src/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRF_VdwLJ_GeomW3P1_sse2_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse2_double.h"
  34 #include "kernelutil_x86_sse2_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sse2_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sse2_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  76     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  77     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     int              nvdwtype;
  80     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  81     int              *vdwtype;
  82     real             *vdwparam;
  83     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  84     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  85     __m128d          dummy_mask,cutoff_mask;
  86     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  87     __m128d          one     = _mm_set1_pd(1.0);
  88     __m128d          two     = _mm_set1_pd(2.0);
  89     x                = xx[0];
  90     f                = ff[0];
  91
  92     nri              = nlist->nri;
  93     iinr             = nlist->iinr;
  94     jindex           = nlist->jindex;
  95     jjnr             = nlist->jjnr;
  96     shiftidx         = nlist->shift;
  97     gid              = nlist->gid;
  98     shiftvec         = fr->shift_vec[0];
  99     fshift           = fr->fshift[0];
 100     facel            = _mm_set1_pd(fr->epsfac);
 101     charge           = mdatoms->chargeA;
 102     krf              = _mm_set1_pd(fr->ic->k_rf);
 103     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 104     crf              = _mm_set1_pd(fr->ic->c_rf);
 105     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     /* Setup water-specific parameters */
 110     inr              = nlist->iinr[0];
 111     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 112     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 113     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 114     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 115
 116     /* Avoid stupid compiler warnings */
 117     jnrA = jnrB = 0;
 118     j_coord_offsetA = 0;
 119     j_coord_offsetB = 0;
 120
 121     outeriter        = 0;
 122     inneriter        = 0;
 123
 124     /* Start outer loop over neighborlists */
 125     for(iidx=0; iidx<nri; iidx++)
 126     {
 127         /* Load shift vector for this list */
 128         i_shift_offset   = DIM*shiftidx[iidx];
 129
 130         /* Load limits for loop over neighbors */
 131         j_index_start    = jindex[iidx];
 132         j_index_end      = jindex[iidx+1];
 133
 134         /* Get outer coordinate index */
 135         inr              = iinr[iidx];
 136         i_coord_offset   = DIM*inr;
 137
 138         /* Load i particle coords and add shift vector */
 139         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 140                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 141
 142         fix0             = _mm_setzero_pd();
 143         fiy0             = _mm_setzero_pd();
 144         fiz0             = _mm_setzero_pd();
 145         fix1             = _mm_setzero_pd();
 146         fiy1             = _mm_setzero_pd();
 147         fiz1             = _mm_setzero_pd();
 148         fix2             = _mm_setzero_pd();
 149         fiy2             = _mm_setzero_pd();
 150         fiz2             = _mm_setzero_pd();
 151
 152         /* Reset potential sums */
 153         velecsum         = _mm_setzero_pd();
 154         vvdwsum          = _mm_setzero_pd();
 155
 156         /* Start inner kernel loop */
 157         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 158         {
 159
 160             /* Get j neighbor index, and coordinate index */
 161             jnrA             = jjnr[jidx];
 162             jnrB             = jjnr[jidx+1];
 163             j_coord_offsetA  = DIM*jnrA;
 164             j_coord_offsetB  = DIM*jnrB;
 165
 166             /* load j atom coordinates */
 167             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 168                                               &jx0,&jy0,&jz0);
 169
 170             /* Calculate displacement vector */
 171             dx00             = _mm_sub_pd(ix0,jx0);
 172             dy00             = _mm_sub_pd(iy0,jy0);
 173             dz00             = _mm_sub_pd(iz0,jz0);
 174             dx10             = _mm_sub_pd(ix1,jx0);
 175             dy10             = _mm_sub_pd(iy1,jy0);
 176             dz10             = _mm_sub_pd(iz1,jz0);
 177             dx20             = _mm_sub_pd(ix2,jx0);
 178             dy20             = _mm_sub_pd(iy2,jy0);
 179             dz20             = _mm_sub_pd(iz2,jz0);
 180
 181             /* Calculate squared distance and things based on it */
 182             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 183             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 184             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 185
 186             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 187             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 188             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 189
 190             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 191             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 192             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 193
 194             /* Load parameters for j particles */
 195             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 196             vdwjidx0A        = 2*vdwtype[jnrA+0];
 197             vdwjidx0B        = 2*vdwtype[jnrB+0];
 198
 199             fjx0             = _mm_setzero_pd();
 200             fjy0             = _mm_setzero_pd();
 201             fjz0             = _mm_setzero_pd();
 202
 203             /**************************
 204              * CALCULATE INTERACTIONS *
 205              **************************/
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             qq00             = _mm_mul_pd(iq0,jq0);
 209             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 210                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 211
 212             /* REACTION-FIELD ELECTROSTATICS */
 213             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 214             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 215
 216             /* LENNARD-JONES DISPERSION/REPULSION */
 217
 218             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 219             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 220             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 221             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 222             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 223
 224             /* Update potential sum for this i atom from the interaction with this j atom. */
 225             velecsum         = _mm_add_pd(velecsum,velec);
 226             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 227
 228             fscal            = _mm_add_pd(felec,fvdw);
 229
 230             /* Calculate temporary vectorial force */
 231             tx               = _mm_mul_pd(fscal,dx00);
 232             ty               = _mm_mul_pd(fscal,dy00);
 233             tz               = _mm_mul_pd(fscal,dz00);
 234
 235             /* Update vectorial force */
 236             fix0             = _mm_add_pd(fix0,tx);
 237             fiy0             = _mm_add_pd(fiy0,ty);
 238             fiz0             = _mm_add_pd(fiz0,tz);
 239
 240             fjx0             = _mm_add_pd(fjx0,tx);
 241             fjy0             = _mm_add_pd(fjy0,ty);
 242             fjz0             = _mm_add_pd(fjz0,tz);
 243
 244             /**************************
 245              * CALCULATE INTERACTIONS *
 246              **************************/
 247
 248             /* Compute parameters for interactions between i and j atoms */
 249             qq10             = _mm_mul_pd(iq1,jq0);
 250
 251             /* REACTION-FIELD ELECTROSTATICS */
 252             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 253             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 254
 255             /* Update potential sum for this i atom from the interaction with this j atom. */
 256             velecsum         = _mm_add_pd(velecsum,velec);
 257
 258             fscal            = felec;
 259
 260             /* Calculate temporary vectorial force */
 261             tx               = _mm_mul_pd(fscal,dx10);
 262             ty               = _mm_mul_pd(fscal,dy10);
 263             tz               = _mm_mul_pd(fscal,dz10);
 264
 265             /* Update vectorial force */
 266             fix1             = _mm_add_pd(fix1,tx);
 267             fiy1             = _mm_add_pd(fiy1,ty);
 268             fiz1             = _mm_add_pd(fiz1,tz);
 269
 270             fjx0             = _mm_add_pd(fjx0,tx);
 271             fjy0             = _mm_add_pd(fjy0,ty);
 272             fjz0             = _mm_add_pd(fjz0,tz);
 273
 274             /**************************
 275              * CALCULATE INTERACTIONS *
 276              **************************/
 277
 278             /* Compute parameters for interactions between i and j atoms */
 279             qq20             = _mm_mul_pd(iq2,jq0);
 280
 281             /* REACTION-FIELD ELECTROSTATICS */
 282             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 283             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velecsum         = _mm_add_pd(velecsum,velec);
 287
 288             fscal            = felec;
 289
 290             /* Calculate temporary vectorial force */
 291             tx               = _mm_mul_pd(fscal,dx20);
 292             ty               = _mm_mul_pd(fscal,dy20);
 293             tz               = _mm_mul_pd(fscal,dz20);
 294
 295             /* Update vectorial force */
 296             fix2             = _mm_add_pd(fix2,tx);
 297             fiy2             = _mm_add_pd(fiy2,ty);
 298             fiz2             = _mm_add_pd(fiz2,tz);
 299
 300             fjx0             = _mm_add_pd(fjx0,tx);
 301             fjy0             = _mm_add_pd(fjy0,ty);
 302             fjz0             = _mm_add_pd(fjz0,tz);
 303
 304             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 305
 306             /* Inner loop uses 111 flops */
 307         }
 308
 309         if(jidx<j_index_end)
 310         {
 311
 312             jnrA             = jjnr[jidx];
 313             j_coord_offsetA  = DIM*jnrA;
 314
 315             /* load j atom coordinates */
 316             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 317                                               &jx0,&jy0,&jz0);
 318
 319             /* Calculate displacement vector */
 320             dx00             = _mm_sub_pd(ix0,jx0);
 321             dy00             = _mm_sub_pd(iy0,jy0);
 322             dz00             = _mm_sub_pd(iz0,jz0);
 323             dx10             = _mm_sub_pd(ix1,jx0);
 324             dy10             = _mm_sub_pd(iy1,jy0);
 325             dz10             = _mm_sub_pd(iz1,jz0);
 326             dx20             = _mm_sub_pd(ix2,jx0);
 327             dy20             = _mm_sub_pd(iy2,jy0);
 328             dz20             = _mm_sub_pd(iz2,jz0);
 329
 330             /* Calculate squared distance and things based on it */
 331             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 332             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 333             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 334
 335             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 336             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 337             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 338
 339             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 340             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 341             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 342
 343             /* Load parameters for j particles */
 344             jq0              = _mm_load_sd(charge+jnrA+0);
 345             vdwjidx0A        = 2*vdwtype[jnrA+0];
 346
 347             fjx0             = _mm_setzero_pd();
 348             fjy0             = _mm_setzero_pd();
 349             fjz0             = _mm_setzero_pd();
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             /* Compute parameters for interactions between i and j atoms */
 356             qq00             = _mm_mul_pd(iq0,jq0);
 357             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 358
 359             /* REACTION-FIELD ELECTROSTATICS */
 360             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 361             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 362
 363             /* LENNARD-JONES DISPERSION/REPULSION */
 364
 365             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 366             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 367             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 368             vvdw             = _mm_sub_pd( _mm_mul_pd(vvdw12,one_twelfth) , _mm_mul_pd(vvdw6,one_sixth) );
 369             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 373             velecsum         = _mm_add_pd(velecsum,velec);
 374             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 375             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 376
 377             fscal            = _mm_add_pd(felec,fvdw);
 378
 379             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm_mul_pd(fscal,dx00);
 383             ty               = _mm_mul_pd(fscal,dy00);
 384             tz               = _mm_mul_pd(fscal,dz00);
 385
 386             /* Update vectorial force */
 387             fix0             = _mm_add_pd(fix0,tx);
 388             fiy0             = _mm_add_pd(fiy0,ty);
 389             fiz0             = _mm_add_pd(fiz0,tz);
 390
 391             fjx0             = _mm_add_pd(fjx0,tx);
 392             fjy0             = _mm_add_pd(fjy0,ty);
 393             fjz0             = _mm_add_pd(fjz0,tz);
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             /* Compute parameters for interactions between i and j atoms */
 400             qq10             = _mm_mul_pd(iq1,jq0);
 401
 402             /* REACTION-FIELD ELECTROSTATICS */
 403             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 404             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 408             velecsum         = _mm_add_pd(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 413
 414             /* Calculate temporary vectorial force */
 415             tx               = _mm_mul_pd(fscal,dx10);
 416             ty               = _mm_mul_pd(fscal,dy10);
 417             tz               = _mm_mul_pd(fscal,dz10);
 418
 419             /* Update vectorial force */
 420             fix1             = _mm_add_pd(fix1,tx);
 421             fiy1             = _mm_add_pd(fiy1,ty);
 422             fiz1             = _mm_add_pd(fiz1,tz);
 423
 424             fjx0             = _mm_add_pd(fjx0,tx);
 425             fjy0             = _mm_add_pd(fjy0,ty);
 426             fjz0             = _mm_add_pd(fjz0,tz);
 427
 428             /**************************
 429              * CALCULATE INTERACTIONS *
 430              **************************/
 431
 432             /* Compute parameters for interactions between i and j atoms */
 433             qq20             = _mm_mul_pd(iq2,jq0);
 434
 435             /* REACTION-FIELD ELECTROSTATICS */
 436             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 437             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 438
 439             /* Update potential sum for this i atom from the interaction with this j atom. */
 440             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 441             velecsum         = _mm_add_pd(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 446
 447             /* Calculate temporary vectorial force */
 448             tx               = _mm_mul_pd(fscal,dx20);
 449             ty               = _mm_mul_pd(fscal,dy20);
 450             tz               = _mm_mul_pd(fscal,dz20);
 451
 452             /* Update vectorial force */
 453             fix2             = _mm_add_pd(fix2,tx);
 454             fiy2             = _mm_add_pd(fiy2,ty);
 455             fiz2             = _mm_add_pd(fiz2,tz);
 456
 457             fjx0             = _mm_add_pd(fjx0,tx);
 458             fjy0             = _mm_add_pd(fjy0,ty);
 459             fjz0             = _mm_add_pd(fjz0,tz);
 460
 461             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 462
 463             /* Inner loop uses 111 flops */
 464         }
 465
 466         /* End of innermost loop */
 467
 468         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 469                                               f+i_coord_offset,fshift+i_shift_offset);
 470
 471         ggid                        = gid[iidx];
 472         /* Update potential energies */
 473         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 474         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 475
 476         /* Increment number of inner iterations */
 477         inneriter                  += j_index_end - j_index_start;
 478
 479         /* Outer loop uses 20 flops */
 480     }
 481
 482     /* Increment number of outer iterations */
 483     outeriter        += nri;
 484
 485     /* Update outer/inner flops */
 486
 487     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*111);
 488 }
 489 /*
 490  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sse2_double
 491  * Electrostatics interaction: ReactionField
 492  * VdW interaction:            LennardJones
 493  * Geometry:                   Water3-Particle
 494  * Calculate force/pot:        Force
 495  */
 496 void
 497 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sse2_double
 498                     (t_nblist * gmx_restrict                nlist,
 499                      rvec * gmx_restrict                    xx,
 500                      rvec * gmx_restrict                    ff,
 501                      t_forcerec * gmx_restrict              fr,
 502                      t_mdatoms * gmx_restrict               mdatoms,
 503                      nb_kernel_data_t * gmx_restrict        kernel_data,
 504                      t_nrnb * gmx_restrict                  nrnb)
 505 {
 506     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 507      * just 0 for non-waters.
 508      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 509      * jnr indices corresponding to data put in the four positions in the SIMD register.
 510      */
 511     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 512     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 513     int              jnrA,jnrB;
 514     int              j_coord_offsetA,j_coord_offsetB;
 515     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 516     real             rcutoff_scalar;
 517     real             *shiftvec,*fshift,*x,*f;
 518     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 519     int              vdwioffset0;
 520     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 521     int              vdwioffset1;
 522     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 523     int              vdwioffset2;
 524     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 525     int              vdwjidx0A,vdwjidx0B;
 526     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 527     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 528     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 529     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 530     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 531     real             *charge;
 532     int              nvdwtype;
 533     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 534     int              *vdwtype;
 535     real             *vdwparam;
 536     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 537     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 538     __m128d          dummy_mask,cutoff_mask;
 539     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 540     __m128d          one     = _mm_set1_pd(1.0);
 541     __m128d          two     = _mm_set1_pd(2.0);
 542     x                = xx[0];
 543     f                = ff[0];
 544
 545     nri              = nlist->nri;
 546     iinr             = nlist->iinr;
 547     jindex           = nlist->jindex;
 548     jjnr             = nlist->jjnr;
 549     shiftidx         = nlist->shift;
 550     gid              = nlist->gid;
 551     shiftvec         = fr->shift_vec[0];
 552     fshift           = fr->fshift[0];
 553     facel            = _mm_set1_pd(fr->epsfac);
 554     charge           = mdatoms->chargeA;
 555     krf              = _mm_set1_pd(fr->ic->k_rf);
 556     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 557     crf              = _mm_set1_pd(fr->ic->c_rf);
 558     nvdwtype         = fr->ntype;
 559     vdwparam         = fr->nbfp;
 560     vdwtype          = mdatoms->typeA;
 561
 562     /* Setup water-specific parameters */
 563     inr              = nlist->iinr[0];
 564     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 565     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 566     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 567     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 568
 569     /* Avoid stupid compiler warnings */
 570     jnrA = jnrB = 0;
 571     j_coord_offsetA = 0;
 572     j_coord_offsetB = 0;
 573
 574     outeriter        = 0;
 575     inneriter        = 0;
 576
 577     /* Start outer loop over neighborlists */
 578     for(iidx=0; iidx<nri; iidx++)
 579     {
 580         /* Load shift vector for this list */
 581         i_shift_offset   = DIM*shiftidx[iidx];
 582
 583         /* Load limits for loop over neighbors */
 584         j_index_start    = jindex[iidx];
 585         j_index_end      = jindex[iidx+1];
 586
 587         /* Get outer coordinate index */
 588         inr              = iinr[iidx];
 589         i_coord_offset   = DIM*inr;
 590
 591         /* Load i particle coords and add shift vector */
 592         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 593                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 594
 595         fix0             = _mm_setzero_pd();
 596         fiy0             = _mm_setzero_pd();
 597         fiz0             = _mm_setzero_pd();
 598         fix1             = _mm_setzero_pd();
 599         fiy1             = _mm_setzero_pd();
 600         fiz1             = _mm_setzero_pd();
 601         fix2             = _mm_setzero_pd();
 602         fiy2             = _mm_setzero_pd();
 603         fiz2             = _mm_setzero_pd();
 604
 605         /* Start inner kernel loop */
 606         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 607         {
 608
 609             /* Get j neighbor index, and coordinate index */
 610             jnrA             = jjnr[jidx];
 611             jnrB             = jjnr[jidx+1];
 612             j_coord_offsetA  = DIM*jnrA;
 613             j_coord_offsetB  = DIM*jnrB;
 614
 615             /* load j atom coordinates */
 616             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 617                                               &jx0,&jy0,&jz0);
 618
 619             /* Calculate displacement vector */
 620             dx00             = _mm_sub_pd(ix0,jx0);
 621             dy00             = _mm_sub_pd(iy0,jy0);
 622             dz00             = _mm_sub_pd(iz0,jz0);
 623             dx10             = _mm_sub_pd(ix1,jx0);
 624             dy10             = _mm_sub_pd(iy1,jy0);
 625             dz10             = _mm_sub_pd(iz1,jz0);
 626             dx20             = _mm_sub_pd(ix2,jx0);
 627             dy20             = _mm_sub_pd(iy2,jy0);
 628             dz20             = _mm_sub_pd(iz2,jz0);
 629
 630             /* Calculate squared distance and things based on it */
 631             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 632             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 633             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 634
 635             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 636             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 637             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 638
 639             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 640             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 641             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 642
 643             /* Load parameters for j particles */
 644             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 645             vdwjidx0A        = 2*vdwtype[jnrA+0];
 646             vdwjidx0B        = 2*vdwtype[jnrB+0];
 647
 648             fjx0             = _mm_setzero_pd();
 649             fjy0             = _mm_setzero_pd();
 650             fjz0             = _mm_setzero_pd();
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             /* Compute parameters for interactions between i and j atoms */
 657             qq00             = _mm_mul_pd(iq0,jq0);
 658             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 659                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 660
 661             /* REACTION-FIELD ELECTROSTATICS */
 662             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 663
 664             /* LENNARD-JONES DISPERSION/REPULSION */
 665
 666             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 667             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 668
 669             fscal            = _mm_add_pd(felec,fvdw);
 670
 671             /* Calculate temporary vectorial force */
 672             tx               = _mm_mul_pd(fscal,dx00);
 673             ty               = _mm_mul_pd(fscal,dy00);
 674             tz               = _mm_mul_pd(fscal,dz00);
 675
 676             /* Update vectorial force */
 677             fix0             = _mm_add_pd(fix0,tx);
 678             fiy0             = _mm_add_pd(fiy0,ty);
 679             fiz0             = _mm_add_pd(fiz0,tz);
 680
 681             fjx0             = _mm_add_pd(fjx0,tx);
 682             fjy0             = _mm_add_pd(fjy0,ty);
 683             fjz0             = _mm_add_pd(fjz0,tz);
 684
 685             /**************************
 686              * CALCULATE INTERACTIONS *
 687              **************************/
 688
 689             /* Compute parameters for interactions between i and j atoms */
 690             qq10             = _mm_mul_pd(iq1,jq0);
 691
 692             /* REACTION-FIELD ELECTROSTATICS */
 693             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 694
 695             fscal            = felec;
 696
 697             /* Calculate temporary vectorial force */
 698             tx               = _mm_mul_pd(fscal,dx10);
 699             ty               = _mm_mul_pd(fscal,dy10);
 700             tz               = _mm_mul_pd(fscal,dz10);
 701
 702             /* Update vectorial force */
 703             fix1             = _mm_add_pd(fix1,tx);
 704             fiy1             = _mm_add_pd(fiy1,ty);
 705             fiz1             = _mm_add_pd(fiz1,tz);
 706
 707             fjx0             = _mm_add_pd(fjx0,tx);
 708             fjy0             = _mm_add_pd(fjy0,ty);
 709             fjz0             = _mm_add_pd(fjz0,tz);
 710
 711             /**************************
 712              * CALCULATE INTERACTIONS *
 713              **************************/
 714
 715             /* Compute parameters for interactions between i and j atoms */
 716             qq20             = _mm_mul_pd(iq2,jq0);
 717
 718             /* REACTION-FIELD ELECTROSTATICS */
 719             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 720
 721             fscal            = felec;
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm_mul_pd(fscal,dx20);
 725             ty               = _mm_mul_pd(fscal,dy20);
 726             tz               = _mm_mul_pd(fscal,dz20);
 727
 728             /* Update vectorial force */
 729             fix2             = _mm_add_pd(fix2,tx);
 730             fiy2             = _mm_add_pd(fiy2,ty);
 731             fiz2             = _mm_add_pd(fiz2,tz);
 732
 733             fjx0             = _mm_add_pd(fjx0,tx);
 734             fjy0             = _mm_add_pd(fjy0,ty);
 735             fjz0             = _mm_add_pd(fjz0,tz);
 736
 737             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 738
 739             /* Inner loop uses 91 flops */
 740         }
 741
 742         if(jidx<j_index_end)
 743         {
 744
 745             jnrA             = jjnr[jidx];
 746             j_coord_offsetA  = DIM*jnrA;
 747
 748             /* load j atom coordinates */
 749             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 750                                               &jx0,&jy0,&jz0);
 751
 752             /* Calculate displacement vector */
 753             dx00             = _mm_sub_pd(ix0,jx0);
 754             dy00             = _mm_sub_pd(iy0,jy0);
 755             dz00             = _mm_sub_pd(iz0,jz0);
 756             dx10             = _mm_sub_pd(ix1,jx0);
 757             dy10             = _mm_sub_pd(iy1,jy0);
 758             dz10             = _mm_sub_pd(iz1,jz0);
 759             dx20             = _mm_sub_pd(ix2,jx0);
 760             dy20             = _mm_sub_pd(iy2,jy0);
 761             dz20             = _mm_sub_pd(iz2,jz0);
 762
 763             /* Calculate squared distance and things based on it */
 764             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 765             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 766             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 767
 768             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 769             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 770             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 771
 772             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 773             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 774             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 775
 776             /* Load parameters for j particles */
 777             jq0              = _mm_load_sd(charge+jnrA+0);
 778             vdwjidx0A        = 2*vdwtype[jnrA+0];
 779
 780             fjx0             = _mm_setzero_pd();
 781             fjy0             = _mm_setzero_pd();
 782             fjz0             = _mm_setzero_pd();
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* Compute parameters for interactions between i and j atoms */
 789             qq00             = _mm_mul_pd(iq0,jq0);
 790             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 791
 792             /* REACTION-FIELD ELECTROSTATICS */
 793             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 794
 795             /* LENNARD-JONES DISPERSION/REPULSION */
 796
 797             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 798             fvdw             = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(c12_00,rinvsix),c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 799
 800             fscal            = _mm_add_pd(felec,fvdw);
 801
 802             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 803
 804             /* Calculate temporary vectorial force */
 805             tx               = _mm_mul_pd(fscal,dx00);
 806             ty               = _mm_mul_pd(fscal,dy00);
 807             tz               = _mm_mul_pd(fscal,dz00);
 808
 809             /* Update vectorial force */
 810             fix0             = _mm_add_pd(fix0,tx);
 811             fiy0             = _mm_add_pd(fiy0,ty);
 812             fiz0             = _mm_add_pd(fiz0,tz);
 813
 814             fjx0             = _mm_add_pd(fjx0,tx);
 815             fjy0             = _mm_add_pd(fjy0,ty);
 816             fjz0             = _mm_add_pd(fjz0,tz);
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             /* Compute parameters for interactions between i and j atoms */
 823             qq10             = _mm_mul_pd(iq1,jq0);
 824
 825             /* REACTION-FIELD ELECTROSTATICS */
 826             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 827
 828             fscal            = felec;
 829
 830             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 831
 832             /* Calculate temporary vectorial force */
 833             tx               = _mm_mul_pd(fscal,dx10);
 834             ty               = _mm_mul_pd(fscal,dy10);
 835             tz               = _mm_mul_pd(fscal,dz10);
 836
 837             /* Update vectorial force */
 838             fix1             = _mm_add_pd(fix1,tx);
 839             fiy1             = _mm_add_pd(fiy1,ty);
 840             fiz1             = _mm_add_pd(fiz1,tz);
 841
 842             fjx0             = _mm_add_pd(fjx0,tx);
 843             fjy0             = _mm_add_pd(fjy0,ty);
 844             fjz0             = _mm_add_pd(fjz0,tz);
 845
 846             /**************************
 847              * CALCULATE INTERACTIONS *
 848              **************************/
 849
 850             /* Compute parameters for interactions between i and j atoms */
 851             qq20             = _mm_mul_pd(iq2,jq0);
 852
 853             /* REACTION-FIELD ELECTROSTATICS */
 854             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 859
 860             /* Calculate temporary vectorial force */
 861             tx               = _mm_mul_pd(fscal,dx20);
 862             ty               = _mm_mul_pd(fscal,dy20);
 863             tz               = _mm_mul_pd(fscal,dz20);
 864
 865             /* Update vectorial force */
 866             fix2             = _mm_add_pd(fix2,tx);
 867             fiy2             = _mm_add_pd(fiy2,ty);
 868             fiz2             = _mm_add_pd(fiz2,tz);
 869
 870             fjx0             = _mm_add_pd(fjx0,tx);
 871             fjy0             = _mm_add_pd(fjy0,ty);
 872             fjz0             = _mm_add_pd(fjz0,tz);
 873
 874             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 875
 876             /* Inner loop uses 91 flops */
 877         }
 878
 879         /* End of innermost loop */
 880
 881         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 882                                               f+i_coord_offset,fshift+i_shift_offset);
 883
 884         /* Increment number of inner iterations */
 885         inneriter                  += j_index_end - j_index_start;
 886
 887         /* Outer loop uses 18 flops */
 888     }
 889
 890     /* Increment number of outer iterations */
 891     outeriter        += nri;
 892
 893     /* Update outer/inner flops */
 894
 895     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*91);
 896 }