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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_double
  52  * Electrostatics interaction: GeneralizedBorn
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwjidx0A,vdwjidx0B;
  83     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  84     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  85     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  86     real             *charge;
  87     __m128i          gbitab;
  88     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
  89     __m128d          minushalf = _mm_set1_pd(-0.5);
  90     real             *invsqrta,*dvda,*gbtab;
  91     int              nvdwtype;
  92     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  96     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  97     __m128i          vfitab;
  98     __m128i          ifour       = _mm_set1_epi32(4);
  99     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 100     real             *vftab;
 101     __m128d          dummy_mask,cutoff_mask;
 102     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 103     __m128d          one     = _mm_set1_pd(1.0);
 104     __m128d          two     = _mm_set1_pd(2.0);
 105     x                = xx[0];
 106     f                = ff[0];
 107
 108     nri              = nlist->nri;
 109     iinr             = nlist->iinr;
 110     jindex           = nlist->jindex;
 111     jjnr             = nlist->jjnr;
 112     shiftidx         = nlist->shift;
 113     gid              = nlist->gid;
 114     shiftvec         = fr->shift_vec[0];
 115     fshift           = fr->fshift[0];
 116     facel            = _mm_set1_pd(fr->epsfac);
 117     charge           = mdatoms->chargeA;
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     invsqrta         = fr->invsqrta;
 123     dvda             = fr->dvda;
 124     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 125     gbtab            = fr->gbtab.data;
 126     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141
 142         /* Load limits for loop over neighbors */
 143         j_index_start    = jindex[iidx];
 144         j_index_end      = jindex[iidx+1];
 145
 146         /* Get outer coordinate index */
 147         inr              = iinr[iidx];
 148         i_coord_offset   = DIM*inr;
 149
 150         /* Load i particle coords and add shift vector */
 151         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 152
 153         fix0             = _mm_setzero_pd();
 154         fiy0             = _mm_setzero_pd();
 155         fiz0             = _mm_setzero_pd();
 156
 157         /* Load parameters for i particles */
 158         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 159         isai0            = _mm_load1_pd(invsqrta+inr+0);
 160         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 161
 162         /* Reset potential sums */
 163         velecsum         = _mm_setzero_pd();
 164         vgbsum           = _mm_setzero_pd();
 165         vvdwsum          = _mm_setzero_pd();
 166         dvdasum          = _mm_setzero_pd();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177
 178             /* load j atom coordinates */
 179             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                               &jx0,&jy0,&jz0);
 181
 182             /* Calculate displacement vector */
 183             dx00             = _mm_sub_pd(ix0,jx0);
 184             dy00             = _mm_sub_pd(iy0,jy0);
 185             dz00             = _mm_sub_pd(iz0,jz0);
 186
 187             /* Calculate squared distance and things based on it */
 188             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 189
 190             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 191
 192             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 193
 194             /* Load parameters for j particles */
 195             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 196             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 197             vdwjidx0A        = 2*vdwtype[jnrA+0];
 198             vdwjidx0B        = 2*vdwtype[jnrB+0];
 199
 200             /**************************
 201              * CALCULATE INTERACTIONS *
 202              **************************/
 203
 204             r00              = _mm_mul_pd(rsq00,rinv00);
 205
 206             /* Compute parameters for interactions between i and j atoms */
 207             qq00             = _mm_mul_pd(iq0,jq0);
 208             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 209                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 210
 211             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 212             isaprod          = _mm_mul_pd(isai0,isaj0);
 213             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 214             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 215
 216             /* Calculate generalized born table index - this is a separate table from the normal one,
 217              * but we use the same procedure by multiplying r with scale and truncating to integer.
 218              */
 219             rt               = _mm_mul_pd(r00,gbscale);
 220             gbitab           = _mm_cvttpd_epi32(rt);
 221 #ifdef __XOP__
 222             gbeps            = _mm_frcz_pd(rt);
 223 #else
 224             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 225 #endif
 226             gbitab           = _mm_slli_epi32(gbitab,2);
 227
 228             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 229             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 230             GMX_MM_TRANSPOSE2_PD(Y,F);
 231             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 232             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 233             GMX_MM_TRANSPOSE2_PD(G,H);
 234             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 235             VV               = _mm_macc_pd(gbeps,Fp,Y);
 236             vgb              = _mm_mul_pd(gbqqfactor,VV);
 237
 238             twogbeps         = _mm_add_pd(gbeps,gbeps);
 239             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 240             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 241             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 242             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 243             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 244             velec            = _mm_mul_pd(qq00,rinv00);
 245             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 246
 247             /* LENNARD-JONES DISPERSION/REPULSION */
 248
 249             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 250             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 251             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 252             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 253             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 254
 255             /* Update potential sum for this i atom from the interaction with this j atom. */
 256             velecsum         = _mm_add_pd(velecsum,velec);
 257             vgbsum           = _mm_add_pd(vgbsum,vgb);
 258             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 259
 260             fscal            = _mm_add_pd(felec,fvdw);
 261
 262             /* Update vectorial force */
 263             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 264             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 265             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 266
 267             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 268                                                    _mm_mul_pd(dx00,fscal),
 269                                                    _mm_mul_pd(dy00,fscal),
 270                                                    _mm_mul_pd(dz00,fscal));
 271
 272             /* Inner loop uses 74 flops */
 273         }
 274
 275         if(jidx<j_index_end)
 276         {
 277
 278             jnrA             = jjnr[jidx];
 279             j_coord_offsetA  = DIM*jnrA;
 280
 281             /* load j atom coordinates */
 282             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 283                                               &jx0,&jy0,&jz0);
 284
 285             /* Calculate displacement vector */
 286             dx00             = _mm_sub_pd(ix0,jx0);
 287             dy00             = _mm_sub_pd(iy0,jy0);
 288             dz00             = _mm_sub_pd(iz0,jz0);
 289
 290             /* Calculate squared distance and things based on it */
 291             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 292
 293             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 294
 295             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 296
 297             /* Load parameters for j particles */
 298             jq0              = _mm_load_sd(charge+jnrA+0);
 299             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 300             vdwjidx0A        = 2*vdwtype[jnrA+0];
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             r00              = _mm_mul_pd(rsq00,rinv00);
 307
 308             /* Compute parameters for interactions between i and j atoms */
 309             qq00             = _mm_mul_pd(iq0,jq0);
 310             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 311
 312             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 313             isaprod          = _mm_mul_pd(isai0,isaj0);
 314             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 315             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 316
 317             /* Calculate generalized born table index - this is a separate table from the normal one,
 318              * but we use the same procedure by multiplying r with scale and truncating to integer.
 319              */
 320             rt               = _mm_mul_pd(r00,gbscale);
 321             gbitab           = _mm_cvttpd_epi32(rt);
 322 #ifdef __XOP__
 323             gbeps            = _mm_frcz_pd(rt);
 324 #else
 325             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 326 #endif
 327             gbitab           = _mm_slli_epi32(gbitab,2);
 328
 329             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 330             F                = _mm_setzero_pd();
 331             GMX_MM_TRANSPOSE2_PD(Y,F);
 332             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 333             H                = _mm_setzero_pd();
 334             GMX_MM_TRANSPOSE2_PD(G,H);
 335             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 336             VV               = _mm_macc_pd(gbeps,Fp,Y);
 337             vgb              = _mm_mul_pd(gbqqfactor,VV);
 338
 339             twogbeps         = _mm_add_pd(gbeps,gbeps);
 340             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 341             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 342             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 343             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 344             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 345             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 346             velec            = _mm_mul_pd(qq00,rinv00);
 347             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 348
 349             /* LENNARD-JONES DISPERSION/REPULSION */
 350
 351             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 352             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 353             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 354             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 355             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 359             velecsum         = _mm_add_pd(velecsum,velec);
 360             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 361             vgbsum           = _mm_add_pd(vgbsum,vgb);
 362             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 363             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 364
 365             fscal            = _mm_add_pd(felec,fvdw);
 366
 367             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 368
 369             /* Update vectorial force */
 370             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 371             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 372             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 373
 374             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 375                                                    _mm_mul_pd(dx00,fscal),
 376                                                    _mm_mul_pd(dy00,fscal),
 377                                                    _mm_mul_pd(dz00,fscal));
 378
 379             /* Inner loop uses 74 flops */
 380         }
 381
 382         /* End of innermost loop */
 383
 384         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 385                                               f+i_coord_offset,fshift+i_shift_offset);
 386
 387         ggid                        = gid[iidx];
 388         /* Update potential energies */
 389         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 390         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 391         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 392         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 393         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 394
 395         /* Increment number of inner iterations */
 396         inneriter                  += j_index_end - j_index_start;
 397
 398         /* Outer loop uses 10 flops */
 399     }
 400
 401     /* Increment number of outer iterations */
 402     outeriter        += nri;
 403
 404     /* Update outer/inner flops */
 405
 406     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*74);
 407 }
 408 /*
 409  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_double
 410  * Electrostatics interaction: GeneralizedBorn
 411  * VdW interaction:            LennardJones
 412  * Geometry:                   Particle-Particle
 413  * Calculate force/pot:        Force
 414  */
 415 void
 416 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_double
 417                     (t_nblist                    * gmx_restrict       nlist,
 418                      rvec                        * gmx_restrict          xx,
 419                      rvec                        * gmx_restrict          ff,
 420                      t_forcerec                  * gmx_restrict          fr,
 421                      t_mdatoms                   * gmx_restrict     mdatoms,
 422                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 423                      t_nrnb                      * gmx_restrict        nrnb)
 424 {
 425     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 426      * just 0 for non-waters.
 427      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 428      * jnr indices corresponding to data put in the four positions in the SIMD register.
 429      */
 430     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 431     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 432     int              jnrA,jnrB;
 433     int              j_coord_offsetA,j_coord_offsetB;
 434     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 435     real             rcutoff_scalar;
 436     real             *shiftvec,*fshift,*x,*f;
 437     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 438     int              vdwioffset0;
 439     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 440     int              vdwjidx0A,vdwjidx0B;
 441     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 442     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 443     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 444     real             *charge;
 445     __m128i          gbitab;
 446     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 447     __m128d          minushalf = _mm_set1_pd(-0.5);
 448     real             *invsqrta,*dvda,*gbtab;
 449     int              nvdwtype;
 450     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 451     int              *vdwtype;
 452     real             *vdwparam;
 453     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 454     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 455     __m128i          vfitab;
 456     __m128i          ifour       = _mm_set1_epi32(4);
 457     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 458     real             *vftab;
 459     __m128d          dummy_mask,cutoff_mask;
 460     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 461     __m128d          one     = _mm_set1_pd(1.0);
 462     __m128d          two     = _mm_set1_pd(2.0);
 463     x                = xx[0];
 464     f                = ff[0];
 465
 466     nri              = nlist->nri;
 467     iinr             = nlist->iinr;
 468     jindex           = nlist->jindex;
 469     jjnr             = nlist->jjnr;
 470     shiftidx         = nlist->shift;
 471     gid              = nlist->gid;
 472     shiftvec         = fr->shift_vec[0];
 473     fshift           = fr->fshift[0];
 474     facel            = _mm_set1_pd(fr->epsfac);
 475     charge           = mdatoms->chargeA;
 476     nvdwtype         = fr->ntype;
 477     vdwparam         = fr->nbfp;
 478     vdwtype          = mdatoms->typeA;
 479
 480     invsqrta         = fr->invsqrta;
 481     dvda             = fr->dvda;
 482     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 483     gbtab            = fr->gbtab.data;
 484     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 485
 486     /* Avoid stupid compiler warnings */
 487     jnrA = jnrB = 0;
 488     j_coord_offsetA = 0;
 489     j_coord_offsetB = 0;
 490
 491     outeriter        = 0;
 492     inneriter        = 0;
 493
 494     /* Start outer loop over neighborlists */
 495     for(iidx=0; iidx<nri; iidx++)
 496     {
 497         /* Load shift vector for this list */
 498         i_shift_offset   = DIM*shiftidx[iidx];
 499
 500         /* Load limits for loop over neighbors */
 501         j_index_start    = jindex[iidx];
 502         j_index_end      = jindex[iidx+1];
 503
 504         /* Get outer coordinate index */
 505         inr              = iinr[iidx];
 506         i_coord_offset   = DIM*inr;
 507
 508         /* Load i particle coords and add shift vector */
 509         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 510
 511         fix0             = _mm_setzero_pd();
 512         fiy0             = _mm_setzero_pd();
 513         fiz0             = _mm_setzero_pd();
 514
 515         /* Load parameters for i particles */
 516         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 517         isai0            = _mm_load1_pd(invsqrta+inr+0);
 518         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 519
 520         dvdasum          = _mm_setzero_pd();
 521
 522         /* Start inner kernel loop */
 523         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 524         {
 525
 526             /* Get j neighbor index, and coordinate index */
 527             jnrA             = jjnr[jidx];
 528             jnrB             = jjnr[jidx+1];
 529             j_coord_offsetA  = DIM*jnrA;
 530             j_coord_offsetB  = DIM*jnrB;
 531
 532             /* load j atom coordinates */
 533             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 534                                               &jx0,&jy0,&jz0);
 535
 536             /* Calculate displacement vector */
 537             dx00             = _mm_sub_pd(ix0,jx0);
 538             dy00             = _mm_sub_pd(iy0,jy0);
 539             dz00             = _mm_sub_pd(iz0,jz0);
 540
 541             /* Calculate squared distance and things based on it */
 542             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 543
 544             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 545
 546             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 547
 548             /* Load parameters for j particles */
 549             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 550             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 551             vdwjidx0A        = 2*vdwtype[jnrA+0];
 552             vdwjidx0B        = 2*vdwtype[jnrB+0];
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             r00              = _mm_mul_pd(rsq00,rinv00);
 559
 560             /* Compute parameters for interactions between i and j atoms */
 561             qq00             = _mm_mul_pd(iq0,jq0);
 562             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 563                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 564
 565             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 566             isaprod          = _mm_mul_pd(isai0,isaj0);
 567             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 568             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 569
 570             /* Calculate generalized born table index - this is a separate table from the normal one,
 571              * but we use the same procedure by multiplying r with scale and truncating to integer.
 572              */
 573             rt               = _mm_mul_pd(r00,gbscale);
 574             gbitab           = _mm_cvttpd_epi32(rt);
 575 #ifdef __XOP__
 576             gbeps            = _mm_frcz_pd(rt);
 577 #else
 578             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 579 #endif
 580             gbitab           = _mm_slli_epi32(gbitab,2);
 581
 582             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 583             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 584             GMX_MM_TRANSPOSE2_PD(Y,F);
 585             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 586             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 587             GMX_MM_TRANSPOSE2_PD(G,H);
 588             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 589             VV               = _mm_macc_pd(gbeps,Fp,Y);
 590             vgb              = _mm_mul_pd(gbqqfactor,VV);
 591
 592             twogbeps         = _mm_add_pd(gbeps,gbeps);
 593             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 594             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 595             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 596             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 597             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 598             velec            = _mm_mul_pd(qq00,rinv00);
 599             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 600
 601             /* LENNARD-JONES DISPERSION/REPULSION */
 602
 603             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 604             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 605
 606             fscal            = _mm_add_pd(felec,fvdw);
 607
 608             /* Update vectorial force */
 609             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 610             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 611             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 612
 613             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 614                                                    _mm_mul_pd(dx00,fscal),
 615                                                    _mm_mul_pd(dy00,fscal),
 616                                                    _mm_mul_pd(dz00,fscal));
 617
 618             /* Inner loop uses 67 flops */
 619         }
 620
 621         if(jidx<j_index_end)
 622         {
 623
 624             jnrA             = jjnr[jidx];
 625             j_coord_offsetA  = DIM*jnrA;
 626
 627             /* load j atom coordinates */
 628             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 629                                               &jx0,&jy0,&jz0);
 630
 631             /* Calculate displacement vector */
 632             dx00             = _mm_sub_pd(ix0,jx0);
 633             dy00             = _mm_sub_pd(iy0,jy0);
 634             dz00             = _mm_sub_pd(iz0,jz0);
 635
 636             /* Calculate squared distance and things based on it */
 637             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 638
 639             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 640
 641             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 642
 643             /* Load parameters for j particles */
 644             jq0              = _mm_load_sd(charge+jnrA+0);
 645             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 646             vdwjidx0A        = 2*vdwtype[jnrA+0];
 647
 648             /**************************
 649              * CALCULATE INTERACTIONS *
 650              **************************/
 651
 652             r00              = _mm_mul_pd(rsq00,rinv00);
 653
 654             /* Compute parameters for interactions between i and j atoms */
 655             qq00             = _mm_mul_pd(iq0,jq0);
 656             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 657
 658             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 659             isaprod          = _mm_mul_pd(isai0,isaj0);
 660             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 661             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 662
 663             /* Calculate generalized born table index - this is a separate table from the normal one,
 664              * but we use the same procedure by multiplying r with scale and truncating to integer.
 665              */
 666             rt               = _mm_mul_pd(r00,gbscale);
 667             gbitab           = _mm_cvttpd_epi32(rt);
 668 #ifdef __XOP__
 669             gbeps            = _mm_frcz_pd(rt);
 670 #else
 671             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 672 #endif
 673             gbitab           = _mm_slli_epi32(gbitab,2);
 674
 675             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 676             F                = _mm_setzero_pd();
 677             GMX_MM_TRANSPOSE2_PD(Y,F);
 678             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 679             H                = _mm_setzero_pd();
 680             GMX_MM_TRANSPOSE2_PD(G,H);
 681             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 682             VV               = _mm_macc_pd(gbeps,Fp,Y);
 683             vgb              = _mm_mul_pd(gbqqfactor,VV);
 684
 685             twogbeps         = _mm_add_pd(gbeps,gbeps);
 686             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 687             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 688             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 689             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 690             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 691             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 692             velec            = _mm_mul_pd(qq00,rinv00);
 693             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 694
 695             /* LENNARD-JONES DISPERSION/REPULSION */
 696
 697             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 698             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 699
 700             fscal            = _mm_add_pd(felec,fvdw);
 701
 702             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 703
 704             /* Update vectorial force */
 705             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 706             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 707             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 708
 709             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 710                                                    _mm_mul_pd(dx00,fscal),
 711                                                    _mm_mul_pd(dy00,fscal),
 712                                                    _mm_mul_pd(dz00,fscal));
 713
 714             /* Inner loop uses 67 flops */
 715         }
 716
 717         /* End of innermost loop */
 718
 719         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 720                                               f+i_coord_offset,fshift+i_shift_offset);
 721
 722         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 723         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 724
 725         /* Increment number of inner iterations */
 726         inneriter                  += j_index_end - j_index_start;
 727
 728         /* Outer loop uses 7 flops */
 729     }
 730
 731     /* Increment number of outer iterations */
 732     outeriter        += nri;
 733
 734     /* Update outer/inner flops */
 735
 736     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*67);
 737 }