src/gromacs/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) 2012,2013,2014,2015,2017, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_double
  51  * Electrostatics interaction: GeneralizedBorn
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_128_fma_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwjidx0A,vdwjidx0B;
  82     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  83     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  84     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  85     real             *charge;
  86     __m128i          gbitab;
  87     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
  88     __m128d          minushalf = _mm_set1_pd(-0.5);
  89     real             *invsqrta,*dvda,*gbtab;
  90     __m128i          vfitab;
  91     __m128i          ifour       = _mm_set1_epi32(4);
  92     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
  93     real             *vftab;
  94     __m128d          dummy_mask,cutoff_mask;
  95     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  96     __m128d          one     = _mm_set1_pd(1.0);
  97     __m128d          two     = _mm_set1_pd(2.0);
  98     x                = xx[0];
  99     f                = ff[0];
 100
 101     nri              = nlist->nri;
 102     iinr             = nlist->iinr;
 103     jindex           = nlist->jindex;
 104     jjnr             = nlist->jjnr;
 105     shiftidx         = nlist->shift;
 106     gid              = nlist->gid;
 107     shiftvec         = fr->shift_vec[0];
 108     fshift           = fr->fshift[0];
 109     facel            = _mm_set1_pd(fr->ic->epsfac);
 110     charge           = mdatoms->chargeA;
 111
 112     invsqrta         = fr->invsqrta;
 113     dvda             = fr->dvda;
 114     gbtabscale       = _mm_set1_pd(fr->gbtab->scale);
 115     gbtab            = fr->gbtab->data;
 116     gbinvepsdiff     = _mm_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 117
 118     /* Avoid stupid compiler warnings */
 119     jnrA = jnrB = 0;
 120     j_coord_offsetA = 0;
 121     j_coord_offsetB = 0;
 122
 123     outeriter        = 0;
 124     inneriter        = 0;
 125
 126     /* Start outer loop over neighborlists */
 127     for(iidx=0; iidx<nri; iidx++)
 128     {
 129         /* Load shift vector for this list */
 130         i_shift_offset   = DIM*shiftidx[iidx];
 131
 132         /* Load limits for loop over neighbors */
 133         j_index_start    = jindex[iidx];
 134         j_index_end      = jindex[iidx+1];
 135
 136         /* Get outer coordinate index */
 137         inr              = iinr[iidx];
 138         i_coord_offset   = DIM*inr;
 139
 140         /* Load i particle coords and add shift vector */
 141         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 142
 143         fix0             = _mm_setzero_pd();
 144         fiy0             = _mm_setzero_pd();
 145         fiz0             = _mm_setzero_pd();
 146
 147         /* Load parameters for i particles */
 148         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 149         isai0            = _mm_load1_pd(invsqrta+inr+0);
 150
 151         /* Reset potential sums */
 152         velecsum         = _mm_setzero_pd();
 153         vgbsum           = _mm_setzero_pd();
 154         dvdasum          = _mm_setzero_pd();
 155
 156         /* Start inner kernel loop */
 157         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 158         {
 159
 160             /* Get j neighbor index, and coordinate index */
 161             jnrA             = jjnr[jidx];
 162             jnrB             = jjnr[jidx+1];
 163             j_coord_offsetA  = DIM*jnrA;
 164             j_coord_offsetB  = DIM*jnrB;
 165
 166             /* load j atom coordinates */
 167             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 168                                               &jx0,&jy0,&jz0);
 169
 170             /* Calculate displacement vector */
 171             dx00             = _mm_sub_pd(ix0,jx0);
 172             dy00             = _mm_sub_pd(iy0,jy0);
 173             dz00             = _mm_sub_pd(iz0,jz0);
 174
 175             /* Calculate squared distance and things based on it */
 176             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 177
 178             rinv00           = avx128fma_invsqrt_d(rsq00);
 179
 180             /* Load parameters for j particles */
 181             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 182             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 183
 184             /**************************
 185              * CALCULATE INTERACTIONS *
 186              **************************/
 187
 188             r00              = _mm_mul_pd(rsq00,rinv00);
 189
 190             /* Compute parameters for interactions between i and j atoms */
 191             qq00             = _mm_mul_pd(iq0,jq0);
 192
 193             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 194             isaprod          = _mm_mul_pd(isai0,isaj0);
 195             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 196             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 197
 198             /* Calculate generalized born table index - this is a separate table from the normal one,
 199              * but we use the same procedure by multiplying r with scale and truncating to integer.
 200              */
 201             rt               = _mm_mul_pd(r00,gbscale);
 202             gbitab           = _mm_cvttpd_epi32(rt);
 203 #ifdef __XOP__
 204             gbeps            = _mm_frcz_pd(rt);
 205 #else
 206             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 207 #endif
 208             gbitab           = _mm_slli_epi32(gbitab,2);
 209
 210             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 211             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 212             GMX_MM_TRANSPOSE2_PD(Y,F);
 213             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 214             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 215             GMX_MM_TRANSPOSE2_PD(G,H);
 216             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 217             VV               = _mm_macc_pd(gbeps,Fp,Y);
 218             vgb              = _mm_mul_pd(gbqqfactor,VV);
 219
 220             twogbeps         = _mm_add_pd(gbeps,gbeps);
 221             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 222             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 223             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 224             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 225             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 226             velec            = _mm_mul_pd(qq00,rinv00);
 227             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 228
 229             /* Update potential sum for this i atom from the interaction with this j atom. */
 230             velecsum         = _mm_add_pd(velecsum,velec);
 231             vgbsum           = _mm_add_pd(vgbsum,vgb);
 232
 233             fscal            = felec;
 234
 235             /* Update vectorial force */
 236             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 237             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 238             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 239
 240             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 241                                                    _mm_mul_pd(dx00,fscal),
 242                                                    _mm_mul_pd(dy00,fscal),
 243                                                    _mm_mul_pd(dz00,fscal));
 244
 245             /* Inner loop uses 61 flops */
 246         }
 247
 248         if(jidx<j_index_end)
 249         {
 250
 251             jnrA             = jjnr[jidx];
 252             j_coord_offsetA  = DIM*jnrA;
 253
 254             /* load j atom coordinates */
 255             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 256                                               &jx0,&jy0,&jz0);
 257
 258             /* Calculate displacement vector */
 259             dx00             = _mm_sub_pd(ix0,jx0);
 260             dy00             = _mm_sub_pd(iy0,jy0);
 261             dz00             = _mm_sub_pd(iz0,jz0);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 265
 266             rinv00           = avx128fma_invsqrt_d(rsq00);
 267
 268             /* Load parameters for j particles */
 269             jq0              = _mm_load_sd(charge+jnrA+0);
 270             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 271
 272             /**************************
 273              * CALCULATE INTERACTIONS *
 274              **************************/
 275
 276             r00              = _mm_mul_pd(rsq00,rinv00);
 277
 278             /* Compute parameters for interactions between i and j atoms */
 279             qq00             = _mm_mul_pd(iq0,jq0);
 280
 281             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 282             isaprod          = _mm_mul_pd(isai0,isaj0);
 283             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 284             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 285
 286             /* Calculate generalized born table index - this is a separate table from the normal one,
 287              * but we use the same procedure by multiplying r with scale and truncating to integer.
 288              */
 289             rt               = _mm_mul_pd(r00,gbscale);
 290             gbitab           = _mm_cvttpd_epi32(rt);
 291 #ifdef __XOP__
 292             gbeps            = _mm_frcz_pd(rt);
 293 #else
 294             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 295 #endif
 296             gbitab           = _mm_slli_epi32(gbitab,2);
 297
 298             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 299             F                = _mm_setzero_pd();
 300             GMX_MM_TRANSPOSE2_PD(Y,F);
 301             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 302             H                = _mm_setzero_pd();
 303             GMX_MM_TRANSPOSE2_PD(G,H);
 304             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 305             VV               = _mm_macc_pd(gbeps,Fp,Y);
 306             vgb              = _mm_mul_pd(gbqqfactor,VV);
 307
 308             twogbeps         = _mm_add_pd(gbeps,gbeps);
 309             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 310             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 311             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 312             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 313             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 314             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 315             velec            = _mm_mul_pd(qq00,rinv00);
 316             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 320             velecsum         = _mm_add_pd(velecsum,velec);
 321             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 322             vgbsum           = _mm_add_pd(vgbsum,vgb);
 323
 324             fscal            = felec;
 325
 326             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 327
 328             /* Update vectorial force */
 329             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 330             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 331             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 332
 333             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 334                                                    _mm_mul_pd(dx00,fscal),
 335                                                    _mm_mul_pd(dy00,fscal),
 336                                                    _mm_mul_pd(dz00,fscal));
 337
 338             /* Inner loop uses 61 flops */
 339         }
 340
 341         /* End of innermost loop */
 342
 343         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 344                                               f+i_coord_offset,fshift+i_shift_offset);
 345
 346         ggid                        = gid[iidx];
 347         /* Update potential energies */
 348         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 349         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 350         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 351         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 352
 353         /* Increment number of inner iterations */
 354         inneriter                  += j_index_end - j_index_start;
 355
 356         /* Outer loop uses 9 flops */
 357     }
 358
 359     /* Increment number of outer iterations */
 360     outeriter        += nri;
 361
 362     /* Update outer/inner flops */
 363
 364     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*61);
 365 }
 366 /*
 367  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_double
 368  * Electrostatics interaction: GeneralizedBorn
 369  * VdW interaction:            None
 370  * Geometry:                   Particle-Particle
 371  * Calculate force/pot:        Force
 372  */
 373 void
 374 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_double
 375                     (t_nblist                    * gmx_restrict       nlist,
 376                      rvec                        * gmx_restrict          xx,
 377                      rvec                        * gmx_restrict          ff,
 378                      struct t_forcerec           * gmx_restrict          fr,
 379                      t_mdatoms                   * gmx_restrict     mdatoms,
 380                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 381                      t_nrnb                      * gmx_restrict        nrnb)
 382 {
 383     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 384      * just 0 for non-waters.
 385      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 386      * jnr indices corresponding to data put in the four positions in the SIMD register.
 387      */
 388     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 389     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 390     int              jnrA,jnrB;
 391     int              j_coord_offsetA,j_coord_offsetB;
 392     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 393     real             rcutoff_scalar;
 394     real             *shiftvec,*fshift,*x,*f;
 395     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 396     int              vdwioffset0;
 397     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 398     int              vdwjidx0A,vdwjidx0B;
 399     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 400     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 401     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 402     real             *charge;
 403     __m128i          gbitab;
 404     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 405     __m128d          minushalf = _mm_set1_pd(-0.5);
 406     real             *invsqrta,*dvda,*gbtab;
 407     __m128i          vfitab;
 408     __m128i          ifour       = _mm_set1_epi32(4);
 409     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 410     real             *vftab;
 411     __m128d          dummy_mask,cutoff_mask;
 412     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 413     __m128d          one     = _mm_set1_pd(1.0);
 414     __m128d          two     = _mm_set1_pd(2.0);
 415     x                = xx[0];
 416     f                = ff[0];
 417
 418     nri              = nlist->nri;
 419     iinr             = nlist->iinr;
 420     jindex           = nlist->jindex;
 421     jjnr             = nlist->jjnr;
 422     shiftidx         = nlist->shift;
 423     gid              = nlist->gid;
 424     shiftvec         = fr->shift_vec[0];
 425     fshift           = fr->fshift[0];
 426     facel            = _mm_set1_pd(fr->ic->epsfac);
 427     charge           = mdatoms->chargeA;
 428
 429     invsqrta         = fr->invsqrta;
 430     dvda             = fr->dvda;
 431     gbtabscale       = _mm_set1_pd(fr->gbtab->scale);
 432     gbtab            = fr->gbtab->data;
 433     gbinvepsdiff     = _mm_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 434
 435     /* Avoid stupid compiler warnings */
 436     jnrA = jnrB = 0;
 437     j_coord_offsetA = 0;
 438     j_coord_offsetB = 0;
 439
 440     outeriter        = 0;
 441     inneriter        = 0;
 442
 443     /* Start outer loop over neighborlists */
 444     for(iidx=0; iidx<nri; iidx++)
 445     {
 446         /* Load shift vector for this list */
 447         i_shift_offset   = DIM*shiftidx[iidx];
 448
 449         /* Load limits for loop over neighbors */
 450         j_index_start    = jindex[iidx];
 451         j_index_end      = jindex[iidx+1];
 452
 453         /* Get outer coordinate index */
 454         inr              = iinr[iidx];
 455         i_coord_offset   = DIM*inr;
 456
 457         /* Load i particle coords and add shift vector */
 458         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 459
 460         fix0             = _mm_setzero_pd();
 461         fiy0             = _mm_setzero_pd();
 462         fiz0             = _mm_setzero_pd();
 463
 464         /* Load parameters for i particles */
 465         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 466         isai0            = _mm_load1_pd(invsqrta+inr+0);
 467
 468         dvdasum          = _mm_setzero_pd();
 469
 470         /* Start inner kernel loop */
 471         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 472         {
 473
 474             /* Get j neighbor index, and coordinate index */
 475             jnrA             = jjnr[jidx];
 476             jnrB             = jjnr[jidx+1];
 477             j_coord_offsetA  = DIM*jnrA;
 478             j_coord_offsetB  = DIM*jnrB;
 479
 480             /* load j atom coordinates */
 481             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 482                                               &jx0,&jy0,&jz0);
 483
 484             /* Calculate displacement vector */
 485             dx00             = _mm_sub_pd(ix0,jx0);
 486             dy00             = _mm_sub_pd(iy0,jy0);
 487             dz00             = _mm_sub_pd(iz0,jz0);
 488
 489             /* Calculate squared distance and things based on it */
 490             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 491
 492             rinv00           = avx128fma_invsqrt_d(rsq00);
 493
 494             /* Load parameters for j particles */
 495             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 496             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 497
 498             /**************************
 499              * CALCULATE INTERACTIONS *
 500              **************************/
 501
 502             r00              = _mm_mul_pd(rsq00,rinv00);
 503
 504             /* Compute parameters for interactions between i and j atoms */
 505             qq00             = _mm_mul_pd(iq0,jq0);
 506
 507             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 508             isaprod          = _mm_mul_pd(isai0,isaj0);
 509             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 510             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 511
 512             /* Calculate generalized born table index - this is a separate table from the normal one,
 513              * but we use the same procedure by multiplying r with scale and truncating to integer.
 514              */
 515             rt               = _mm_mul_pd(r00,gbscale);
 516             gbitab           = _mm_cvttpd_epi32(rt);
 517 #ifdef __XOP__
 518             gbeps            = _mm_frcz_pd(rt);
 519 #else
 520             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 521 #endif
 522             gbitab           = _mm_slli_epi32(gbitab,2);
 523
 524             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 525             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 526             GMX_MM_TRANSPOSE2_PD(Y,F);
 527             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 528             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 529             GMX_MM_TRANSPOSE2_PD(G,H);
 530             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 531             VV               = _mm_macc_pd(gbeps,Fp,Y);
 532             vgb              = _mm_mul_pd(gbqqfactor,VV);
 533
 534             twogbeps         = _mm_add_pd(gbeps,gbeps);
 535             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 536             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 537             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 538             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 539             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 540             velec            = _mm_mul_pd(qq00,rinv00);
 541             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 542
 543             fscal            = felec;
 544
 545             /* Update vectorial force */
 546             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 547             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 548             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 549
 550             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 551                                                    _mm_mul_pd(dx00,fscal),
 552                                                    _mm_mul_pd(dy00,fscal),
 553                                                    _mm_mul_pd(dz00,fscal));
 554
 555             /* Inner loop uses 59 flops */
 556         }
 557
 558         if(jidx<j_index_end)
 559         {
 560
 561             jnrA             = jjnr[jidx];
 562             j_coord_offsetA  = DIM*jnrA;
 563
 564             /* load j atom coordinates */
 565             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 566                                               &jx0,&jy0,&jz0);
 567
 568             /* Calculate displacement vector */
 569             dx00             = _mm_sub_pd(ix0,jx0);
 570             dy00             = _mm_sub_pd(iy0,jy0);
 571             dz00             = _mm_sub_pd(iz0,jz0);
 572
 573             /* Calculate squared distance and things based on it */
 574             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 575
 576             rinv00           = avx128fma_invsqrt_d(rsq00);
 577
 578             /* Load parameters for j particles */
 579             jq0              = _mm_load_sd(charge+jnrA+0);
 580             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             r00              = _mm_mul_pd(rsq00,rinv00);
 587
 588             /* Compute parameters for interactions between i and j atoms */
 589             qq00             = _mm_mul_pd(iq0,jq0);
 590
 591             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 592             isaprod          = _mm_mul_pd(isai0,isaj0);
 593             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 594             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 595
 596             /* Calculate generalized born table index - this is a separate table from the normal one,
 597              * but we use the same procedure by multiplying r with scale and truncating to integer.
 598              */
 599             rt               = _mm_mul_pd(r00,gbscale);
 600             gbitab           = _mm_cvttpd_epi32(rt);
 601 #ifdef __XOP__
 602             gbeps            = _mm_frcz_pd(rt);
 603 #else
 604             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 605 #endif
 606             gbitab           = _mm_slli_epi32(gbitab,2);
 607
 608             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 609             F                = _mm_setzero_pd();
 610             GMX_MM_TRANSPOSE2_PD(Y,F);
 611             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 612             H                = _mm_setzero_pd();
 613             GMX_MM_TRANSPOSE2_PD(G,H);
 614             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 615             VV               = _mm_macc_pd(gbeps,Fp,Y);
 616             vgb              = _mm_mul_pd(gbqqfactor,VV);
 617
 618             twogbeps         = _mm_add_pd(gbeps,gbeps);
 619             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 620             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 621             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 622             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 623             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 624             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 625             velec            = _mm_mul_pd(qq00,rinv00);
 626             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 627
 628             fscal            = felec;
 629
 630             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 631
 632             /* Update vectorial force */
 633             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 634             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 635             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 636
 637             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 638                                                    _mm_mul_pd(dx00,fscal),
 639                                                    _mm_mul_pd(dy00,fscal),
 640                                                    _mm_mul_pd(dz00,fscal));
 641
 642             /* Inner loop uses 59 flops */
 643         }
 644
 645         /* End of innermost loop */
 646
 647         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 648                                               f+i_coord_offset,fshift+i_shift_offset);
 649
 650         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 651         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 652
 653         /* Increment number of inner iterations */
 654         inneriter                  += j_index_end - j_index_start;
 655
 656         /* Outer loop uses 7 flops */
 657     }
 658
 659     /* Increment number of outer iterations */
 660     outeriter        += nri;
 661
 662     /* Update outer/inner flops */
 663
 664     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*59);
 665 }