src/gromacs/gmxlib/nonbonded/nb_kernel_sparc64_hpc_ace_double/nb_kernel_ElecRF_VdwLJ_GeomW3P1_sparc64_hpc_ace_double.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 sparc64_hpc_ace_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 "gromacs/legacyheaders/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "kernelutil_sparc64_hpc_ace_double.h"
  50
  51 /*
  52  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sparc64_hpc_ace_double
  53  * Electrostatics interaction: ReactionField
  54  * VdW interaction:            LennardJones
  55  * Geometry:                   Water3-Particle
  56  * Calculate force/pot:        PotentialAndForce
  57  */
  58 void
  59 nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_sparc64_hpc_ace_double
  60                     (t_nblist                    * gmx_restrict       nlist,
  61                      rvec                        * gmx_restrict          xx,
  62                      rvec                        * gmx_restrict          ff,
  63                      t_forcerec                  * gmx_restrict          fr,
  64                      t_mdatoms                   * gmx_restrict     mdatoms,
  65                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  66                      t_nrnb                      * gmx_restrict        nrnb)
  67 {
  68     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  69      * just 0 for non-waters.
  70      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
  71      * jnr indices corresponding to data put in the four positions in the SIMD register.
  72      */
  73     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  74     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  75     int              jnrA,jnrB;
  76     int              j_coord_offsetA,j_coord_offsetB;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  81     int              vdwioffset0;
  82     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  83     int              vdwioffset1;
  84     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwjidx0A,vdwjidx0B;
  88     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  91     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  92     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
  99     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 100     _fjsp_v2r8       itab_tmp;
 101     _fjsp_v2r8       dummy_mask,cutoff_mask;
 102     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 103     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 104     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 105
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 120     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 121     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 129     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 130     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     /* Avoid stupid compiler warnings */
 134     jnrA = jnrB = 0;
 135     j_coord_offsetA = 0;
 136     j_coord_offsetB = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 157                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 158
 159         fix0             = _fjsp_setzero_v2r8();
 160         fiy0             = _fjsp_setzero_v2r8();
 161         fiz0             = _fjsp_setzero_v2r8();
 162         fix1             = _fjsp_setzero_v2r8();
 163         fiy1             = _fjsp_setzero_v2r8();
 164         fiz1             = _fjsp_setzero_v2r8();
 165         fix2             = _fjsp_setzero_v2r8();
 166         fiy2             = _fjsp_setzero_v2r8();
 167         fiz2             = _fjsp_setzero_v2r8();
 168
 169         /* Reset potential sums */
 170         velecsum         = _fjsp_setzero_v2r8();
 171         vvdwsum          = _fjsp_setzero_v2r8();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182
 183             /* load j atom coordinates */
 184             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 185                                               &jx0,&jy0,&jz0);
 186
 187             /* Calculate displacement vector */
 188             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 189             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 190             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 191             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 192             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 193             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 194             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 195             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 196             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 197
 198             /* Calculate squared distance and things based on it */
 199             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 200             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 201             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 202
 203             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 204             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 205             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 206
 207             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 208             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 209             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 213             vdwjidx0A        = 2*vdwtype[jnrA+0];
 214             vdwjidx0B        = 2*vdwtype[jnrB+0];
 215
 216             fjx0             = _fjsp_setzero_v2r8();
 217             fjy0             = _fjsp_setzero_v2r8();
 218             fjz0             = _fjsp_setzero_v2r8();
 219
 220             /**************************
 221              * CALCULATE INTERACTIONS *
 222              **************************/
 223
 224             /* Compute parameters for interactions between i and j atoms */
 225             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 226             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 227                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 228
 229             /* REACTION-FIELD ELECTROSTATICS */
 230             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 231             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 232
 233             /* LENNARD-JONES DISPERSION/REPULSION */
 234
 235             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 236             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 237             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 238             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 239             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 240
 241             /* Update potential sum for this i atom from the interaction with this j atom. */
 242             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 243             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 244
 245             fscal            = _fjsp_add_v2r8(felec,fvdw);
 246
 247             /* Update vectorial force */
 248             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 249             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 250             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 251
 252             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 253             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 254             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 255
 256             /**************************
 257              * CALCULATE INTERACTIONS *
 258              **************************/
 259
 260             /* Compute parameters for interactions between i and j atoms */
 261             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 262
 263             /* REACTION-FIELD ELECTROSTATICS */
 264             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 265             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 266
 267             /* Update potential sum for this i atom from the interaction with this j atom. */
 268             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 269
 270             fscal            = felec;
 271
 272             /* Update vectorial force */
 273             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 274             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 275             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 276
 277             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 278             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 279             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 280
 281             /**************************
 282              * CALCULATE INTERACTIONS *
 283              **************************/
 284
 285             /* Compute parameters for interactions between i and j atoms */
 286             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 287
 288             /* REACTION-FIELD ELECTROSTATICS */
 289             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 290             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 291
 292             /* Update potential sum for this i atom from the interaction with this j atom. */
 293             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 294
 295             fscal            = felec;
 296
 297             /* Update vectorial force */
 298             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 299             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 300             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 301
 302             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 303             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 304             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 305
 306             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 307
 308             /* Inner loop uses 120 flops */
 309         }
 310
 311         if(jidx<j_index_end)
 312         {
 313
 314             jnrA             = jjnr[jidx];
 315             j_coord_offsetA  = DIM*jnrA;
 316
 317             /* load j atom coordinates */
 318             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 319                                               &jx0,&jy0,&jz0);
 320
 321             /* Calculate displacement vector */
 322             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 323             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 324             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 325             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 326             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 327             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 328             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 329             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 330             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 331
 332             /* Calculate squared distance and things based on it */
 333             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 334             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 335             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 336
 337             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 338             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 339             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 340
 341             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 342             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 343             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 344
 345             /* Load parameters for j particles */
 346             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 347             vdwjidx0A        = 2*vdwtype[jnrA+0];
 348
 349             fjx0             = _fjsp_setzero_v2r8();
 350             fjy0             = _fjsp_setzero_v2r8();
 351             fjz0             = _fjsp_setzero_v2r8();
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             /* Compute parameters for interactions between i and j atoms */
 358             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 359             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 360                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 361
 362             /* REACTION-FIELD ELECTROSTATICS */
 363             velec            = _fjsp_mul_v2r8(qq00,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq00,rinv00),crf));
 364             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 365
 366             /* LENNARD-JONES DISPERSION/REPULSION */
 367
 368             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 369             vvdw6            = _fjsp_mul_v2r8(c6_00,rinvsix);
 370             vvdw12           = _fjsp_mul_v2r8(c12_00,_fjsp_mul_v2r8(rinvsix,rinvsix));
 371             vvdw             = _fjsp_msub_v2r8( vvdw12,one_twelfth, _fjsp_mul_v2r8(vvdw6,one_sixth) );
 372             fvdw             = _fjsp_mul_v2r8(_fjsp_sub_v2r8(vvdw12,vvdw6),rinvsq00);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 376             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 377             vvdw             = _fjsp_unpacklo_v2r8(vvdw,_fjsp_setzero_v2r8());
 378             vvdwsum          = _fjsp_add_v2r8(vvdwsum,vvdw);
 379
 380             fscal            = _fjsp_add_v2r8(felec,fvdw);
 381
 382             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 383
 384             /* Update vectorial force */
 385             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 386             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 387             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 388
 389             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 390             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 391             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 392
 393             /**************************
 394              * CALCULATE INTERACTIONS *
 395              **************************/
 396
 397             /* Compute parameters for interactions between i and j atoms */
 398             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 399
 400             /* REACTION-FIELD ELECTROSTATICS */
 401             velec            = _fjsp_mul_v2r8(qq10,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq10,rinv10),crf));
 402             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 406             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 411
 412             /* Update vectorial force */
 413             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 414             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 415             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 416
 417             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 418             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 419             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* Compute parameters for interactions between i and j atoms */
 426             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 427
 428             /* REACTION-FIELD ELECTROSTATICS */
 429             velec            = _fjsp_mul_v2r8(qq20,_fjsp_sub_v2r8(_fjsp_madd_v2r8(krf,rsq20,rinv20),crf));
 430             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _fjsp_unpacklo_v2r8(velec,_fjsp_setzero_v2r8());
 434             velecsum         = _fjsp_add_v2r8(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 439
 440             /* Update vectorial force */
 441             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 442             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 443             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 444
 445             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 446             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 447             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 448
 449             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 450
 451             /* Inner loop uses 120 flops */
 452         }
 453
 454         /* End of innermost loop */
 455
 456         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 457                                               f+i_coord_offset,fshift+i_shift_offset);
 458
 459         ggid                        = gid[iidx];
 460         /* Update potential energies */
 461         gmx_fjsp_update_1pot_v2r8(velecsum,kernel_data->energygrp_elec+ggid);
 462         gmx_fjsp_update_1pot_v2r8(vvdwsum,kernel_data->energygrp_vdw+ggid);
 463
 464         /* Increment number of inner iterations */
 465         inneriter                  += j_index_end - j_index_start;
 466
 467         /* Outer loop uses 20 flops */
 468     }
 469
 470     /* Increment number of outer iterations */
 471     outeriter        += nri;
 472
 473     /* Update outer/inner flops */
 474
 475     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*120);
 476 }
 477 /*
 478  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 479  * Electrostatics interaction: ReactionField
 480  * VdW interaction:            LennardJones
 481  * Geometry:                   Water3-Particle
 482  * Calculate force/pot:        Force
 483  */
 484 void
 485 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_sparc64_hpc_ace_double
 486                     (t_nblist                    * gmx_restrict       nlist,
 487                      rvec                        * gmx_restrict          xx,
 488                      rvec                        * gmx_restrict          ff,
 489                      t_forcerec                  * gmx_restrict          fr,
 490                      t_mdatoms                   * gmx_restrict     mdatoms,
 491                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 492                      t_nrnb                      * gmx_restrict        nrnb)
 493 {
 494     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 495      * just 0 for non-waters.
 496      * Suffixes A,B refer to j loop unrolling done with double precision SIMD, e.g. for the two different
 497      * jnr indices corresponding to data put in the four positions in the SIMD register.
 498      */
 499     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 500     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 501     int              jnrA,jnrB;
 502     int              j_coord_offsetA,j_coord_offsetB;
 503     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 504     real             rcutoff_scalar;
 505     real             *shiftvec,*fshift,*x,*f;
 506     _fjsp_v2r8       tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 507     int              vdwioffset0;
 508     _fjsp_v2r8       ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 509     int              vdwioffset1;
 510     _fjsp_v2r8       ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 511     int              vdwioffset2;
 512     _fjsp_v2r8       ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 513     int              vdwjidx0A,vdwjidx0B;
 514     _fjsp_v2r8       jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 515     _fjsp_v2r8       dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 516     _fjsp_v2r8       dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 517     _fjsp_v2r8       dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 518     _fjsp_v2r8       velec,felec,velecsum,facel,crf,krf,krf2;
 519     real             *charge;
 520     int              nvdwtype;
 521     _fjsp_v2r8       rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 522     int              *vdwtype;
 523     real             *vdwparam;
 524     _fjsp_v2r8       one_sixth   = gmx_fjsp_set1_v2r8(1.0/6.0);
 525     _fjsp_v2r8       one_twelfth = gmx_fjsp_set1_v2r8(1.0/12.0);
 526     _fjsp_v2r8       itab_tmp;
 527     _fjsp_v2r8       dummy_mask,cutoff_mask;
 528     _fjsp_v2r8       one     = gmx_fjsp_set1_v2r8(1.0);
 529     _fjsp_v2r8       two     = gmx_fjsp_set1_v2r8(2.0);
 530     union { _fjsp_v2r8 simd; long long int i[2]; } vfconv,gbconv,ewconv;
 531
 532     x                = xx[0];
 533     f                = ff[0];
 534
 535     nri              = nlist->nri;
 536     iinr             = nlist->iinr;
 537     jindex           = nlist->jindex;
 538     jjnr             = nlist->jjnr;
 539     shiftidx         = nlist->shift;
 540     gid              = nlist->gid;
 541     shiftvec         = fr->shift_vec[0];
 542     fshift           = fr->fshift[0];
 543     facel            = gmx_fjsp_set1_v2r8(fr->epsfac);
 544     charge           = mdatoms->chargeA;
 545     krf              = gmx_fjsp_set1_v2r8(fr->ic->k_rf);
 546     krf2             = gmx_fjsp_set1_v2r8(fr->ic->k_rf*2.0);
 547     crf              = gmx_fjsp_set1_v2r8(fr->ic->c_rf);
 548     nvdwtype         = fr->ntype;
 549     vdwparam         = fr->nbfp;
 550     vdwtype          = mdatoms->typeA;
 551
 552     /* Setup water-specific parameters */
 553     inr              = nlist->iinr[0];
 554     iq0              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+0]));
 555     iq1              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+1]));
 556     iq2              = _fjsp_mul_v2r8(facel,gmx_fjsp_set1_v2r8(charge[inr+2]));
 557     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 558
 559     /* Avoid stupid compiler warnings */
 560     jnrA = jnrB = 0;
 561     j_coord_offsetA = 0;
 562     j_coord_offsetB = 0;
 563
 564     outeriter        = 0;
 565     inneriter        = 0;
 566
 567     /* Start outer loop over neighborlists */
 568     for(iidx=0; iidx<nri; iidx++)
 569     {
 570         /* Load shift vector for this list */
 571         i_shift_offset   = DIM*shiftidx[iidx];
 572
 573         /* Load limits for loop over neighbors */
 574         j_index_start    = jindex[iidx];
 575         j_index_end      = jindex[iidx+1];
 576
 577         /* Get outer coordinate index */
 578         inr              = iinr[iidx];
 579         i_coord_offset   = DIM*inr;
 580
 581         /* Load i particle coords and add shift vector */
 582         gmx_fjsp_load_shift_and_3rvec_broadcast_v2r8(shiftvec+i_shift_offset,x+i_coord_offset,
 583                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 584
 585         fix0             = _fjsp_setzero_v2r8();
 586         fiy0             = _fjsp_setzero_v2r8();
 587         fiz0             = _fjsp_setzero_v2r8();
 588         fix1             = _fjsp_setzero_v2r8();
 589         fiy1             = _fjsp_setzero_v2r8();
 590         fiz1             = _fjsp_setzero_v2r8();
 591         fix2             = _fjsp_setzero_v2r8();
 592         fiy2             = _fjsp_setzero_v2r8();
 593         fiz2             = _fjsp_setzero_v2r8();
 594
 595         /* Start inner kernel loop */
 596         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 597         {
 598
 599             /* Get j neighbor index, and coordinate index */
 600             jnrA             = jjnr[jidx];
 601             jnrB             = jjnr[jidx+1];
 602             j_coord_offsetA  = DIM*jnrA;
 603             j_coord_offsetB  = DIM*jnrB;
 604
 605             /* load j atom coordinates */
 606             gmx_fjsp_load_1rvec_2ptr_swizzle_v2r8(x+j_coord_offsetA,x+j_coord_offsetB,
 607                                               &jx0,&jy0,&jz0);
 608
 609             /* Calculate displacement vector */
 610             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 611             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 612             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 613             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 614             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 615             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 616             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 617             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 618             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 619
 620             /* Calculate squared distance and things based on it */
 621             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 622             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 623             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 624
 625             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 626             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 627             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 628
 629             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 630             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 631             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 632
 633             /* Load parameters for j particles */
 634             jq0              = gmx_fjsp_load_2real_swizzle_v2r8(charge+jnrA+0,charge+jnrB+0);
 635             vdwjidx0A        = 2*vdwtype[jnrA+0];
 636             vdwjidx0B        = 2*vdwtype[jnrB+0];
 637
 638             fjx0             = _fjsp_setzero_v2r8();
 639             fjy0             = _fjsp_setzero_v2r8();
 640             fjz0             = _fjsp_setzero_v2r8();
 641
 642             /**************************
 643              * CALCULATE INTERACTIONS *
 644              **************************/
 645
 646             /* Compute parameters for interactions between i and j atoms */
 647             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 648             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 649                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 650
 651             /* REACTION-FIELD ELECTROSTATICS */
 652             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 653
 654             /* LENNARD-JONES DISPERSION/REPULSION */
 655
 656             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 657             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 658
 659             fscal            = _fjsp_add_v2r8(felec,fvdw);
 660
 661             /* Update vectorial force */
 662             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 663             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 664             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 665
 666             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 667             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 668             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             /* Compute parameters for interactions between i and j atoms */
 675             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 676
 677             /* REACTION-FIELD ELECTROSTATICS */
 678             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 679
 680             fscal            = felec;
 681
 682             /* Update vectorial force */
 683             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 684             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 685             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 686
 687             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 688             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 689             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 690
 691             /**************************
 692              * CALCULATE INTERACTIONS *
 693              **************************/
 694
 695             /* Compute parameters for interactions between i and j atoms */
 696             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 697
 698             /* REACTION-FIELD ELECTROSTATICS */
 699             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 700
 701             fscal            = felec;
 702
 703             /* Update vectorial force */
 704             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 705             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 706             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 707
 708             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 709             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 710             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 711
 712             gmx_fjsp_decrement_1rvec_2ptr_swizzle_v2r8(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 713
 714             /* Inner loop uses 100 flops */
 715         }
 716
 717         if(jidx<j_index_end)
 718         {
 719
 720             jnrA             = jjnr[jidx];
 721             j_coord_offsetA  = DIM*jnrA;
 722
 723             /* load j atom coordinates */
 724             gmx_fjsp_load_1rvec_1ptr_swizzle_v2r8(x+j_coord_offsetA,
 725                                               &jx0,&jy0,&jz0);
 726
 727             /* Calculate displacement vector */
 728             dx00             = _fjsp_sub_v2r8(ix0,jx0);
 729             dy00             = _fjsp_sub_v2r8(iy0,jy0);
 730             dz00             = _fjsp_sub_v2r8(iz0,jz0);
 731             dx10             = _fjsp_sub_v2r8(ix1,jx0);
 732             dy10             = _fjsp_sub_v2r8(iy1,jy0);
 733             dz10             = _fjsp_sub_v2r8(iz1,jz0);
 734             dx20             = _fjsp_sub_v2r8(ix2,jx0);
 735             dy20             = _fjsp_sub_v2r8(iy2,jy0);
 736             dz20             = _fjsp_sub_v2r8(iz2,jz0);
 737
 738             /* Calculate squared distance and things based on it */
 739             rsq00            = gmx_fjsp_calc_rsq_v2r8(dx00,dy00,dz00);
 740             rsq10            = gmx_fjsp_calc_rsq_v2r8(dx10,dy10,dz10);
 741             rsq20            = gmx_fjsp_calc_rsq_v2r8(dx20,dy20,dz20);
 742
 743             rinv00           = gmx_fjsp_invsqrt_v2r8(rsq00);
 744             rinv10           = gmx_fjsp_invsqrt_v2r8(rsq10);
 745             rinv20           = gmx_fjsp_invsqrt_v2r8(rsq20);
 746
 747             rinvsq00         = _fjsp_mul_v2r8(rinv00,rinv00);
 748             rinvsq10         = _fjsp_mul_v2r8(rinv10,rinv10);
 749             rinvsq20         = _fjsp_mul_v2r8(rinv20,rinv20);
 750
 751             /* Load parameters for j particles */
 752             jq0              = _fjsp_loadl_v2r8(_fjsp_setzero_v2r8(),charge+jnrA+0);
 753             vdwjidx0A        = 2*vdwtype[jnrA+0];
 754
 755             fjx0             = _fjsp_setzero_v2r8();
 756             fjy0             = _fjsp_setzero_v2r8();
 757             fjz0             = _fjsp_setzero_v2r8();
 758
 759             /**************************
 760              * CALCULATE INTERACTIONS *
 761              **************************/
 762
 763             /* Compute parameters for interactions between i and j atoms */
 764             qq00             = _fjsp_mul_v2r8(iq0,jq0);
 765             gmx_fjsp_load_2pair_swizzle_v2r8(vdwparam+vdwioffset0+vdwjidx0A,
 766                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 767
 768             /* REACTION-FIELD ELECTROSTATICS */
 769             felec            = _fjsp_mul_v2r8(qq00,_fjsp_msub_v2r8(rinv00,rinvsq00,krf2));
 770
 771             /* LENNARD-JONES DISPERSION/REPULSION */
 772
 773             rinvsix          = _fjsp_mul_v2r8(_fjsp_mul_v2r8(rinvsq00,rinvsq00),rinvsq00);
 774             fvdw             = _fjsp_mul_v2r8(_fjsp_msub_v2r8(c12_00,rinvsix,c6_00),_fjsp_mul_v2r8(rinvsix,rinvsq00));
 775
 776             fscal            = _fjsp_add_v2r8(felec,fvdw);
 777
 778             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 779
 780             /* Update vectorial force */
 781             fix0             = _fjsp_madd_v2r8(dx00,fscal,fix0);
 782             fiy0             = _fjsp_madd_v2r8(dy00,fscal,fiy0);
 783             fiz0             = _fjsp_madd_v2r8(dz00,fscal,fiz0);
 784
 785             fjx0             = _fjsp_madd_v2r8(dx00,fscal,fjx0);
 786             fjy0             = _fjsp_madd_v2r8(dy00,fscal,fjy0);
 787             fjz0             = _fjsp_madd_v2r8(dz00,fscal,fjz0);
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             /* Compute parameters for interactions between i and j atoms */
 794             qq10             = _fjsp_mul_v2r8(iq1,jq0);
 795
 796             /* REACTION-FIELD ELECTROSTATICS */
 797             felec            = _fjsp_mul_v2r8(qq10,_fjsp_msub_v2r8(rinv10,rinvsq10,krf2));
 798
 799             fscal            = felec;
 800
 801             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 802
 803             /* Update vectorial force */
 804             fix1             = _fjsp_madd_v2r8(dx10,fscal,fix1);
 805             fiy1             = _fjsp_madd_v2r8(dy10,fscal,fiy1);
 806             fiz1             = _fjsp_madd_v2r8(dz10,fscal,fiz1);
 807
 808             fjx0             = _fjsp_madd_v2r8(dx10,fscal,fjx0);
 809             fjy0             = _fjsp_madd_v2r8(dy10,fscal,fjy0);
 810             fjz0             = _fjsp_madd_v2r8(dz10,fscal,fjz0);
 811
 812             /**************************
 813              * CALCULATE INTERACTIONS *
 814              **************************/
 815
 816             /* Compute parameters for interactions between i and j atoms */
 817             qq20             = _fjsp_mul_v2r8(iq2,jq0);
 818
 819             /* REACTION-FIELD ELECTROSTATICS */
 820             felec            = _fjsp_mul_v2r8(qq20,_fjsp_msub_v2r8(rinv20,rinvsq20,krf2));
 821
 822             fscal            = felec;
 823
 824             fscal            = _fjsp_unpacklo_v2r8(fscal,_fjsp_setzero_v2r8());
 825
 826             /* Update vectorial force */
 827             fix2             = _fjsp_madd_v2r8(dx20,fscal,fix2);
 828             fiy2             = _fjsp_madd_v2r8(dy20,fscal,fiy2);
 829             fiz2             = _fjsp_madd_v2r8(dz20,fscal,fiz2);
 830
 831             fjx0             = _fjsp_madd_v2r8(dx20,fscal,fjx0);
 832             fjy0             = _fjsp_madd_v2r8(dy20,fscal,fjy0);
 833             fjz0             = _fjsp_madd_v2r8(dz20,fscal,fjz0);
 834
 835             gmx_fjsp_decrement_1rvec_1ptr_swizzle_v2r8(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 836
 837             /* Inner loop uses 100 flops */
 838         }
 839
 840         /* End of innermost loop */
 841
 842         gmx_fjsp_update_iforce_3atom_swizzle_v2r8(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 843                                               f+i_coord_offset,fshift+i_shift_offset);
 844
 845         /* Increment number of inner iterations */
 846         inneriter                  += j_index_end - j_index_start;
 847
 848         /* Outer loop uses 18 flops */
 849     }
 850
 851     /* Increment number of outer iterations */
 852     outeriter        += nri;
 853
 854     /* Update outer/inner flops */
 855
 856     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*100);
 857 }