* Update vectorial force */
 839             fix3             = _mm_macc_ps(dx30,fscal,fix3);
 840             fiy3             = _mm_macc_ps(dy30,fscal,fiy3);
 841             fiz3             = _mm_macc_ps(dz30,fscal,fiz3);
 842
 843             fjx0             = _mm_macc_ps(dx30,fscal,fjx0);
 844             fjy0             = _mm_macc_ps(dy30,fscal,fjy0);
 845             fjz0             = _mm_macc_ps(dz30,fscal,fjz0);
 846
 847             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 848             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 849             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 850             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 851
 852             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 853
 854             /* Inner loop uses 90 flops */
 855         }
 856
 857         /* End of innermost loop */
 858
 859         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 860                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 861
 862         /* Increment number of inner iterations */
 863         inneriter                  += j_index_end - j_index_start;
 864
 865         /* Outer loop uses 18 flops */
 866     }
 867
 868     /* Increment number of outer iterations */
 869     outeriter        += nri;
 870
 871     /* Update outer/inner flops */
 872
 873     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4_F,outeriter*18 + inneriter*90);
 874 }