src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecGB_VdwNone_GeomP1P1_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012, by the GROMACS development team, led by
   5  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   6  * others, as listed in the AUTHORS file in the top-level source
   7  * 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 "gmx_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_VdwNone_GeomP1P1_VF_avx_128_fma_double
  54  * Electrostatics interaction: GeneralizedBorn
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecGB_VdwNone_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_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     __m128i          vfitab;
  94     __m128i          ifour       = _mm_set1_epi32(4);
  95     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
  96     real             *vftab;
  97     __m128d          dummy_mask,cutoff_mask;
  98     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  99     __m128d          one     = _mm_set1_pd(1.0);
 100     __m128d          two     = _mm_set1_pd(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm_set1_pd(fr->epsfac);
 113     charge           = mdatoms->chargeA;
 114
 115     invsqrta         = fr->invsqrta;
 116     dvda             = fr->dvda;
 117     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 118     gbtab            = fr->gbtab.data;
 119     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 120
 121     /* Avoid stupid compiler warnings */
 122     jnrA = jnrB = 0;
 123     j_coord_offsetA = 0;
 124     j_coord_offsetB = 0;
 125
 126     outeriter        = 0;
 127     inneriter        = 0;
 128
 129     /* Start outer loop over neighborlists */
 130     for(iidx=0; iidx<nri; iidx++)
 131     {
 132         /* Load shift vector for this list */
 133         i_shift_offset   = DIM*shiftidx[iidx];
 134
 135         /* Load limits for loop over neighbors */
 136         j_index_start    = jindex[iidx];
 137         j_index_end      = jindex[iidx+1];
 138
 139         /* Get outer coordinate index */
 140         inr              = iinr[iidx];
 141         i_coord_offset   = DIM*inr;
 142
 143         /* Load i particle coords and add shift vector */
 144         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 145
 146         fix0             = _mm_setzero_pd();
 147         fiy0             = _mm_setzero_pd();
 148         fiz0             = _mm_setzero_pd();
 149
 150         /* Load parameters for i particles */
 151         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 152         isai0            = _mm_load1_pd(invsqrta+inr+0);
 153
 154         /* Reset potential sums */
 155         velecsum         = _mm_setzero_pd();
 156         vgbsum           = _mm_setzero_pd();
 157         dvdasum          = _mm_setzero_pd();
 158
 159         /* Start inner kernel loop */
 160         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 161         {
 162
 163             /* Get j neighbor index, and coordinate index */
 164             jnrA             = jjnr[jidx];
 165             jnrB             = jjnr[jidx+1];
 166             j_coord_offsetA  = DIM*jnrA;
 167             j_coord_offsetB  = DIM*jnrB;
 168
 169             /* load j atom coordinates */
 170             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 171                                               &jx0,&jy0,&jz0);
 172
 173             /* Calculate displacement vector */
 174             dx00             = _mm_sub_pd(ix0,jx0);
 175             dy00             = _mm_sub_pd(iy0,jy0);
 176             dz00             = _mm_sub_pd(iz0,jz0);
 177
 178             /* Calculate squared distance and things based on it */
 179             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 180
 181             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 182
 183             /* Load parameters for j particles */
 184             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 185             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 186
 187             /**************************
 188              * CALCULATE INTERACTIONS *
 189              **************************/
 190
 191             r00              = _mm_mul_pd(rsq00,rinv00);
 192
 193             /* Compute parameters for interactions between i and j atoms */
 194             qq00             = _mm_mul_pd(iq0,jq0);
 195
 196             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 197             isaprod          = _mm_mul_pd(isai0,isaj0);
 198             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 199             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 200
 201             /* Calculate generalized born table index - this is a separate table from the normal one,
 202              * but we use the same procedure by multiplying r with scale and truncating to integer.
 203              */
 204             rt               = _mm_mul_pd(r00,gbscale);
 205             gbitab           = _mm_cvttpd_epi32(rt);
 206 #ifdef __XOP__
 207             gbeps            = _mm_frcz_pd(rt);
 208 #else
 209             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 210 #endif
 211             gbitab           = _mm_slli_epi32(gbitab,2);
 212
 213             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 214             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 215             GMX_MM_TRANSPOSE2_PD(Y,F);
 216             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 217             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 218             GMX_MM_TRANSPOSE2_PD(G,H);
 219             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 220             VV               = _mm_macc_pd(gbeps,Fp,Y);
 221             vgb              = _mm_mul_pd(gbqqfactor,VV);
 222
 223             twogbeps         = _mm_add_pd(gbeps,gbeps);
 224             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 225             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 226             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 227             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 228             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 229             velec            = _mm_mul_pd(qq00,rinv00);
 230             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 231
 232             /* Update potential sum for this i atom from the interaction with this j atom. */
 233             velecsum         = _mm_add_pd(velecsum,velec);
 234             vgbsum           = _mm_add_pd(vgbsum,vgb);
 235
 236             fscal            = felec;
 237
 238             /* Update vectorial force */
 239             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 240             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 241             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 242
 243             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 244                                                    _mm_mul_pd(dx00,fscal),
 245                                                    _mm_mul_pd(dy00,fscal),
 246                                                    _mm_mul_pd(dz00,fscal));
 247
 248             /* Inner loop uses 61 flops */
 249         }
 250
 251         if(jidx<j_index_end)
 252         {
 253
 254             jnrA             = jjnr[jidx];
 255             j_coord_offsetA  = DIM*jnrA;
 256
 257             /* load j atom coordinates */
 258             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 259                                               &jx0,&jy0,&jz0);
 260
 261             /* Calculate displacement vector */
 262             dx00             = _mm_sub_pd(ix0,jx0);
 263             dy00             = _mm_sub_pd(iy0,jy0);
 264             dz00             = _mm_sub_pd(iz0,jz0);
 265
 266             /* Calculate squared distance and things based on it */
 267             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 268
 269             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 270
 271             /* Load parameters for j particles */
 272             jq0              = _mm_load_sd(charge+jnrA+0);
 273             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 274
 275             /**************************
 276              * CALCULATE INTERACTIONS *
 277              **************************/
 278
 279             r00              = _mm_mul_pd(rsq00,rinv00);
 280
 281             /* Compute parameters for interactions between i and j atoms */
 282             qq00             = _mm_mul_pd(iq0,jq0);
 283
 284             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 285             isaprod          = _mm_mul_pd(isai0,isaj0);
 286             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 287             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 288
 289             /* Calculate generalized born table index - this is a separate table from the normal one,
 290              * but we use the same procedure by multiplying r with scale and truncating to integer.
 291              */
 292             rt               = _mm_mul_pd(r00,gbscale);
 293             gbitab           = _mm_cvttpd_epi32(rt);
 294 #ifdef __XOP__
 295             gbeps            = _mm_frcz_pd(rt);
 296 #else
 297             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 298 #endif
 299             gbitab           = _mm_slli_epi32(gbitab,2);
 300
 301             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 302             F                = _mm_setzero_pd();
 303             GMX_MM_TRANSPOSE2_PD(Y,F);
 304             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 305             H                = _mm_setzero_pd();
 306             GMX_MM_TRANSPOSE2_PD(G,H);
 307             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 308             VV               = _mm_macc_pd(gbeps,Fp,Y);
 309             vgb              = _mm_mul_pd(gbqqfactor,VV);
 310
 311             twogbeps         = _mm_add_pd(gbeps,gbeps);
 312             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 313             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 314             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 315             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 316             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 317             velec            = _mm_mul_pd(qq00,rinv00);
 318             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 319
 320             /* Update potential sum for this i atom from the interaction with this j atom. */
 321             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 322             velecsum         = _mm_add_pd(velecsum,velec);
 323             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 324             vgbsum           = _mm_add_pd(vgbsum,vgb);
 325
 326             fscal            = felec;
 327
 328             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 329
 330             /* Update vectorial force */
 331             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 332             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 333             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 334
 335             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 336                                                    _mm_mul_pd(dx00,fscal),
 337                                                    _mm_mul_pd(dy00,fscal),
 338                                                    _mm_mul_pd(dz00,fscal));
 339
 340             /* Inner loop uses 61 flops */
 341         }
 342
 343         /* End of innermost loop */
 344
 345         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 346                                               f+i_coord_offset,fshift+i_shift_offset);
 347
 348         ggid                        = gid[iidx];
 349         /* Update potential energies */
 350         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 351         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 352         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 353         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 354
 355         /* Increment number of inner iterations */
 356         inneriter                  += j_index_end - j_index_start;
 357
 358         /* Outer loop uses 9 flops */
 359     }
 360
 361     /* Increment number of outer iterations */
 362     outeriter        += nri;
 363
 364     /* Update outer/inner flops */
 365
 366     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*61);
 367 }
 368 /*
 369  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_double
 370  * Electrostatics interaction: GeneralizedBorn
 371  * VdW interaction:            None
 372  * Geometry:                   Particle-Particle
 373  * Calculate force/pot:        Force
 374  */
 375 void
 376 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_double
 377                     (t_nblist * gmx_restrict                nlist,
 378                      rvec * gmx_restrict                    xx,
 379                      rvec * gmx_restrict                    ff,
 380                      t_forcerec * gmx_restrict              fr,
 381                      t_mdatoms * gmx_restrict               mdatoms,
 382                      nb_kernel_data_t * gmx_restrict        kernel_data,
 383                      t_nrnb * gmx_restrict                  nrnb)
 384 {
 385     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 386      * just 0 for non-waters.
 387      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 388      * jnr indices corresponding to data put in the four positions in the SIMD register.
 389      */
 390     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 391     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 392     int              jnrA,jnrB;
 393     int              j_coord_offsetA,j_coord_offsetB;
 394     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 395     real             rcutoff_scalar;
 396     real             *shiftvec,*fshift,*x,*f;
 397     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 398     int              vdwioffset0;
 399     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 400     int              vdwjidx0A,vdwjidx0B;
 401     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 402     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 403     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 404     real             *charge;
 405     __m128i          gbitab;
 406     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 407     __m128d          minushalf = _mm_set1_pd(-0.5);
 408     real             *invsqrta,*dvda,*gbtab;
 409     __m128i          vfitab;
 410     __m128i          ifour       = _mm_set1_epi32(4);
 411     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 412     real             *vftab;
 413     __m128d          dummy_mask,cutoff_mask;
 414     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 415     __m128d          one     = _mm_set1_pd(1.0);
 416     __m128d          two     = _mm_set1_pd(2.0);
 417     x                = xx[0];
 418     f                = ff[0];
 419
 420     nri              = nlist->nri;
 421     iinr             = nlist->iinr;
 422     jindex           = nlist->jindex;
 423     jjnr             = nlist->jjnr;
 424     shiftidx         = nlist->shift;
 425     gid              = nlist->gid;
 426     shiftvec         = fr->shift_vec[0];
 427     fshift           = fr->fshift[0];
 428     facel            = _mm_set1_pd(fr->epsfac);
 429     charge           = mdatoms->chargeA;
 430
 431     invsqrta         = fr->invsqrta;
 432     dvda             = fr->dvda;
 433     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 434     gbtab            = fr->gbtab.data;
 435     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 436
 437     /* Avoid stupid compiler warnings */
 438     jnrA = jnrB = 0;
 439     j_coord_offsetA = 0;
 440     j_coord_offsetB = 0;
 441
 442     outeriter        = 0;
 443     inneriter        = 0;
 444
 445     /* Start outer loop over neighborlists */
 446     for(iidx=0; iidx<nri; iidx++)
 447     {
 448         /* Load shift vector for this list */
 449         i_shift_offset   = DIM*shiftidx[iidx];
 450
 451         /* Load limits for loop over neighbors */
 452         j_index_start    = jindex[iidx];
 453         j_index_end      = jindex[iidx+1];
 454
 455         /* Get outer coordinate index */
 456         inr              = iinr[iidx];
 457         i_coord_offset   = DIM*inr;
 458
 459         /* Load i particle coords and add shift vector */
 460         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 461
 462         fix0             = _mm_setzero_pd();
 463         fiy0             = _mm_setzero_pd();
 464         fiz0             = _mm_setzero_pd();
 465
 466         /* Load parameters for i particles */
 467         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 468         isai0            = _mm_load1_pd(invsqrta+inr+0);
 469
 470         dvdasum          = _mm_setzero_pd();
 471
 472         /* Start inner kernel loop */
 473         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 474         {
 475
 476             /* Get j neighbor index, and coordinate index */
 477             jnrA             = jjnr[jidx];
 478             jnrB             = jjnr[jidx+1];
 479             j_coord_offsetA  = DIM*jnrA;
 480             j_coord_offsetB  = DIM*jnrB;
 481
 482             /* load j atom coordinates */
 483             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 484                                               &jx0,&jy0,&jz0);
 485
 486             /* Calculate displacement vector */
 487             dx00             = _mm_sub_pd(ix0,jx0);
 488             dy00             = _mm_sub_pd(iy0,jy0);
 489             dz00             = _mm_sub_pd(iz0,jz0);
 490
 491             /* Calculate squared distance and things based on it */
 492             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 493
 494             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 495
 496             /* Load parameters for j particles */
 497             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 498             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 499
 500             /**************************
 501              * CALCULATE INTERACTIONS *
 502              **************************/
 503
 504             r00              = _mm_mul_pd(rsq00,rinv00);
 505
 506             /* Compute parameters for interactions between i and j atoms */
 507             qq00             = _mm_mul_pd(iq0,jq0);
 508
 509             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 510             isaprod          = _mm_mul_pd(isai0,isaj0);
 511             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 512             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 513
 514             /* Calculate generalized born table index - this is a separate table from the normal one,
 515              * but we use the same procedure by multiplying r with scale and truncating to integer.
 516              */
 517             rt               = _mm_mul_pd(r00,gbscale);
 518             gbitab           = _mm_cvttpd_epi32(rt);
 519 #ifdef __XOP__
 520             gbeps            = _mm_frcz_pd(rt);
 521 #else
 522             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 523 #endif
 524             gbitab           = _mm_slli_epi32(gbitab,2);
 525
 526             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 527             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 528             GMX_MM_TRANSPOSE2_PD(Y,F);
 529             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 530             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 531             GMX_MM_TRANSPOSE2_PD(G,H);
 532             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 533             VV               = _mm_macc_pd(gbeps,Fp,Y);
 534             vgb              = _mm_mul_pd(gbqqfactor,VV);
 535
 536             twogbeps         = _mm_add_pd(gbeps,gbeps);
 537             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 538             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 539             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 540             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 541             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 542             velec            = _mm_mul_pd(qq00,rinv00);
 543             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 544
 545             fscal            = felec;
 546
 547             /* Update vectorial force */
 548             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 549             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 550             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 551
 552             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 553                                                    _mm_mul_pd(dx00,fscal),
 554                                                    _mm_mul_pd(dy00,fscal),
 555                                                    _mm_mul_pd(dz00,fscal));
 556
 557             /* Inner loop uses 59 flops */
 558         }
 559
 560         if(jidx<j_index_end)
 561         {
 562
 563             jnrA             = jjnr[jidx];
 564             j_coord_offsetA  = DIM*jnrA;
 565
 566             /* load j atom coordinates */
 567             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 568                                               &jx0,&jy0,&jz0);
 569
 570             /* Calculate displacement vector */
 571             dx00             = _mm_sub_pd(ix0,jx0);
 572             dy00             = _mm_sub_pd(iy0,jy0);
 573             dz00             = _mm_sub_pd(iz0,jz0);
 574
 575             /* Calculate squared distance and things based on it */
 576             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 577
 578             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 579
 580             /* Load parameters for j particles */
 581             jq0              = _mm_load_sd(charge+jnrA+0);
 582             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             r00              = _mm_mul_pd(rsq00,rinv00);
 589
 590             /* Compute parameters for interactions between i and j atoms */
 591             qq00             = _mm_mul_pd(iq0,jq0);
 592
 593             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 594             isaprod          = _mm_mul_pd(isai0,isaj0);
 595             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 596             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 597
 598             /* Calculate generalized born table index - this is a separate table from the normal one,
 599              * but we use the same procedure by multiplying r with scale and truncating to integer.
 600              */
 601             rt               = _mm_mul_pd(r00,gbscale);
 602             gbitab           = _mm_cvttpd_epi32(rt);
 603 #ifdef __XOP__
 604             gbeps            = _mm_frcz_pd(rt);
 605 #else
 606             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 607 #endif
 608             gbitab           = _mm_slli_epi32(gbitab,2);
 609
 610             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 611             F                = _mm_setzero_pd();
 612             GMX_MM_TRANSPOSE2_PD(Y,F);
 613             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 614             H                = _mm_setzero_pd();
 615             GMX_MM_TRANSPOSE2_PD(G,H);
 616             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 617             VV               = _mm_macc_pd(gbeps,Fp,Y);
 618             vgb              = _mm_mul_pd(gbqqfactor,VV);
 619
 620             twogbeps         = _mm_add_pd(gbeps,gbeps);
 621             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 622             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 623             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 624             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 625             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 626             velec            = _mm_mul_pd(qq00,rinv00);
 627             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 628
 629             fscal            = felec;
 630
 631             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 632
 633             /* Update vectorial force */
 634             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 635             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 636             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 637
 638             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 639                                                    _mm_mul_pd(dx00,fscal),
 640                                                    _mm_mul_pd(dy00,fscal),
 641                                                    _mm_mul_pd(dz00,fscal));
 642
 643             /* Inner loop uses 59 flops */
 644         }
 645
 646         /* End of innermost loop */
 647
 648         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 649                                               f+i_coord_offset,fshift+i_shift_offset);
 650
 651         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 652         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 653
 654         /* Increment number of inner iterations */
 655         inneriter                  += j_index_end - j_index_start;
 656
 657         /* Outer loop uses 7 flops */
 658     }
 659
 660     /* Increment number of outer iterations */
 661     outeriter        += nri;
 662
 663     /* Update outer/inner flops */
 664
 665     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*59);
 666 }