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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            None
  56  * Geometry:                   Water3-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwNone_GeomW3P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  94     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  96     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  97     real             *charge;
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     /* Setup water-specific parameters */
 117     inr              = nlist->iinr[0];
 118     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 119     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 120     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 153                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 154
 155         fix0             = _mm_setzero_ps();
 156         fiy0             = _mm_setzero_ps();
 157         fiz0             = _mm_setzero_ps();
 158         fix1             = _mm_setzero_ps();
 159         fiy1             = _mm_setzero_ps();
 160         fiz1             = _mm_setzero_ps();
 161         fix2             = _mm_setzero_ps();
 162         fiy2             = _mm_setzero_ps();
 163         fiz2             = _mm_setzero_ps();
 164
 165         /* Reset potential sums */
 166         velecsum         = _mm_setzero_ps();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179             j_coord_offsetC  = DIM*jnrC;
 180             j_coord_offsetD  = DIM*jnrD;
 181
 182             /* load j atom coordinates */
 183             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 184                                               x+j_coord_offsetC,x+j_coord_offsetD,
 185                                               &jx0,&jy0,&jz0);
 186
 187             /* Calculate displacement vector */
 188             dx00             = _mm_sub_ps(ix0,jx0);
 189             dy00             = _mm_sub_ps(iy0,jy0);
 190             dz00             = _mm_sub_ps(iz0,jz0);
 191             dx10             = _mm_sub_ps(ix1,jx0);
 192             dy10             = _mm_sub_ps(iy1,jy0);
 193             dz10             = _mm_sub_ps(iz1,jz0);
 194             dx20             = _mm_sub_ps(ix2,jx0);
 195             dy20             = _mm_sub_ps(iy2,jy0);
 196             dz20             = _mm_sub_ps(iz2,jz0);
 197
 198             /* Calculate squared distance and things based on it */
 199             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 200             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 201             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 202
 203             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 204             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 205             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 206
 207             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 208             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 209             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 213                                                               charge+jnrC+0,charge+jnrD+0);
 214
 215             fjx0             = _mm_setzero_ps();
 216             fjy0             = _mm_setzero_ps();
 217             fjz0             = _mm_setzero_ps();
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             /* Compute parameters for interactions between i and j atoms */
 224             qq00             = _mm_mul_ps(iq0,jq0);
 225
 226             /* COULOMB ELECTROSTATICS */
 227             velec            = _mm_mul_ps(qq00,rinv00);
 228             felec            = _mm_mul_ps(velec,rinvsq00);
 229
 230             /* Update potential sum for this i atom from the interaction with this j atom. */
 231             velecsum         = _mm_add_ps(velecsum,velec);
 232
 233             fscal            = felec;
 234
 235              /* Update vectorial force */
 236             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 237             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 238             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 239
 240             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 241             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 242             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 243
 244             /**************************
 245              * CALCULATE INTERACTIONS *
 246              **************************/
 247
 248             /* Compute parameters for interactions between i and j atoms */
 249             qq10             = _mm_mul_ps(iq1,jq0);
 250
 251             /* COULOMB ELECTROSTATICS */
 252             velec            = _mm_mul_ps(qq10,rinv10);
 253             felec            = _mm_mul_ps(velec,rinvsq10);
 254
 255             /* Update potential sum for this i atom from the interaction with this j atom. */
 256             velecsum         = _mm_add_ps(velecsum,velec);
 257
 258             fscal            = felec;
 259
 260              /* Update vectorial force */
 261             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 262             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 263             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 264
 265             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 266             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 267             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 268
 269             /**************************
 270              * CALCULATE INTERACTIONS *
 271              **************************/
 272
 273             /* Compute parameters for interactions between i and j atoms */
 274             qq20             = _mm_mul_ps(iq2,jq0);
 275
 276             /* COULOMB ELECTROSTATICS */
 277             velec            = _mm_mul_ps(qq20,rinv20);
 278             felec            = _mm_mul_ps(velec,rinvsq20);
 279
 280             /* Update potential sum for this i atom from the interaction with this j atom. */
 281             velecsum         = _mm_add_ps(velecsum,velec);
 282
 283             fscal            = felec;
 284
 285              /* Update vectorial force */
 286             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 287             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 288             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 289
 290             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 291             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 292             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 293
 294             fjptrA             = f+j_coord_offsetA;
 295             fjptrB             = f+j_coord_offsetB;
 296             fjptrC             = f+j_coord_offsetC;
 297             fjptrD             = f+j_coord_offsetD;
 298
 299             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 300
 301             /* Inner loop uses 93 flops */
 302         }
 303
 304         if(jidx<j_index_end)
 305         {
 306
 307             /* Get j neighbor index, and coordinate index */
 308             jnrlistA         = jjnr[jidx];
 309             jnrlistB         = jjnr[jidx+1];
 310             jnrlistC         = jjnr[jidx+2];
 311             jnrlistD         = jjnr[jidx+3];
 312             /* Sign of each element will be negative for non-real atoms.
 313              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 314              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 315              */
 316             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 317             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 318             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 319             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 320             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 321             j_coord_offsetA  = DIM*jnrA;
 322             j_coord_offsetB  = DIM*jnrB;
 323             j_coord_offsetC  = DIM*jnrC;
 324             j_coord_offsetD  = DIM*jnrD;
 325
 326             /* load j atom coordinates */
 327             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 328                                               x+j_coord_offsetC,x+j_coord_offsetD,
 329                                               &jx0,&jy0,&jz0);
 330
 331             /* Calculate displacement vector */
 332             dx00             = _mm_sub_ps(ix0,jx0);
 333             dy00             = _mm_sub_ps(iy0,jy0);
 334             dz00             = _mm_sub_ps(iz0,jz0);
 335             dx10             = _mm_sub_ps(ix1,jx0);
 336             dy10             = _mm_sub_ps(iy1,jy0);
 337             dz10             = _mm_sub_ps(iz1,jz0);
 338             dx20             = _mm_sub_ps(ix2,jx0);
 339             dy20             = _mm_sub_ps(iy2,jy0);
 340             dz20             = _mm_sub_ps(iz2,jz0);
 341
 342             /* Calculate squared distance and things based on it */
 343             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 344             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 345             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 346
 347             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 348             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 349             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 350
 351             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 352             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 353             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 354
 355             /* Load parameters for j particles */
 356             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 357                                                               charge+jnrC+0,charge+jnrD+0);
 358
 359             fjx0             = _mm_setzero_ps();
 360             fjy0             = _mm_setzero_ps();
 361             fjz0             = _mm_setzero_ps();
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             /* Compute parameters for interactions between i and j atoms */
 368             qq00             = _mm_mul_ps(iq0,jq0);
 369
 370             /* COULOMB ELECTROSTATICS */
 371             velec            = _mm_mul_ps(qq00,rinv00);
 372             felec            = _mm_mul_ps(velec,rinvsq00);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velec            = _mm_andnot_ps(dummy_mask,velec);
 376             velecsum         = _mm_add_ps(velecsum,velec);
 377
 378             fscal            = felec;
 379
 380             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 381
 382              /* Update vectorial force */
 383             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 384             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 385             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 386
 387             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 388             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 389             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* Compute parameters for interactions between i and j atoms */
 396             qq10             = _mm_mul_ps(iq1,jq0);
 397
 398             /* COULOMB ELECTROSTATICS */
 399             velec            = _mm_mul_ps(qq10,rinv10);
 400             felec            = _mm_mul_ps(velec,rinvsq10);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velec            = _mm_andnot_ps(dummy_mask,velec);
 404             velecsum         = _mm_add_ps(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 409
 410              /* Update vectorial force */
 411             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 412             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 413             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 414
 415             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 416             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 417             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             /* Compute parameters for interactions between i and j atoms */
 424             qq20             = _mm_mul_ps(iq2,jq0);
 425
 426             /* COULOMB ELECTROSTATICS */
 427             velec            = _mm_mul_ps(qq20,rinv20);
 428             felec            = _mm_mul_ps(velec,rinvsq20);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velec            = _mm_andnot_ps(dummy_mask,velec);
 432             velecsum         = _mm_add_ps(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 437
 438              /* Update vectorial force */
 439             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 440             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 441             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 442
 443             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 444             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 445             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 446
 447             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 448             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 449             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 450             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 451
 452             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 453
 454             /* Inner loop uses 93 flops */
 455         }
 456
 457         /* End of innermost loop */
 458
 459         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 460                                               f+i_coord_offset,fshift+i_shift_offset);
 461
 462         ggid                        = gid[iidx];
 463         /* Update potential energies */
 464         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 465
 466         /* Increment number of inner iterations */
 467         inneriter                  += j_index_end - j_index_start;
 468
 469         /* Outer loop uses 19 flops */
 470     }
 471
 472     /* Increment number of outer iterations */
 473     outeriter        += nri;
 474
 475     /* Update outer/inner flops */
 476
 477     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*19 + inneriter*93);
 478 }
 479 /*
 480  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3P1_F_avx_128_fma_single
 481  * Electrostatics interaction: Coulomb
 482  * VdW interaction:            None
 483  * Geometry:                   Water3-Particle
 484  * Calculate force/pot:        Force
 485  */
 486 void
 487 nb_kernel_ElecCoul_VdwNone_GeomW3P1_F_avx_128_fma_single
 488                     (t_nblist                    * gmx_restrict       nlist,
 489                      rvec                        * gmx_restrict          xx,
 490                      rvec                        * gmx_restrict          ff,
 491                      t_forcerec                  * gmx_restrict          fr,
 492                      t_mdatoms                   * gmx_restrict     mdatoms,
 493                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 494                      t_nrnb                      * gmx_restrict        nrnb)
 495 {
 496     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 497      * just 0 for non-waters.
 498      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 499      * jnr indices corresponding to data put in the four positions in the SIMD register.
 500      */
 501     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 502     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 503     int              jnrA,jnrB,jnrC,jnrD;
 504     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 505     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 506     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 507     real             rcutoff_scalar;
 508     real             *shiftvec,*fshift,*x,*f;
 509     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 510     real             scratch[4*DIM];
 511     __m128           fscal,rcutoff,rcutoff2,jidxall;
 512     int              vdwioffset0;
 513     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 514     int              vdwioffset1;
 515     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 516     int              vdwioffset2;
 517     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 518     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 519     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 520     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 521     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 522     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 523     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 524     real             *charge;
 525     __m128           dummy_mask,cutoff_mask;
 526     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 527     __m128           one     = _mm_set1_ps(1.0);
 528     __m128           two     = _mm_set1_ps(2.0);
 529     x                = xx[0];
 530     f                = ff[0];
 531
 532     nri              = nlist->nri;
 533     iinr             = nlist->iinr;
 534     jindex           = nlist->jindex;
 535     jjnr             = nlist->jjnr;
 536     shiftidx         = nlist->shift;
 537     gid              = nlist->gid;
 538     shiftvec         = fr->shift_vec[0];
 539     fshift           = fr->fshift[0];
 540     facel            = _mm_set1_ps(fr->epsfac);
 541     charge           = mdatoms->chargeA;
 542
 543     /* Setup water-specific parameters */
 544     inr              = nlist->iinr[0];
 545     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 546     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 547     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 548
 549     /* Avoid stupid compiler warnings */
 550     jnrA = jnrB = jnrC = jnrD = 0;
 551     j_coord_offsetA = 0;
 552     j_coord_offsetB = 0;
 553     j_coord_offsetC = 0;
 554     j_coord_offsetD = 0;
 555
 556     outeriter        = 0;
 557     inneriter        = 0;
 558
 559     for(iidx=0;iidx<4*DIM;iidx++)
 560     {
 561         scratch[iidx] = 0.0;
 562     }
 563
 564     /* Start outer loop over neighborlists */
 565     for(iidx=0; iidx<nri; iidx++)
 566     {
 567         /* Load shift vector for this list */
 568         i_shift_offset   = DIM*shiftidx[iidx];
 569
 570         /* Load limits for loop over neighbors */
 571         j_index_start    = jindex[iidx];
 572         j_index_end      = jindex[iidx+1];
 573
 574         /* Get outer coordinate index */
 575         inr              = iinr[iidx];
 576         i_coord_offset   = DIM*inr;
 577
 578         /* Load i particle coords and add shift vector */
 579         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 580                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 581
 582         fix0             = _mm_setzero_ps();
 583         fiy0             = _mm_setzero_ps();
 584         fiz0             = _mm_setzero_ps();
 585         fix1             = _mm_setzero_ps();
 586         fiy1             = _mm_setzero_ps();
 587         fiz1             = _mm_setzero_ps();
 588         fix2             = _mm_setzero_ps();
 589         fiy2             = _mm_setzero_ps();
 590         fiz2             = _mm_setzero_ps();
 591
 592         /* Start inner kernel loop */
 593         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 594         {
 595
 596             /* Get j neighbor index, and coordinate index */
 597             jnrA             = jjnr[jidx];
 598             jnrB             = jjnr[jidx+1];
 599             jnrC             = jjnr[jidx+2];
 600             jnrD             = jjnr[jidx+3];
 601             j_coord_offsetA  = DIM*jnrA;
 602             j_coord_offsetB  = DIM*jnrB;
 603             j_coord_offsetC  = DIM*jnrC;
 604             j_coord_offsetD  = DIM*jnrD;
 605
 606             /* load j atom coordinates */
 607             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 608                                               x+j_coord_offsetC,x+j_coord_offsetD,
 609                                               &jx0,&jy0,&jz0);
 610
 611             /* Calculate displacement vector */
 612             dx00             = _mm_sub_ps(ix0,jx0);
 613             dy00             = _mm_sub_ps(iy0,jy0);
 614             dz00             = _mm_sub_ps(iz0,jz0);
 615             dx10             = _mm_sub_ps(ix1,jx0);
 616             dy10             = _mm_sub_ps(iy1,jy0);
 617             dz10             = _mm_sub_ps(iz1,jz0);
 618             dx20             = _mm_sub_ps(ix2,jx0);
 619             dy20             = _mm_sub_ps(iy2,jy0);
 620             dz20             = _mm_sub_ps(iz2,jz0);
 621
 622             /* Calculate squared distance and things based on it */
 623             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 624             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 625             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 626
 627             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 628             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 629             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 630
 631             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 632             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 633             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 634
 635             /* Load parameters for j particles */
 636             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 637                                                               charge+jnrC+0,charge+jnrD+0);
 638
 639             fjx0             = _mm_setzero_ps();
 640             fjy0             = _mm_setzero_ps();
 641             fjz0             = _mm_setzero_ps();
 642
 643             /**************************
 644              * CALCULATE INTERACTIONS *
 645              **************************/
 646
 647             /* Compute parameters for interactions between i and j atoms */
 648             qq00             = _mm_mul_ps(iq0,jq0);
 649
 650             /* COULOMB ELECTROSTATICS */
 651             velec            = _mm_mul_ps(qq00,rinv00);
 652             felec            = _mm_mul_ps(velec,rinvsq00);
 653
 654             fscal            = felec;
 655
 656              /* Update vectorial force */
 657             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 658             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 659             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 660
 661             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 662             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 663             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 664
 665             /**************************
 666              * CALCULATE INTERACTIONS *
 667              **************************/
 668
 669             /* Compute parameters for interactions between i and j atoms */
 670             qq10             = _mm_mul_ps(iq1,jq0);
 671
 672             /* COULOMB ELECTROSTATICS */
 673             velec            = _mm_mul_ps(qq10,rinv10);
 674             felec            = _mm_mul_ps(velec,rinvsq10);
 675
 676             fscal            = felec;
 677
 678              /* Update vectorial force */
 679             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 680             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 681             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 682
 683             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 684             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 685             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 686
 687             /**************************
 688              * CALCULATE INTERACTIONS *
 689              **************************/
 690
 691             /* Compute parameters for interactions between i and j atoms */
 692             qq20             = _mm_mul_ps(iq2,jq0);
 693
 694             /* COULOMB ELECTROSTATICS */
 695             velec            = _mm_mul_ps(qq20,rinv20);
 696             felec            = _mm_mul_ps(velec,rinvsq20);
 697
 698             fscal            = felec;
 699
 700              /* Update vectorial force */
 701             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 702             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 703             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 704
 705             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 706             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 707             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 708
 709             fjptrA             = f+j_coord_offsetA;
 710             fjptrB             = f+j_coord_offsetB;
 711             fjptrC             = f+j_coord_offsetC;
 712             fjptrD             = f+j_coord_offsetD;
 713
 714             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 715
 716             /* Inner loop uses 90 flops */
 717         }
 718
 719         if(jidx<j_index_end)
 720         {
 721
 722             /* Get j neighbor index, and coordinate index */
 723             jnrlistA         = jjnr[jidx];
 724             jnrlistB         = jjnr[jidx+1];
 725             jnrlistC         = jjnr[jidx+2];
 726             jnrlistD         = jjnr[jidx+3];
 727             /* Sign of each element will be negative for non-real atoms.
 728              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 729              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 730              */
 731             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 732             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 733             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 734             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 735             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 736             j_coord_offsetA  = DIM*jnrA;
 737             j_coord_offsetB  = DIM*jnrB;
 738             j_coord_offsetC  = DIM*jnrC;
 739             j_coord_offsetD  = DIM*jnrD;
 740
 741             /* load j atom coordinates */
 742             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 743                                               x+j_coord_offsetC,x+j_coord_offsetD,
 744                                               &jx0,&jy0,&jz0);
 745
 746             /* Calculate displacement vector */
 747             dx00             = _mm_sub_ps(ix0,jx0);
 748             dy00             = _mm_sub_ps(iy0,jy0);
 749             dz00             = _mm_sub_ps(iz0,jz0);
 750             dx10             = _mm_sub_ps(ix1,jx0);
 751             dy10             = _mm_sub_ps(iy1,jy0);
 752             dz10             = _mm_sub_ps(iz1,jz0);
 753             dx20             = _mm_sub_ps(ix2,jx0);
 754             dy20             = _mm_sub_ps(iy2,jy0);
 755             dz20             = _mm_sub_ps(iz2,jz0);
 756
 757             /* Calculate squared distance and things based on it */
 758             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 759             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 760             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 761
 762             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 763             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 764             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 765
 766             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 767             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 768             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 769
 770             /* Load parameters for j particles */
 771             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 772                                                               charge+jnrC+0,charge+jnrD+0);
 773
 774             fjx0             = _mm_setzero_ps();
 775             fjy0             = _mm_setzero_ps();
 776             fjz0             = _mm_setzero_ps();
 777
 778             /**************************
 779              * CALCULATE INTERACTIONS *
 780              **************************/
 781
 782             /* Compute parameters for interactions between i and j atoms */
 783             qq00             = _mm_mul_ps(iq0,jq0);
 784
 785             /* COULOMB ELECTROSTATICS */
 786             velec            = _mm_mul_ps(qq00,rinv00);
 787             felec            = _mm_mul_ps(velec,rinvsq00);
 788
 789             fscal            = felec;
 790
 791             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 792
 793              /* Update vectorial force */
 794             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 795             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 796             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 797
 798             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 799             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 800             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             /* Compute parameters for interactions between i and j atoms */
 807             qq10             = _mm_mul_ps(iq1,jq0);
 808
 809             /* COULOMB ELECTROSTATICS */
 810             velec            = _mm_mul_ps(qq10,rinv10);
 811             felec            = _mm_mul_ps(velec,rinvsq10);
 812
 813             fscal            = felec;
 814
 815             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 816
 817              /* Update vectorial force */
 818             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 819             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 820             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 821
 822             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 823             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 824             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 825
 826             /**************************
 827              * CALCULATE INTERACTIONS *
 828              **************************/
 829
 830             /* Compute parameters for interactions between i and j atoms */
 831             qq20             = _mm_mul_ps(iq2,jq0);
 832
 833             /* COULOMB ELECTROSTATICS */
 834             velec            = _mm_mul_ps(qq20,rinv20);
 835             felec            = _mm_mul_ps(velec,rinvsq20);
 836
 837             fscal            = felec;
 838
 839             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 840
 841              /* Update vectorial force */
 842             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 843             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 844             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 845
 846             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 847             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 848             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 849
 850             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 851             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 852             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 853             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 854
 855             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
 856
 857             /* Inner loop uses 90 flops */
 858         }
 859
 860         /* End of innermost loop */
 861
 862         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 863                                               f+i_coord_offset,fshift+i_shift_offset);
 864
 865         /* Increment number of inner iterations */
 866         inneriter                  += j_index_end - j_index_start;
 867
 868         /* Outer loop uses 18 flops */
 869     }
 870
 871     /* Increment number of outer iterations */
 872     outeriter        += nri;
 873
 874     /* Update outer/inner flops */
 875
 876     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*18 + inneriter*90);
 877 }