src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecGB_VdwLJ_GeomP1P1_avx_128_fma_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 avx_128_fma_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_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_double
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_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              vdwjidx0A,vdwjidx0B;
  85     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128i          gbitab;
  90     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
  91     __m128d          minushalf = _mm_set1_pd(-0.5);
  92     real             *invsqrta,*dvda,*gbtab;
  93     int              nvdwtype;
  94     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  98     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  99     __m128i          vfitab;
 100     __m128i          ifour       = _mm_set1_epi32(4);
 101     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 102     real             *vftab;
 103     __m128d          dummy_mask,cutoff_mask;
 104     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 105     __m128d          one     = _mm_set1_pd(1.0);
 106     __m128d          two     = _mm_set1_pd(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_pd(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123
 124     invsqrta         = fr->invsqrta;
 125     dvda             = fr->dvda;
 126     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 127     gbtab            = fr->gbtab.data;
 128     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 129
 130     /* Avoid stupid compiler warnings */
 131     jnrA = jnrB = 0;
 132     j_coord_offsetA = 0;
 133     j_coord_offsetB = 0;
 134
 135     outeriter        = 0;
 136     inneriter        = 0;
 137
 138     /* Start outer loop over neighborlists */
 139     for(iidx=0; iidx<nri; iidx++)
 140     {
 141         /* Load shift vector for this list */
 142         i_shift_offset   = DIM*shiftidx[iidx];
 143
 144         /* Load limits for loop over neighbors */
 145         j_index_start    = jindex[iidx];
 146         j_index_end      = jindex[iidx+1];
 147
 148         /* Get outer coordinate index */
 149         inr              = iinr[iidx];
 150         i_coord_offset   = DIM*inr;
 151
 152         /* Load i particle coords and add shift vector */
 153         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 154
 155         fix0             = _mm_setzero_pd();
 156         fiy0             = _mm_setzero_pd();
 157         fiz0             = _mm_setzero_pd();
 158
 159         /* Load parameters for i particles */
 160         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 161         isai0            = _mm_load1_pd(invsqrta+inr+0);
 162         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_pd();
 166         vgbsum           = _mm_setzero_pd();
 167         vvdwsum          = _mm_setzero_pd();
 168         dvdasum          = _mm_setzero_pd();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179
 180             /* load j atom coordinates */
 181             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                               &jx0,&jy0,&jz0);
 183
 184             /* Calculate displacement vector */
 185             dx00             = _mm_sub_pd(ix0,jx0);
 186             dy00             = _mm_sub_pd(iy0,jy0);
 187             dz00             = _mm_sub_pd(iz0,jz0);
 188
 189             /* Calculate squared distance and things based on it */
 190             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 191
 192             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 193
 194             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 195
 196             /* Load parameters for j particles */
 197             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 198             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 199             vdwjidx0A        = 2*vdwtype[jnrA+0];
 200             vdwjidx0B        = 2*vdwtype[jnrB+0];
 201
 202             /**************************
 203              * CALCULATE INTERACTIONS *
 204              **************************/
 205
 206             r00              = _mm_mul_pd(rsq00,rinv00);
 207
 208             /* Compute parameters for interactions between i and j atoms */
 209             qq00             = _mm_mul_pd(iq0,jq0);
 210             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 211                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 212
 213             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 214             isaprod          = _mm_mul_pd(isai0,isaj0);
 215             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 216             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 217
 218             /* Calculate generalized born table index - this is a separate table from the normal one,
 219              * but we use the same procedure by multiplying r with scale and truncating to integer.
 220              */
 221             rt               = _mm_mul_pd(r00,gbscale);
 222             gbitab           = _mm_cvttpd_epi32(rt);
 223 #ifdef __XOP__
 224             gbeps            = _mm_frcz_pd(rt);
 225 #else
 226             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 227 #endif
 228             gbitab           = _mm_slli_epi32(gbitab,2);
 229
 230             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 231             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 232             GMX_MM_TRANSPOSE2_PD(Y,F);
 233             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 234             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 235             GMX_MM_TRANSPOSE2_PD(G,H);
 236             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 237             VV               = _mm_macc_pd(gbeps,Fp,Y);
 238             vgb              = _mm_mul_pd(gbqqfactor,VV);
 239
 240             twogbeps         = _mm_add_pd(gbeps,gbeps);
 241             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 242             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 243             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 244             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 245             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 246             velec            = _mm_mul_pd(qq00,rinv00);
 247             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 248
 249             /* LENNARD-JONES DISPERSION/REPULSION */
 250
 251             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 252             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 253             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 254             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 255             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 256
 257             /* Update potential sum for this i atom from the interaction with this j atom. */
 258             velecsum         = _mm_add_pd(velecsum,velec);
 259             vgbsum           = _mm_add_pd(vgbsum,vgb);
 260             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 261
 262             fscal            = _mm_add_pd(felec,fvdw);
 263
 264             /* Update vectorial force */
 265             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 266             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 267             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 268
 269             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 270                                                    _mm_mul_pd(dx00,fscal),
 271                                                    _mm_mul_pd(dy00,fscal),
 272                                                    _mm_mul_pd(dz00,fscal));
 273
 274             /* Inner loop uses 74 flops */
 275         }
 276
 277         if(jidx<j_index_end)
 278         {
 279
 280             jnrA             = jjnr[jidx];
 281             j_coord_offsetA  = DIM*jnrA;
 282
 283             /* load j atom coordinates */
 284             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 285                                               &jx0,&jy0,&jz0);
 286
 287             /* Calculate displacement vector */
 288             dx00             = _mm_sub_pd(ix0,jx0);
 289             dy00             = _mm_sub_pd(iy0,jy0);
 290             dz00             = _mm_sub_pd(iz0,jz0);
 291
 292             /* Calculate squared distance and things based on it */
 293             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 294
 295             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 296
 297             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 298
 299             /* Load parameters for j particles */
 300             jq0              = _mm_load_sd(charge+jnrA+0);
 301             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 302             vdwjidx0A        = 2*vdwtype[jnrA+0];
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             r00              = _mm_mul_pd(rsq00,rinv00);
 309
 310             /* Compute parameters for interactions between i and j atoms */
 311             qq00             = _mm_mul_pd(iq0,jq0);
 312             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 313
 314             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 315             isaprod          = _mm_mul_pd(isai0,isaj0);
 316             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 317             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 318
 319             /* Calculate generalized born table index - this is a separate table from the normal one,
 320              * but we use the same procedure by multiplying r with scale and truncating to integer.
 321              */
 322             rt               = _mm_mul_pd(r00,gbscale);
 323             gbitab           = _mm_cvttpd_epi32(rt);
 324 #ifdef __XOP__
 325             gbeps            = _mm_frcz_pd(rt);
 326 #else
 327             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 328 #endif
 329             gbitab           = _mm_slli_epi32(gbitab,2);
 330
 331             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 332             F                = _mm_setzero_pd();
 333             GMX_MM_TRANSPOSE2_PD(Y,F);
 334             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 335             H                = _mm_setzero_pd();
 336             GMX_MM_TRANSPOSE2_PD(G,H);
 337             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 338             VV               = _mm_macc_pd(gbeps,Fp,Y);
 339             vgb              = _mm_mul_pd(gbqqfactor,VV);
 340
 341             twogbeps         = _mm_add_pd(gbeps,gbeps);
 342             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 343             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 344             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 345             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 346             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 347             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 348             velec            = _mm_mul_pd(qq00,rinv00);
 349             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 350
 351             /* LENNARD-JONES DISPERSION/REPULSION */
 352
 353             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 354             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 355             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 356             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 357             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 361             velecsum         = _mm_add_pd(velecsum,velec);
 362             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 363             vgbsum           = _mm_add_pd(vgbsum,vgb);
 364             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 365             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 366
 367             fscal            = _mm_add_pd(felec,fvdw);
 368
 369             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 370
 371             /* Update vectorial force */
 372             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 373             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 374             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 375
 376             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 377                                                    _mm_mul_pd(dx00,fscal),
 378                                                    _mm_mul_pd(dy00,fscal),
 379                                                    _mm_mul_pd(dz00,fscal));
 380
 381             /* Inner loop uses 74 flops */
 382         }
 383
 384         /* End of innermost loop */
 385
 386         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 387                                               f+i_coord_offset,fshift+i_shift_offset);
 388
 389         ggid                        = gid[iidx];
 390         /* Update potential energies */
 391         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 392         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 393         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 394         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 395         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 396
 397         /* Increment number of inner iterations */
 398         inneriter                  += j_index_end - j_index_start;
 399
 400         /* Outer loop uses 10 flops */
 401     }
 402
 403     /* Increment number of outer iterations */
 404     outeriter        += nri;
 405
 406     /* Update outer/inner flops */
 407
 408     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*74);
 409 }
 410 /*
 411  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_double
 412  * Electrostatics interaction: GeneralizedBorn
 413  * VdW interaction:            LennardJones
 414  * Geometry:                   Particle-Particle
 415  * Calculate force/pot:        Force
 416  */
 417 void
 418 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_double
 419                     (t_nblist                    * gmx_restrict       nlist,
 420                      rvec                        * gmx_restrict          xx,
 421                      rvec                        * gmx_restrict          ff,
 422                      t_forcerec                  * gmx_restrict          fr,
 423                      t_mdatoms                   * gmx_restrict     mdatoms,
 424                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 425                      t_nrnb                      * gmx_restrict        nrnb)
 426 {
 427     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 428      * just 0 for non-waters.
 429      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 430      * jnr indices corresponding to data put in the four positions in the SIMD register.
 431      */
 432     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 433     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 434     int              jnrA,jnrB;
 435     int              j_coord_offsetA,j_coord_offsetB;
 436     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 437     real             rcutoff_scalar;
 438     real             *shiftvec,*fshift,*x,*f;
 439     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 440     int              vdwioffset0;
 441     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 442     int              vdwjidx0A,vdwjidx0B;
 443     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 444     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 445     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 446     real             *charge;
 447     __m128i          gbitab;
 448     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 449     __m128d          minushalf = _mm_set1_pd(-0.5);
 450     real             *invsqrta,*dvda,*gbtab;
 451     int              nvdwtype;
 452     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 453     int              *vdwtype;
 454     real             *vdwparam;
 455     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 456     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 457     __m128i          vfitab;
 458     __m128i          ifour       = _mm_set1_epi32(4);
 459     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 460     real             *vftab;
 461     __m128d          dummy_mask,cutoff_mask;
 462     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 463     __m128d          one     = _mm_set1_pd(1.0);
 464     __m128d          two     = _mm_set1_pd(2.0);
 465     x                = xx[0];
 466     f                = ff[0];
 467
 468     nri              = nlist->nri;
 469     iinr             = nlist->iinr;
 470     jindex           = nlist->jindex;
 471     jjnr             = nlist->jjnr;
 472     shiftidx         = nlist->shift;
 473     gid              = nlist->gid;
 474     shiftvec         = fr->shift_vec[0];
 475     fshift           = fr->fshift[0];
 476     facel            = _mm_set1_pd(fr->epsfac);
 477     charge           = mdatoms->chargeA;
 478     nvdwtype         = fr->ntype;
 479     vdwparam         = fr->nbfp;
 480     vdwtype          = mdatoms->typeA;
 481
 482     invsqrta         = fr->invsqrta;
 483     dvda             = fr->dvda;
 484     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 485     gbtab            = fr->gbtab.data;
 486     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 487
 488     /* Avoid stupid compiler warnings */
 489     jnrA = jnrB = 0;
 490     j_coord_offsetA = 0;
 491     j_coord_offsetB = 0;
 492
 493     outeriter        = 0;
 494     inneriter        = 0;
 495
 496     /* Start outer loop over neighborlists */
 497     for(iidx=0; iidx<nri; iidx++)
 498     {
 499         /* Load shift vector for this list */
 500         i_shift_offset   = DIM*shiftidx[iidx];
 501
 502         /* Load limits for loop over neighbors */
 503         j_index_start    = jindex[iidx];
 504         j_index_end      = jindex[iidx+1];
 505
 506         /* Get outer coordinate index */
 507         inr              = iinr[iidx];
 508         i_coord_offset   = DIM*inr;
 509
 510         /* Load i particle coords and add shift vector */
 511         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 512
 513         fix0             = _mm_setzero_pd();
 514         fiy0             = _mm_setzero_pd();
 515         fiz0             = _mm_setzero_pd();
 516
 517         /* Load parameters for i particles */
 518         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 519         isai0            = _mm_load1_pd(invsqrta+inr+0);
 520         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 521
 522         dvdasum          = _mm_setzero_pd();
 523
 524         /* Start inner kernel loop */
 525         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 526         {
 527
 528             /* Get j neighbor index, and coordinate index */
 529             jnrA             = jjnr[jidx];
 530             jnrB             = jjnr[jidx+1];
 531             j_coord_offsetA  = DIM*jnrA;
 532             j_coord_offsetB  = DIM*jnrB;
 533
 534             /* load j atom coordinates */
 535             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 536                                               &jx0,&jy0,&jz0);
 537
 538             /* Calculate displacement vector */
 539             dx00             = _mm_sub_pd(ix0,jx0);
 540             dy00             = _mm_sub_pd(iy0,jy0);
 541             dz00             = _mm_sub_pd(iz0,jz0);
 542
 543             /* Calculate squared distance and things based on it */
 544             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 545
 546             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 547
 548             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 549
 550             /* Load parameters for j particles */
 551             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 552             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 553             vdwjidx0A        = 2*vdwtype[jnrA+0];
 554             vdwjidx0B        = 2*vdwtype[jnrB+0];
 555
 556             /**************************
 557              * CALCULATE INTERACTIONS *
 558              **************************/
 559
 560             r00              = _mm_mul_pd(rsq00,rinv00);
 561
 562             /* Compute parameters for interactions between i and j atoms */
 563             qq00             = _mm_mul_pd(iq0,jq0);
 564             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 565                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 566
 567             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 568             isaprod          = _mm_mul_pd(isai0,isaj0);
 569             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 570             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 571
 572             /* Calculate generalized born table index - this is a separate table from the normal one,
 573              * but we use the same procedure by multiplying r with scale and truncating to integer.
 574              */
 575             rt               = _mm_mul_pd(r00,gbscale);
 576             gbitab           = _mm_cvttpd_epi32(rt);
 577 #ifdef __XOP__
 578             gbeps            = _mm_frcz_pd(rt);
 579 #else
 580             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 581 #endif
 582             gbitab           = _mm_slli_epi32(gbitab,2);
 583
 584             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 585             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 586             GMX_MM_TRANSPOSE2_PD(Y,F);
 587             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 588             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 589             GMX_MM_TRANSPOSE2_PD(G,H);
 590             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 591             VV               = _mm_macc_pd(gbeps,Fp,Y);
 592             vgb              = _mm_mul_pd(gbqqfactor,VV);
 593
 594             twogbeps         = _mm_add_pd(gbeps,gbeps);
 595             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 596             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 597             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 598             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 599             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 600             velec            = _mm_mul_pd(qq00,rinv00);
 601             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 602
 603             /* LENNARD-JONES DISPERSION/REPULSION */
 604
 605             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 606             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 607
 608             fscal            = _mm_add_pd(felec,fvdw);
 609
 610             /* Update vectorial force */
 611             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 612             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 613             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 614
 615             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 616                                                    _mm_mul_pd(dx00,fscal),
 617                                                    _mm_mul_pd(dy00,fscal),
 618                                                    _mm_mul_pd(dz00,fscal));
 619
 620             /* Inner loop uses 67 flops */
 621         }
 622
 623         if(jidx<j_index_end)
 624         {
 625
 626             jnrA             = jjnr[jidx];
 627             j_coord_offsetA  = DIM*jnrA;
 628
 629             /* load j atom coordinates */
 630             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 631                                               &jx0,&jy0,&jz0);
 632
 633             /* Calculate displacement vector */
 634             dx00             = _mm_sub_pd(ix0,jx0);
 635             dy00             = _mm_sub_pd(iy0,jy0);
 636             dz00             = _mm_sub_pd(iz0,jz0);
 637
 638             /* Calculate squared distance and things based on it */
 639             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 640
 641             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 642
 643             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 644
 645             /* Load parameters for j particles */
 646             jq0              = _mm_load_sd(charge+jnrA+0);
 647             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 648             vdwjidx0A        = 2*vdwtype[jnrA+0];
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             r00              = _mm_mul_pd(rsq00,rinv00);
 655
 656             /* Compute parameters for interactions between i and j atoms */
 657             qq00             = _mm_mul_pd(iq0,jq0);
 658             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 659
 660             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 661             isaprod          = _mm_mul_pd(isai0,isaj0);
 662             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 663             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 664
 665             /* Calculate generalized born table index - this is a separate table from the normal one,
 666              * but we use the same procedure by multiplying r with scale and truncating to integer.
 667              */
 668             rt               = _mm_mul_pd(r00,gbscale);
 669             gbitab           = _mm_cvttpd_epi32(rt);
 670 #ifdef __XOP__
 671             gbeps            = _mm_frcz_pd(rt);
 672 #else
 673             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 674 #endif
 675             gbitab           = _mm_slli_epi32(gbitab,2);
 676
 677             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 678             F                = _mm_setzero_pd();
 679             GMX_MM_TRANSPOSE2_PD(Y,F);
 680             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 681             H                = _mm_setzero_pd();
 682             GMX_MM_TRANSPOSE2_PD(G,H);
 683             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 684             VV               = _mm_macc_pd(gbeps,Fp,Y);
 685             vgb              = _mm_mul_pd(gbqqfactor,VV);
 686
 687             twogbeps         = _mm_add_pd(gbeps,gbeps);
 688             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 689             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 690             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 691             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 692             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 693             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 694             velec            = _mm_mul_pd(qq00,rinv00);
 695             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 696
 697             /* LENNARD-JONES DISPERSION/REPULSION */
 698
 699             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 700             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 701
 702             fscal            = _mm_add_pd(felec,fvdw);
 703
 704             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 705
 706             /* Update vectorial force */
 707             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 708             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 709             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 710
 711             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 712                                                    _mm_mul_pd(dx00,fscal),
 713                                                    _mm_mul_pd(dy00,fscal),
 714                                                    _mm_mul_pd(dz00,fscal));
 715
 716             /* Inner loop uses 67 flops */
 717         }
 718
 719         /* End of innermost loop */
 720
 721         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 722                                               f+i_coord_offset,fshift+i_shift_offset);
 723
 724         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 725         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 726
 727         /* Increment number of inner iterations */
 728         inneriter                  += j_index_end - j_index_start;
 729
 730         /* Outer loop uses 7 flops */
 731     }
 732
 733     /* Increment number of outer iterations */
 734     outeriter        += nri;
 735
 736     /* Update outer/inner flops */
 737
 738     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*67);
 739 }