src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwNone_GeomW3P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            None
  54  * Geometry:                   Water3-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwNone_GeomW3P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  93     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  94     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  95     real             *charge;
  96     __m128           dummy_mask,cutoff_mask;
  97     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  98     __m128           one     = _mm_set1_ps(1.0);
  99     __m128           two     = _mm_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm_set1_ps(fr->epsfac);
 112     charge           = mdatoms->chargeA;
 113
 114     /* Setup water-specific parameters */
 115     inr              = nlist->iinr[0];
 116     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 117     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 118     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 119
 120     /* Avoid stupid compiler warnings */
 121     jnrA = jnrB = jnrC = jnrD = 0;
 122     j_coord_offsetA = 0;
 123     j_coord_offsetB = 0;
 124     j_coord_offsetC = 0;
 125     j_coord_offsetD = 0;
 126
 127     outeriter        = 0;
 128     inneriter        = 0;
 129
 130     for(iidx=0;iidx<4*DIM;iidx++)
 131     {
 132         scratch[iidx] = 0.0;
 133     }
 134
 135     /* Start outer loop over neighborlists */
 136     for(iidx=0; iidx<nri; iidx++)
 137     {
 138         /* Load shift vector for this list */
 139         i_shift_offset   = DIM*shiftidx[iidx];
 140
 141         /* Load limits for loop over neighbors */
 142         j_index_start    = jindex[iidx];
 143         j_index_end      = jindex[iidx+1];
 144
 145         /* Get outer coordinate index */
 146         inr              = iinr[iidx];
 147         i_coord_offset   = DIM*inr;
 148
 149         /* Load i particle coords and add shift vector */
 150         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 151                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 152
 153         fix0             = _mm_setzero_ps();
 154         fiy0             = _mm_setzero_ps();
 155         fiz0             = _mm_setzero_ps();
 156         fix1             = _mm_setzero_ps();
 157         fiy1             = _mm_setzero_ps();
 158         fiz1             = _mm_setzero_ps();
 159         fix2             = _mm_setzero_ps();
 160         fiy2             = _mm_setzero_ps();
 161         fiz2             = _mm_setzero_ps();
 162
 163         /* Reset potential sums */
 164         velecsum         = _mm_setzero_ps();
 165
 166         /* Start inner kernel loop */
 167         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 168         {
 169
 170             /* Get j neighbor index, and coordinate index */
 171             jnrA             = jjnr[jidx];
 172             jnrB             = jjnr[jidx+1];
 173             jnrC             = jjnr[jidx+2];
 174             jnrD             = jjnr[jidx+3];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177             j_coord_offsetC  = DIM*jnrC;
 178             j_coord_offsetD  = DIM*jnrD;
 179
 180             /* load j atom coordinates */
 181             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                               x+j_coord_offsetC,x+j_coord_offsetD,
 183                                               &jx0,&jy0,&jz0);
 184
 185             /* Calculate displacement vector */
 186             dx00             = _mm_sub_ps(ix0,jx0);
 187             dy00             = _mm_sub_ps(iy0,jy0);
 188             dz00             = _mm_sub_ps(iz0,jz0);
 189             dx10             = _mm_sub_ps(ix1,jx0);
 190             dy10             = _mm_sub_ps(iy1,jy0);
 191             dz10             = _mm_sub_ps(iz1,jz0);
 192             dx20             = _mm_sub_ps(ix2,jx0);
 193             dy20             = _mm_sub_ps(iy2,jy0);
 194             dz20             = _mm_sub_ps(iz2,jz0);
 195
 196             /* Calculate squared distance and things based on it */
 197             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 198             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 199             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 200
 201             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 202             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 203             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 204
 205             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 206             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 207             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 208
 209             /* Load parameters for j particles */
 210             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 211                                                               charge+jnrC+0,charge+jnrD+0);
 212
 213             fjx0             = _mm_setzero_ps();
 214             fjy0             = _mm_setzero_ps();
 215             fjz0             = _mm_setzero_ps();
 216
 217             /**************************
 218              * CALCULATE INTERACTIONS *
 219              **************************/
 220
 221             /* Compute parameters for interactions between i and j atoms */
 222             qq00             = _mm_mul_ps(iq0,jq0);
 223
 224             /* COULOMB ELECTROSTATICS */
 225             velec            = _mm_mul_ps(qq00,rinv00);
 226             felec            = _mm_mul_ps(velec,rinvsq00);
 227
 228             /* Update potential sum for this i atom from the interaction with this j atom. */
 229             velecsum         = _mm_add_ps(velecsum,velec);
 230
 231             fscal            = felec;
 232
 233              /* Update vectorial force */
 234             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 235             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 236             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 237
 238             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 239             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 240             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 241
 242             /**************************
 243              * CALCULATE INTERACTIONS *
 244              **************************/
 245
 246             /* Compute parameters for interactions between i and j atoms */
 247             qq10             = _mm_mul_ps(iq1,jq0);
 248
 249             /* COULOMB ELECTROSTATICS */
 250             velec            = _mm_mul_ps(qq10,rinv10);
 251             felec            = _mm_mul_ps(velec,rinvsq10);
 252
 253             /* Update potential sum for this i atom from the interaction with this j atom. */
 254             velecsum         = _mm_add_ps(velecsum,velec);
 255
 256             fscal            = felec;
 257
 258              /* Update vectorial force */
 259             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 260             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 261             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 262
 263             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 264             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 265             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 266
 267             /**************************
 268              * CALCULATE INTERACTIONS *
 269              **************************/
 270
 271             /* Compute parameters for interactions between i and j atoms */
 272             qq20             = _mm_mul_ps(iq2,jq0);
 273
 274             /* COULOMB ELECTROSTATICS */
 275             velec            = _mm_mul_ps(qq20,rinv20);
 276             felec            = _mm_mul_ps(velec,rinvsq20);
 277
 278             /* Update potential sum for this i atom from the interaction with this j atom. */
 279             velecsum         = _mm_add_ps(velecsum,velec);
 280
 281             fscal            = felec;
 282
 283              /* Update vectorial force */
 284             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 285             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 286             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 287
 288             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 289             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 290             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 291
 292             fjptrA             = f+j_coord_offsetA;
 293             fjptrB             = f+j_coord_offsetB;
 294             fjptrC             = f+j_coord_offsetC;
 295             fjptrD             = f+j_coord_offsetD;
 296
 297             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 298
 299             /* Inner loop uses 93 flops */
 300         }
 301
 302         if(jidx<j_index_end)
 303         {
 304
 305             /* Get j neighbor index, and coordinate index */
 306             jnrlistA         = jjnr[jidx];
 307             jnrlistB         = jjnr[jidx+1];
 308             jnrlistC         = jjnr[jidx+2];
 309             jnrlistD         = jjnr[jidx+3];
 310             /* Sign of each element will be negative for non-real atoms.
 311              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 312              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 313              */
 314             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 315             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 316             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 317             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 318             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 319             j_coord_offsetA  = DIM*jnrA;
 320             j_coord_offsetB  = DIM*jnrB;
 321             j_coord_offsetC  = DIM*jnrC;
 322             j_coord_offsetD  = DIM*jnrD;
 323
 324             /* load j atom coordinates */
 325             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 326                                               x+j_coord_offsetC,x+j_coord_offsetD,
 327                                               &jx0,&jy0,&jz0);
 328
 329             /* Calculate displacement vector */
 330             dx00             = _mm_sub_ps(ix0,jx0);
 331             dy00             = _mm_sub_ps(iy0,jy0);
 332             dz00             = _mm_sub_ps(iz0,jz0);
 333             dx10             = _mm_sub_ps(ix1,jx0);
 334             dy10             = _mm_sub_ps(iy1,jy0);
 335             dz10             = _mm_sub_ps(iz1,jz0);
 336             dx20             = _mm_sub_ps(ix2,jx0);
 337             dy20             = _mm_sub_ps(iy2,jy0);
 338             dz20             = _mm_sub_ps(iz2,jz0);
 339
 340             /* Calculate squared distance and things based on it */
 341             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 342             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 343             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 344
 345             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 346             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 347             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 348
 349             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 350             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 351             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 352
 353             /* Load parameters for j particles */
 354             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 355                                                               charge+jnrC+0,charge+jnrD+0);
 356
 357             fjx0             = _mm_setzero_ps();
 358             fjy0             = _mm_setzero_ps();
 359             fjz0             = _mm_setzero_ps();
 360
 361             /**************************
 362              * CALCULATE INTERACTIONS *
 363              **************************/
 364
 365             /* Compute parameters for interactions between i and j atoms */
 366             qq00             = _mm_mul_ps(iq0,jq0);
 367
 368             /* COULOMB ELECTROSTATICS */
 369             velec            = _mm_mul_ps(qq00,rinv00);
 370             felec            = _mm_mul_ps(velec,rinvsq00);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velec            = _mm_andnot_ps(dummy_mask,velec);
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 379
 380              /* Update vectorial force */
 381             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 382             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 383             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 384
 385             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 386             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 387             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 388
 389             /**************************
 390              * CALCULATE INTERACTIONS *
 391              **************************/
 392
 393             /* Compute parameters for interactions between i and j atoms */
 394             qq10             = _mm_mul_ps(iq1,jq0);
 395
 396             /* COULOMB ELECTROSTATICS */
 397             velec            = _mm_mul_ps(qq10,rinv10);
 398             felec            = _mm_mul_ps(velec,rinvsq10);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velec            = _mm_andnot_ps(dummy_mask,velec);
 402             velecsum         = _mm_add_ps(velecsum,velec);
 403
 404             fscal            = felec;
 405
 406             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 407
 408              /* Update vectorial force */
 409             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 410             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 411             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 412
 413             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 414             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 415             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             /* Compute parameters for interactions between i and j atoms */
 422             qq20             = _mm_mul_ps(iq2,jq0);
 423
 424             /* COULOMB ELECTROSTATICS */
 425             velec            = _mm_mul_ps(qq20,rinv20);
 426             felec            = _mm_mul_ps(velec,rinvsq20);
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velec            = _mm_andnot_ps(dummy_mask,velec);
 430             velecsum         = _mm_add_ps(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 435
 436              /* Update vectorial force */
 437             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 438             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 439             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 440
 441             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 442             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 443             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 444
 445             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 446             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 447             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 448             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 449
 450             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 451
 452             /* Inner loop uses 93 flops */
 453         }
 454
 455         /* End of innermost loop */
 456
 457         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 458                                               f+i_coord_offset,fshift+i_shift_offset);
 459
 460         ggid                        = gid[iidx];
 461         /* Update potential energies */
 462         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 463
 464         /* Increment number of inner iterations */
 465         inneriter                  += j_index_end - j_index_start;
 466
 467         /* Outer loop uses 19 flops */
 468     }
 469
 470     /* Increment number of outer iterations */
 471     outeriter        += nri;
 472
 473     /* Update outer/inner flops */
 474
 475     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*19 + inneriter*93);
 476 }
 477 /*
 478  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3P1_F_avx_128_fma_single
 479  * Electrostatics interaction: Coulomb
 480  * VdW interaction:            None
 481  * Geometry:                   Water3-Particle
 482  * Calculate force/pot:        Force
 483  */
 484 void
 485 nb_kernel_ElecCoul_VdwNone_GeomW3P1_F_avx_128_fma_single
 486                     (t_nblist                    * gmx_restrict       nlist,
 487                      rvec                        * gmx_restrict          xx,
 488                      rvec                        * gmx_restrict          ff,
 489                      t_forcerec                  * gmx_restrict          fr,
 490                      t_mdatoms                   * gmx_restrict     mdatoms,
 491                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 492                      t_nrnb                      * gmx_restrict        nrnb)
 493 {
 494     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 495      * just 0 for non-waters.
 496      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 497      * jnr indices corresponding to data put in the four positions in the SIMD register.
 498      */
 499     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 500     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 501     int              jnrA,jnrB,jnrC,jnrD;
 502     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 503     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 504     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 505     real             rcutoff_scalar;
 506     real             *shiftvec,*fshift,*x,*f;
 507     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 508     real             scratch[4*DIM];
 509     __m128           fscal,rcutoff,rcutoff2,jidxall;
 510     int              vdwioffset0;
 511     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 512     int              vdwioffset1;
 513     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 514     int              vdwioffset2;
 515     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 516     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 517     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 518     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 519     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 520     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 521     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 522     real             *charge;
 523     __m128           dummy_mask,cutoff_mask;
 524     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 525     __m128           one     = _mm_set1_ps(1.0);
 526     __m128           two     = _mm_set1_ps(2.0);
 527     x                = xx[0];
 528     f                = ff[0];
 529
 530     nri              = nlist->nri;
 531     iinr             = nlist->iinr;
 532     jindex           = nlist->jindex;
 533     jjnr             = nlist->jjnr;
 534     shiftidx         = nlist->shift;
 535     gid              = nlist->gid;
 536     shiftvec         = fr->shift_vec[0];
 537     fshift           = fr->fshift[0];
 538     facel            = _mm_set1_ps(fr->epsfac);
 539     charge           = mdatoms->chargeA;
 540
 541     /* Setup water-specific parameters */
 542     inr              = nlist->iinr[0];
 543     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 544     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 545     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 546
 547     /* Avoid stupid compiler warnings */
 548     jnrA = jnrB = jnrC = jnrD = 0;
 549     j_coord_offsetA = 0;
 550     j_coord_offsetB = 0;
 551     j_coord_offsetC = 0;
 552     j_coord_offsetD = 0;
 553
 554     outeriter        = 0;
 555     inneriter        = 0;
 556
 557     for(iidx=0;iidx<4*DIM;iidx++)
 558     {
 559         scratch[iidx] = 0.0;
 560     }
 561
 562     /* Start outer loop over neighborlists */
 563     for(iidx=0; iidx<nri; iidx++)
 564     {
 565         /* Load shift vector for this list */
 566         i_shift_offset   = DIM*shiftidx[iidx];
 567
 568         /* Load limits for loop over neighbors */
 569         j_index_start    = jindex[iidx];
 570         j_index_end      = jindex[iidx+1];
 571
 572         /* Get outer coordinate index */
 573         inr              = iinr[iidx];
 574         i_coord_offset   = DIM*inr;
 575
 576         /* Load i particle coords and add shift vector */
 577         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 578                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 579
 580         fix0             = _mm_setzero_ps();
 581         fiy0             = _mm_setzero_ps();
 582         fiz0             = _mm_setzero_ps();
 583         fix1             = _mm_setzero_ps();
 584         fiy1             = _mm_setzero_ps();
 585         fiz1             = _mm_setzero_ps();
 586         fix2             = _mm_setzero_ps();
 587         fiy2             = _mm_setzero_ps();
 588         fiz2             = _mm_setzero_ps();
 589
 590         /* Start inner kernel loop */
 591         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 592         {
 593
 594             /* Get j neighbor index, and coordinate index */
 595             jnrA             = jjnr[jidx];
 596             jnrB             = jjnr[jidx+1];
 597             jnrC             = jjnr[jidx+2];
 598             jnrD             = jjnr[jidx+3];
 599             j_coord_offsetA  = DIM*jnrA;
 600             j_coord_offsetB  = DIM*jnrB;
 601             j_coord_offsetC  = DIM*jnrC;
 602             j_coord_offsetD  = DIM*jnrD;
 603
 604             /* load j atom coordinates */
 605             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 606                                               x+j_coord_offsetC,x+j_coord_offsetD,
 607                                               &jx0,&jy0,&jz0);
 608
 609             /* Calculate displacement vector */
 610             dx00             = _mm_sub_ps(ix0,jx0);
 611             dy00             = _mm_sub_ps(iy0,jy0);
 612             dz00             = _mm_sub_ps(iz0,jz0);
 613             dx10             = _mm_sub_ps(ix1,jx0);
 614             dy10             = _mm_sub_ps(iy1,jy0);
 615             dz10             = _mm_sub_ps(iz1,jz0);
 616             dx20             = _mm_sub_ps(ix2,jx0);
 617             dy20             = _mm_sub_ps(iy2,jy0);
 618             dz20             = _mm_sub_ps(iz2,jz0);
 619
 620             /* Calculate squared distance and things based on it */
 621             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 622             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 623             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 624
 625             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 626             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 627             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 628
 629             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 630             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 631             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 632
 633             /* Load parameters for j particles */
 634             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 635                                                               charge+jnrC+0,charge+jnrD+0);
 636
 637             fjx0             = _mm_setzero_ps();
 638             fjy0             = _mm_setzero_ps();
 639             fjz0             = _mm_setzero_ps();
 640
 641             /**************************
 642              * CALCULATE INTERACTIONS *
 643              **************************/
 644
 645             /* Compute parameters for interactions between i and j atoms */
 646             qq00             = _mm_mul_ps(iq0,jq0);
 647
 648             /* COULOMB ELECTROSTATICS */
 649             velec            = _mm_mul_ps(qq00,rinv00);
 650             felec            = _mm_mul_ps(velec,rinvsq00);
 651
 652             fscal            = felec;
 653
 654              /* Update vectorial force */
 655             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 656             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 657             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 658
 659             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 660             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 661             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 662
 663             /**************************
 664              * CALCULATE INTERACTIONS *
 665              **************************/
 666
 667             /* Compute parameters for interactions between i and j atoms */
 668             qq10             = _mm_mul_ps(iq1,jq0);
 669
 670             /* COULOMB ELECTROSTATICS */
 671             velec            = _mm_mul_ps(qq10,rinv10);
 672             felec            = _mm_mul_ps(velec,rinvsq10);
 673
 674             fscal            = felec;
 675
 676              /* Update vectorial force */
 677             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 678             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 679             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 680
 681             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 682             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 683             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             /* Compute parameters for interactions between i and j atoms */
 690             qq20             = _mm_mul_ps(iq2,jq0);
 691
 692             /* COULOMB ELECTROSTATICS */
 693             velec            = _mm_mul_ps(qq20,rinv20);
 694             felec            = _mm_mul_ps(velec,rinvsq20);
 695
 696             fscal            = felec;
 697
 698              /* Update vectorial force */
 699             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 700             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 701             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 702
 703             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 704             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 705             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 706
 707             fjptrA             = f+j_coord_offsetA;
 708             fjptrB             = f+j_coord_offsetB;
 709             fjptrC             = f+j_coord_offsetC;
 710             fjptrD             = f+j_coord_offsetD;
 711
 712             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 713
 714             /* Inner loop uses 90 flops */
 715         }
 716
 717         if(jidx<j_index_end)
 718         {
 719
 720             /* Get j neighbor index, and coordinate index */
 721             jnrlistA         = jjnr[jidx];
 722             jnrlistB         = jjnr[jidx+1];
 723             jnrlistC         = jjnr[jidx+2];
 724             jnrlistD         = jjnr[jidx+3];
 725             /* Sign of each element will be negative for non-real atoms.
 726              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 727              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 728              */
 729             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 730             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 731             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 732             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 733             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 734             j_coord_offsetA  = DIM*jnrA;
 735             j_coord_offsetB  = DIM*jnrB;
 736             j_coord_offsetC  = DIM*jnrC;
 737             j_coord_offsetD  = DIM*jnrD;
 738
 739             /* load j atom coordinates */
 740             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 741                                               x+j_coord_offsetC,x+j_coord_offsetD,
 742                                               &jx0,&jy0,&jz0);
 743
 744             /* Calculate displacement vector */
 745             dx00             = _mm_sub_ps(ix0,jx0);
 746             dy00             = _mm_sub_ps(iy0,jy0);
 747             dz00             = _mm_sub_ps(iz0,jz0);
 748             dx10             = _mm_sub_ps(ix1,jx0);
 749             dy10             = _mm_sub_ps(iy1,jy0);
 750             dz10             = _mm_sub_ps(iz1,jz0);
 751             dx20             = _mm_sub_ps(ix2,jx0);
 752             dy20             = _mm_sub_ps(iy2,jy0);
 753             dz20             = _mm_sub_ps(iz2,jz0);
 754
 755             /* Calculate squared distance and things based on it */
 756             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 757             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 758             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 759
 760             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 761             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 762             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 763
 764             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 765             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 766             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 767
 768             /* Load parameters for j particles */
 769             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 770                                                               charge+jnrC+0,charge+jnrD+0);
 771
 772             fjx0             = _mm_setzero_ps();
 773             fjy0             = _mm_setzero_ps();
 774             fjz0             = _mm_setzero_ps();
 775
 776             /**************************
 777              * CALCULATE INTERACTIONS *
 778              **************************/
 779
 780             /* Compute parameters for interactions between i and j atoms */
 781             qq00             = _mm_mul_ps(iq0,jq0);
 782
 783             /* COULOMB ELECTROSTATICS */
 784             velec            = _mm_mul_ps(qq00,rinv00);
 785             felec            = _mm_mul_ps(velec,rinvsq00);
 786
 787             fscal            = felec;
 788
 789             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 790
 791              /* Update vectorial force */
 792             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 793             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 794             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 795
 796             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 797             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 798             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             /* Compute parameters for interactions between i and j atoms */
 805             qq10             = _mm_mul_ps(iq1,jq0);
 806
 807             /* COULOMB ELECTROSTATICS */
 808             velec            = _mm_mul_ps(qq10,rinv10);
 809             felec            = _mm_mul_ps(velec,rinvsq10);
 810
 811             fscal            = felec;
 812
 813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 814
 815              /* Update vectorial force */
 816             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 817             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 818             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 819
 820             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 821             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 822             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             /* Compute parameters for interactions between i and j atoms */
 829             qq20             = _mm_mul_ps(iq2,jq0);
 830
 831             /* COULOMB ELECTROSTATICS */
 832             velec            = _mm_mul_ps(qq20,rinv20);
 833             felec            = _mm_mul_ps(velec,rinvsq20);
 834
 835             fscal            = felec;
 836
 837             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 838
 839              /* Update vectorial force */
 840             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 841             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 842             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 843
 844             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 845             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 846             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 847
 848             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 849             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 850             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 851             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 852
 853             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 854
 855             /* Inner loop uses 90 flops */
 856         }
 857
 858         /* End of innermost loop */
 859
 860         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 861                                               f+i_coord_offset,fshift+i_shift_offset);
 862
 863         /* Increment number of inner iterations */
 864         inneriter                  += j_index_end - j_index_start;
 865
 866         /* Outer loop uses 18 flops */
 867     }
 868
 869     /* Increment number of outer iterations */
 870     outeriter        += nri;
 871
 872     /* Update outer/inner flops */
 873
 874     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*18 + inneriter*90);
 875 }