src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRF_VdwLJ_GeomW3P1_sse2_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 sse2_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_double.h"
  50 #include "kernelutil_x86_sse2_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sse2_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sse2_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  92     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  93     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  94     real             *charge;
  95     int              nvdwtype;
  96     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  97     int              *vdwtype;
  98     real             *vdwparam;
  99     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 100     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 101     __m128d          dummy_mask,cutoff_mask;
 102     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 103     __m128d          one     = _mm_set1_pd(1.0);
 104     __m128d          two     = _mm_set1_pd(2.0);
 105     x                = xx[0];
 106     f                = ff[0];
 107
 108     nri              = nlist->nri;
 109     iinr             = nlist->iinr;
 110     jindex           = nlist->jindex;
 111     jjnr             = nlist->jjnr;
 112     shiftidx         = nlist->shift;
 113     gid              = nlist->gid;
 114     shiftvec         = fr->shift_vec[0];
 115     fshift           = fr->fshift[0];
 116     facel            = _mm_set1_pd(fr->epsfac);
 117     charge           = mdatoms->chargeA;
 118     krf              = _mm_set1_pd(fr->ic->k_rf);
 119     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 120     crf              = _mm_set1_pd(fr->ic->c_rf);
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     /* Setup water-specific parameters */
 126     inr              = nlist->iinr[0];
 127     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 128     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 129     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 130     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136
 137     outeriter        = 0;
 138     inneriter        = 0;
 139
 140     /* Start outer loop over neighborlists */
 141     for(iidx=0; iidx<nri; iidx++)
 142     {
 143         /* Load shift vector for this list */
 144         i_shift_offset   = DIM*shiftidx[iidx];
 145
 146         /* Load limits for loop over neighbors */
 147         j_index_start    = jindex[iidx];
 148         j_index_end      = jindex[iidx+1];
 149
 150         /* Get outer coordinate index */
 151         inr              = iinr[iidx];
 152         i_coord_offset   = DIM*inr;
 153
 154         /* Load i particle coords and add shift vector */
 155         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 156                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 157
 158         fix0             = _mm_setzero_pd();
 159         fiy0             = _mm_setzero_pd();
 160         fiz0             = _mm_setzero_pd();
 161         fix1             = _mm_setzero_pd();
 162         fiy1             = _mm_setzero_pd();
 163         fiz1             = _mm_setzero_pd();
 164         fix2             = _mm_setzero_pd();
 165         fiy2             = _mm_setzero_pd();
 166         fiz2             = _mm_setzero_pd();
 167
 168         /* Reset potential sums */
 169         velecsum         = _mm_setzero_pd();
 170         vvdwsum          = _mm_setzero_pd();
 171
 172         /* Start inner kernel loop */
 173         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 174         {
 175
 176             /* Get j neighbor index, and coordinate index */
 177             jnrA             = jjnr[jidx];
 178             jnrB             = jjnr[jidx+1];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181
 182             /* load j atom coordinates */
 183             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_pd(ix0,jx0);
 188             dy00             = _mm_sub_pd(iy0,jy0);
 189             dz00             = _mm_sub_pd(iz0,jz0);
 190             dx10             = _mm_sub_pd(ix1,jx0);
 191             dy10             = _mm_sub_pd(iy1,jy0);
 192             dz10             = _mm_sub_pd(iz1,jz0);
 193             dx20             = _mm_sub_pd(ix2,jx0);
 194             dy20             = _mm_sub_pd(iy2,jy0);
 195             dz20             = _mm_sub_pd(iz2,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 199             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 200             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 201
 202             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 203             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 204             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 205
 206             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 207             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 208             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 209
 210             /* Load parameters for j particles */
 211             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 212             vdwjidx0A        = 2*vdwtype[jnrA+0];
 213             vdwjidx0B        = 2*vdwtype[jnrB+0];
 214
 215             fjx0             = _mm_setzero_pd();
 216             fjy0             = _mm_setzero_pd();
 217             fjz0             = _mm_setzero_pd();
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             /* Compute parameters for interactions between i and j atoms */
 224             qq00             = _mm_mul_pd(iq0,jq0);
 225             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 226                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 227
 228             /* REACTION-FIELD ELECTROSTATICS */
 229             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 230             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 231
 232             /* LENNARD-JONES DISPERSION/REPULSION */
 233
 234             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 235             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 236             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 237             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 238             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 239
 240             /* Update potential sum for this i atom from the interaction with this j atom. */
 241             velecsum         = _mm_add_pd(velecsum,velec);
 242             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 243
 244             fscal            = _mm_add_pd(felec,fvdw);
 245
 246             /* Calculate temporary vectorial force */
 247             tx               = _mm_mul_pd(fscal,dx00);
 248             ty               = _mm_mul_pd(fscal,dy00);
 249             tz               = _mm_mul_pd(fscal,dz00);
 250
 251             /* Update vectorial force */
 252             fix0             = _mm_add_pd(fix0,tx);
 253             fiy0             = _mm_add_pd(fiy0,ty);
 254             fiz0             = _mm_add_pd(fiz0,tz);
 255
 256             fjx0             = _mm_add_pd(fjx0,tx);
 257             fjy0             = _mm_add_pd(fjy0,ty);
 258             fjz0             = _mm_add_pd(fjz0,tz);
 259
 260             /**************************
 261              * CALCULATE INTERACTIONS *
 262              **************************/
 263
 264             /* Compute parameters for interactions between i and j atoms */
 265             qq10             = _mm_mul_pd(iq1,jq0);
 266
 267             /* REACTION-FIELD ELECTROSTATICS */
 268             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 269             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 270
 271             /* Update potential sum for this i atom from the interaction with this j atom. */
 272             velecsum         = _mm_add_pd(velecsum,velec);
 273
 274             fscal            = felec;
 275
 276             /* Calculate temporary vectorial force */
 277             tx               = _mm_mul_pd(fscal,dx10);
 278             ty               = _mm_mul_pd(fscal,dy10);
 279             tz               = _mm_mul_pd(fscal,dz10);
 280
 281             /* Update vectorial force */
 282             fix1             = _mm_add_pd(fix1,tx);
 283             fiy1             = _mm_add_pd(fiy1,ty);
 284             fiz1             = _mm_add_pd(fiz1,tz);
 285
 286             fjx0             = _mm_add_pd(fjx0,tx);
 287             fjy0             = _mm_add_pd(fjy0,ty);
 288             fjz0             = _mm_add_pd(fjz0,tz);
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             /* Compute parameters for interactions between i and j atoms */
 295             qq20             = _mm_mul_pd(iq2,jq0);
 296
 297             /* REACTION-FIELD ELECTROSTATICS */
 298             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 299             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 300
 301             /* Update potential sum for this i atom from the interaction with this j atom. */
 302             velecsum         = _mm_add_pd(velecsum,velec);
 303
 304             fscal            = felec;
 305
 306             /* Calculate temporary vectorial force */
 307             tx               = _mm_mul_pd(fscal,dx20);
 308             ty               = _mm_mul_pd(fscal,dy20);
 309             tz               = _mm_mul_pd(fscal,dz20);
 310
 311             /* Update vectorial force */
 312             fix2             = _mm_add_pd(fix2,tx);
 313             fiy2             = _mm_add_pd(fiy2,ty);
 314             fiz2             = _mm_add_pd(fiz2,tz);
 315
 316             fjx0             = _mm_add_pd(fjx0,tx);
 317             fjy0             = _mm_add_pd(fjy0,ty);
 318             fjz0             = _mm_add_pd(fjz0,tz);
 319
 320             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 321
 322             /* Inner loop uses 111 flops */
 323         }
 324
 325         if(jidx<j_index_end)
 326         {
 327
 328             jnrA             = jjnr[jidx];
 329             j_coord_offsetA  = DIM*jnrA;
 330
 331             /* load j atom coordinates */
 332             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 333                                               &jx0,&jy0,&jz0);
 334
 335             /* Calculate displacement vector */
 336             dx00             = _mm_sub_pd(ix0,jx0);
 337             dy00             = _mm_sub_pd(iy0,jy0);
 338             dz00             = _mm_sub_pd(iz0,jz0);
 339             dx10             = _mm_sub_pd(ix1,jx0);
 340             dy10             = _mm_sub_pd(iy1,jy0);
 341             dz10             = _mm_sub_pd(iz1,jz0);
 342             dx20             = _mm_sub_pd(ix2,jx0);
 343             dy20             = _mm_sub_pd(iy2,jy0);
 344             dz20             = _mm_sub_pd(iz2,jz0);
 345
 346             /* Calculate squared distance and things based on it */
 347             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 348             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 349             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 350
 351             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 352             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 353             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 354
 355             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 356             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 357             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 358
 359             /* Load parameters for j particles */
 360             jq0              = _mm_load_sd(charge+jnrA+0);
 361             vdwjidx0A        = 2*vdwtype[jnrA+0];
 362
 363             fjx0             = _mm_setzero_pd();
 364             fjy0             = _mm_setzero_pd();
 365             fjz0             = _mm_setzero_pd();
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             /* Compute parameters for interactions between i and j atoms */
 372             qq00             = _mm_mul_pd(iq0,jq0);
 373             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 374
 375             /* REACTION-FIELD ELECTROSTATICS */
 376             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 377             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 378
 379             /* LENNARD-JONES DISPERSION/REPULSION */
 380
 381             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 382             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 383             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 384             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 385             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 386
 387             /* Update potential sum for this i atom from the interaction with this j atom. */
 388             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 389             velecsum         = _mm_add_pd(velecsum,velec);
 390             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 391             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 392
 393             fscal            = _mm_add_pd(felec,fvdw);
 394
 395             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 396
 397             /* Calculate temporary vectorial force */
 398             tx               = _mm_mul_pd(fscal,dx00);
 399             ty               = _mm_mul_pd(fscal,dy00);
 400             tz               = _mm_mul_pd(fscal,dz00);
 401
 402             /* Update vectorial force */
 403             fix0             = _mm_add_pd(fix0,tx);
 404             fiy0             = _mm_add_pd(fiy0,ty);
 405             fiz0             = _mm_add_pd(fiz0,tz);
 406
 407             fjx0             = _mm_add_pd(fjx0,tx);
 408             fjy0             = _mm_add_pd(fjy0,ty);
 409             fjz0             = _mm_add_pd(fjz0,tz);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             /* Compute parameters for interactions between i and j atoms */
 416             qq10             = _mm_mul_pd(iq1,jq0);
 417
 418             /* REACTION-FIELD ELECTROSTATICS */
 419             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 420             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 421
 422             /* Update potential sum for this i atom from the interaction with this j atom. */
 423             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 424             velecsum         = _mm_add_pd(velecsum,velec);
 425
 426             fscal            = felec;
 427
 428             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 429
 430             /* Calculate temporary vectorial force */
 431             tx               = _mm_mul_pd(fscal,dx10);
 432             ty               = _mm_mul_pd(fscal,dy10);
 433             tz               = _mm_mul_pd(fscal,dz10);
 434
 435             /* Update vectorial force */
 436             fix1             = _mm_add_pd(fix1,tx);
 437             fiy1             = _mm_add_pd(fiy1,ty);
 438             fiz1             = _mm_add_pd(fiz1,tz);
 439
 440             fjx0             = _mm_add_pd(fjx0,tx);
 441             fjy0             = _mm_add_pd(fjy0,ty);
 442             fjz0             = _mm_add_pd(fjz0,tz);
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             /* Compute parameters for interactions between i and j atoms */
 449             qq20             = _mm_mul_pd(iq2,jq0);
 450
 451             /* REACTION-FIELD ELECTROSTATICS */
 452             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 453             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 457             velecsum         = _mm_add_pd(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 462
 463             /* Calculate temporary vectorial force */
 464             tx               = _mm_mul_pd(fscal,dx20);
 465             ty               = _mm_mul_pd(fscal,dy20);
 466             tz               = _mm_mul_pd(fscal,dz20);
 467
 468             /* Update vectorial force */
 469             fix2             = _mm_add_pd(fix2,tx);
 470             fiy2             = _mm_add_pd(fiy2,ty);
 471             fiz2             = _mm_add_pd(fiz2,tz);
 472
 473             fjx0             = _mm_add_pd(fjx0,tx);
 474             fjy0             = _mm_add_pd(fjy0,ty);
 475             fjz0             = _mm_add_pd(fjz0,tz);
 476
 477             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 478
 479             /* Inner loop uses 111 flops */
 480         }
 481
 482         /* End of innermost loop */
 483
 484         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 485                                               f+i_coord_offset,fshift+i_shift_offset);
 486
 487         ggid                        = gid[iidx];
 488         /* Update potential energies */
 489         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 490         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 491
 492         /* Increment number of inner iterations */
 493         inneriter                  += j_index_end - j_index_start;
 494
 495         /* Outer loop uses 20 flops */
 496     }
 497
 498     /* Increment number of outer iterations */
 499     outeriter        += nri;
 500
 501     /* Update outer/inner flops */
 502
 503     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*111);
 504 }
 505 /*
 506  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sse2_double
 507  * Electrostatics interaction: ReactionField
 508  * VdW interaction:            LennardJones
 509  * Geometry:                   Water3-Particle
 510  * Calculate force/pot:        Force
 511  */
 512 void
 513 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sse2_double
 514                     (t_nblist                    * gmx_restrict       nlist,
 515                      rvec                        * gmx_restrict          xx,
 516                      rvec                        * gmx_restrict          ff,
 517                      t_forcerec                  * gmx_restrict          fr,
 518                      t_mdatoms                   * gmx_restrict     mdatoms,
 519                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 520                      t_nrnb                      * gmx_restrict        nrnb)
 521 {
 522     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 523      * just 0 for non-waters.
 524      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 525      * jnr indices corresponding to data put in the four positions in the SIMD register.
 526      */
 527     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 528     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 529     int              jnrA,jnrB;
 530     int              j_coord_offsetA,j_coord_offsetB;
 531     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 532     real             rcutoff_scalar;
 533     real             *shiftvec,*fshift,*x,*f;
 534     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 535     int              vdwioffset0;
 536     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 537     int              vdwioffset1;
 538     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 539     int              vdwioffset2;
 540     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 541     int              vdwjidx0A,vdwjidx0B;
 542     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 543     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 544     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 545     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 546     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 547     real             *charge;
 548     int              nvdwtype;
 549     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 550     int              *vdwtype;
 551     real             *vdwparam;
 552     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 553     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 554     __m128d          dummy_mask,cutoff_mask;
 555     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 556     __m128d          one     = _mm_set1_pd(1.0);
 557     __m128d          two     = _mm_set1_pd(2.0);
 558     x                = xx[0];
 559     f                = ff[0];
 560
 561     nri              = nlist->nri;
 562     iinr             = nlist->iinr;
 563     jindex           = nlist->jindex;
 564     jjnr             = nlist->jjnr;
 565     shiftidx         = nlist->shift;
 566     gid              = nlist->gid;
 567     shiftvec         = fr->shift_vec[0];
 568     fshift           = fr->fshift[0];
 569     facel            = _mm_set1_pd(fr->epsfac);
 570     charge           = mdatoms->chargeA;
 571     krf              = _mm_set1_pd(fr->ic->k_rf);
 572     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 573     crf              = _mm_set1_pd(fr->ic->c_rf);
 574     nvdwtype         = fr->ntype;
 575     vdwparam         = fr->nbfp;
 576     vdwtype          = mdatoms->typeA;
 577
 578     /* Setup water-specific parameters */
 579     inr              = nlist->iinr[0];
 580     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 581     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 582     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 583     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 584
 585     /* Avoid stupid compiler warnings */
 586     jnrA = jnrB = 0;
 587     j_coord_offsetA = 0;
 588     j_coord_offsetB = 0;
 589
 590     outeriter        = 0;
 591     inneriter        = 0;
 592
 593     /* Start outer loop over neighborlists */
 594     for(iidx=0; iidx<nri; iidx++)
 595     {
 596         /* Load shift vector for this list */
 597         i_shift_offset   = DIM*shiftidx[iidx];
 598
 599         /* Load limits for loop over neighbors */
 600         j_index_start    = jindex[iidx];
 601         j_index_end      = jindex[iidx+1];
 602
 603         /* Get outer coordinate index */
 604         inr              = iinr[iidx];
 605         i_coord_offset   = DIM*inr;
 606
 607         /* Load i particle coords and add shift vector */
 608         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 609                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 610
 611         fix0             = _mm_setzero_pd();
 612         fiy0             = _mm_setzero_pd();
 613         fiz0             = _mm_setzero_pd();
 614         fix1             = _mm_setzero_pd();
 615         fiy1             = _mm_setzero_pd();
 616         fiz1             = _mm_setzero_pd();
 617         fix2             = _mm_setzero_pd();
 618         fiy2             = _mm_setzero_pd();
 619         fiz2             = _mm_setzero_pd();
 620
 621         /* Start inner kernel loop */
 622         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 623         {
 624
 625             /* Get j neighbor index, and coordinate index */
 626             jnrA             = jjnr[jidx];
 627             jnrB             = jjnr[jidx+1];
 628             j_coord_offsetA  = DIM*jnrA;
 629             j_coord_offsetB  = DIM*jnrB;
 630
 631             /* load j atom coordinates */
 632             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 633                                               &jx0,&jy0,&jz0);
 634
 635             /* Calculate displacement vector */
 636             dx00             = _mm_sub_pd(ix0,jx0);
 637             dy00             = _mm_sub_pd(iy0,jy0);
 638             dz00             = _mm_sub_pd(iz0,jz0);
 639             dx10             = _mm_sub_pd(ix1,jx0);
 640             dy10             = _mm_sub_pd(iy1,jy0);
 641             dz10             = _mm_sub_pd(iz1,jz0);
 642             dx20             = _mm_sub_pd(ix2,jx0);
 643             dy20             = _mm_sub_pd(iy2,jy0);
 644             dz20             = _mm_sub_pd(iz2,jz0);
 645
 646             /* Calculate squared distance and things based on it */
 647             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 648             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 649             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 650
 651             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 652             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 653             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 654
 655             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 656             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 657             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 658
 659             /* Load parameters for j particles */
 660             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 661             vdwjidx0A        = 2*vdwtype[jnrA+0];
 662             vdwjidx0B        = 2*vdwtype[jnrB+0];
 663
 664             fjx0             = _mm_setzero_pd();
 665             fjy0             = _mm_setzero_pd();
 666             fjz0             = _mm_setzero_pd();
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             /* Compute parameters for interactions between i and j atoms */
 673             qq00             = _mm_mul_pd(iq0,jq0);
 674             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 675                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 676
 677             /* REACTION-FIELD ELECTROSTATICS */
 678             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 679
 680             /* LENNARD-JONES DISPERSION/REPULSION */
 681
 682             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 683             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 684
 685             fscal            = _mm_add_pd(felec,fvdw);
 686
 687             /* Calculate temporary vectorial force */
 688             tx               = _mm_mul_pd(fscal,dx00);
 689             ty               = _mm_mul_pd(fscal,dy00);
 690             tz               = _mm_mul_pd(fscal,dz00);
 691
 692             /* Update vectorial force */
 693             fix0             = _mm_add_pd(fix0,tx);
 694             fiy0             = _mm_add_pd(fiy0,ty);
 695             fiz0             = _mm_add_pd(fiz0,tz);
 696
 697             fjx0             = _mm_add_pd(fjx0,tx);
 698             fjy0             = _mm_add_pd(fjy0,ty);
 699             fjz0             = _mm_add_pd(fjz0,tz);
 700
 701             /**************************
 702              * CALCULATE INTERACTIONS *
 703              **************************/
 704
 705             /* Compute parameters for interactions between i and j atoms */
 706             qq10             = _mm_mul_pd(iq1,jq0);
 707
 708             /* REACTION-FIELD ELECTROSTATICS */
 709             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 710
 711             fscal            = felec;
 712
 713             /* Calculate temporary vectorial force */
 714             tx               = _mm_mul_pd(fscal,dx10);
 715             ty               = _mm_mul_pd(fscal,dy10);
 716             tz               = _mm_mul_pd(fscal,dz10);
 717
 718             /* Update vectorial force */
 719             fix1             = _mm_add_pd(fix1,tx);
 720             fiy1             = _mm_add_pd(fiy1,ty);
 721             fiz1             = _mm_add_pd(fiz1,tz);
 722
 723             fjx0             = _mm_add_pd(fjx0,tx);
 724             fjy0             = _mm_add_pd(fjy0,ty);
 725             fjz0             = _mm_add_pd(fjz0,tz);
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             /* Compute parameters for interactions between i and j atoms */
 732             qq20             = _mm_mul_pd(iq2,jq0);
 733
 734             /* REACTION-FIELD ELECTROSTATICS */
 735             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 736
 737             fscal            = felec;
 738
 739             /* Calculate temporary vectorial force */
 740             tx               = _mm_mul_pd(fscal,dx20);
 741             ty               = _mm_mul_pd(fscal,dy20);
 742             tz               = _mm_mul_pd(fscal,dz20);
 743
 744             /* Update vectorial force */
 745             fix2             = _mm_add_pd(fix2,tx);
 746             fiy2             = _mm_add_pd(fiy2,ty);
 747             fiz2             = _mm_add_pd(fiz2,tz);
 748
 749             fjx0             = _mm_add_pd(fjx0,tx);
 750             fjy0             = _mm_add_pd(fjy0,ty);
 751             fjz0             = _mm_add_pd(fjz0,tz);
 752
 753             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 754
 755             /* Inner loop uses 91 flops */
 756         }
 757
 758         if(jidx<j_index_end)
 759         {
 760
 761             jnrA             = jjnr[jidx];
 762             j_coord_offsetA  = DIM*jnrA;
 763
 764             /* load j atom coordinates */
 765             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 766                                               &jx0,&jy0,&jz0);
 767
 768             /* Calculate displacement vector */
 769             dx00             = _mm_sub_pd(ix0,jx0);
 770             dy00             = _mm_sub_pd(iy0,jy0);
 771             dz00             = _mm_sub_pd(iz0,jz0);
 772             dx10             = _mm_sub_pd(ix1,jx0);
 773             dy10             = _mm_sub_pd(iy1,jy0);
 774             dz10             = _mm_sub_pd(iz1,jz0);
 775             dx20             = _mm_sub_pd(ix2,jx0);
 776             dy20             = _mm_sub_pd(iy2,jy0);
 777             dz20             = _mm_sub_pd(iz2,jz0);
 778
 779             /* Calculate squared distance and things based on it */
 780             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 781             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 782             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 783
 784             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 785             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 786             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 787
 788             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 789             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 790             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 791
 792             /* Load parameters for j particles */
 793             jq0              = _mm_load_sd(charge+jnrA+0);
 794             vdwjidx0A        = 2*vdwtype[jnrA+0];
 795
 796             fjx0             = _mm_setzero_pd();
 797             fjy0             = _mm_setzero_pd();
 798             fjz0             = _mm_setzero_pd();
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             /* Compute parameters for interactions between i and j atoms */
 805             qq00             = _mm_mul_pd(iq0,jq0);
 806             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 807
 808             /* REACTION-FIELD ELECTROSTATICS */
 809             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 810
 811             /* LENNARD-JONES DISPERSION/REPULSION */
 812
 813             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 814             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 815
 816             fscal            = _mm_add_pd(felec,fvdw);
 817
 818             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 819
 820             /* Calculate temporary vectorial force */
 821             tx               = _mm_mul_pd(fscal,dx00);
 822             ty               = _mm_mul_pd(fscal,dy00);
 823             tz               = _mm_mul_pd(fscal,dz00);
 824
 825             /* Update vectorial force */
 826             fix0             = _mm_add_pd(fix0,tx);
 827             fiy0             = _mm_add_pd(fiy0,ty);
 828             fiz0             = _mm_add_pd(fiz0,tz);
 829
 830             fjx0             = _mm_add_pd(fjx0,tx);
 831             fjy0             = _mm_add_pd(fjy0,ty);
 832             fjz0             = _mm_add_pd(fjz0,tz);
 833
 834             /**************************
 835              * CALCULATE INTERACTIONS *
 836              **************************/
 837
 838             /* Compute parameters for interactions between i and j atoms */
 839             qq10             = _mm_mul_pd(iq1,jq0);
 840
 841             /* REACTION-FIELD ELECTROSTATICS */
 842             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 843
 844             fscal            = felec;
 845
 846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 847
 848             /* Calculate temporary vectorial force */
 849             tx               = _mm_mul_pd(fscal,dx10);
 850             ty               = _mm_mul_pd(fscal,dy10);
 851             tz               = _mm_mul_pd(fscal,dz10);
 852
 853             /* Update vectorial force */
 854             fix1             = _mm_add_pd(fix1,tx);
 855             fiy1             = _mm_add_pd(fiy1,ty);
 856             fiz1             = _mm_add_pd(fiz1,tz);
 857
 858             fjx0             = _mm_add_pd(fjx0,tx);
 859             fjy0             = _mm_add_pd(fjy0,ty);
 860             fjz0             = _mm_add_pd(fjz0,tz);
 861
 862             /**************************
 863              * CALCULATE INTERACTIONS *
 864              **************************/
 865
 866             /* Compute parameters for interactions between i and j atoms */
 867             qq20             = _mm_mul_pd(iq2,jq0);
 868
 869             /* REACTION-FIELD ELECTROSTATICS */
 870             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 871
 872             fscal            = felec;
 873
 874             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 875
 876             /* Calculate temporary vectorial force */
 877             tx               = _mm_mul_pd(fscal,dx20);
 878             ty               = _mm_mul_pd(fscal,dy20);
 879             tz               = _mm_mul_pd(fscal,dz20);
 880
 881             /* Update vectorial force */
 882             fix2             = _mm_add_pd(fix2,tx);
 883             fiy2             = _mm_add_pd(fiy2,ty);
 884             fiz2             = _mm_add_pd(fiz2,tz);
 885
 886             fjx0             = _mm_add_pd(fjx0,tx);
 887             fjy0             = _mm_add_pd(fjy0,ty);
 888             fjz0             = _mm_add_pd(fjz0,tz);
 889
 890             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 891
 892             /* Inner loop uses 91 flops */
 893         }
 894
 895         /* End of innermost loop */
 896
 897         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 898                                               f+i_coord_offset,fshift+i_shift_offset);
 899
 900         /* Increment number of inner iterations */
 901         inneriter                  += j_index_end - j_index_start;
 902
 903         /* Outer loop uses 18 flops */
 904     }
 905
 906     /* Increment number of outer iterations */
 907     outeriter        += nri;
 908
 909     /* Update outer/inner flops */
 910
 911     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*91);
 912 }