src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCSTab_VdwNone_GeomP1P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_single 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_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_128_fma_single
  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,C,D refer to j loop unrolling done with AVX_128, e.g. for the four 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,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  85     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128i          vfitab;
  90     __m128i          ifour       = _mm_set1_epi32(4);
  91     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  92     real             *vftab;
  93     __m128           dummy_mask,cutoff_mask;
  94     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  95     __m128           one     = _mm_set1_ps(1.0);
  96     __m128           two     = _mm_set1_ps(2.0);
  97     x                = xx[0];
  98     f                = ff[0];
  99
 100     nri              = nlist->nri;
 101     iinr             = nlist->iinr;
 102     jindex           = nlist->jindex;
 103     jjnr             = nlist->jjnr;
 104     shiftidx         = nlist->shift;
 105     gid              = nlist->gid;
 106     shiftvec         = fr->shift_vec[0];
 107     fshift           = fr->fshift[0];
 108     facel            = _mm_set1_ps(fr->ic->epsfac);
 109     charge           = mdatoms->chargeA;
 110
 111     vftab            = kernel_data->table_elec->data;
 112     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 113
 114     /* Avoid stupid compiler warnings */
 115     jnrA = jnrB = jnrC = jnrD = 0;
 116     j_coord_offsetA = 0;
 117     j_coord_offsetB = 0;
 118     j_coord_offsetC = 0;
 119     j_coord_offsetD = 0;
 120
 121     outeriter        = 0;
 122     inneriter        = 0;
 123
 124     for(iidx=0;iidx<4*DIM;iidx++)
 125     {
 126         scratch[iidx] = 0.0;
 127     }
 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_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 145
 146         fix0             = _mm_setzero_ps();
 147         fiy0             = _mm_setzero_ps();
 148         fiz0             = _mm_setzero_ps();
 149
 150         /* Load parameters for i particles */
 151         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 152
 153         /* Reset potential sums */
 154         velecsum         = _mm_setzero_ps();
 155
 156         /* Start inner kernel loop */
 157         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 158         {
 159
 160             /* Get j neighbor index, and coordinate index */
 161             jnrA             = jjnr[jidx];
 162             jnrB             = jjnr[jidx+1];
 163             jnrC             = jjnr[jidx+2];
 164             jnrD             = jjnr[jidx+3];
 165             j_coord_offsetA  = DIM*jnrA;
 166             j_coord_offsetB  = DIM*jnrB;
 167             j_coord_offsetC  = DIM*jnrC;
 168             j_coord_offsetD  = DIM*jnrD;
 169
 170             /* load j atom coordinates */
 171             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 172                                               x+j_coord_offsetC,x+j_coord_offsetD,
 173                                               &jx0,&jy0,&jz0);
 174
 175             /* Calculate displacement vector */
 176             dx00             = _mm_sub_ps(ix0,jx0);
 177             dy00             = _mm_sub_ps(iy0,jy0);
 178             dz00             = _mm_sub_ps(iz0,jz0);
 179
 180             /* Calculate squared distance and things based on it */
 181             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 182
 183             rinv00           = avx128fma_invsqrt_f(rsq00);
 184
 185             /* Load parameters for j particles */
 186             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 187                                                               charge+jnrC+0,charge+jnrD+0);
 188
 189             /**************************
 190              * CALCULATE INTERACTIONS *
 191              **************************/
 192
 193             r00              = _mm_mul_ps(rsq00,rinv00);
 194
 195             /* Compute parameters for interactions between i and j atoms */
 196             qq00             = _mm_mul_ps(iq0,jq0);
 197
 198             /* Calculate table index by multiplying r with table scale and truncate to integer */
 199             rt               = _mm_mul_ps(r00,vftabscale);
 200             vfitab           = _mm_cvttps_epi32(rt);
 201 #ifdef __XOP__
 202             vfeps            = _mm_frcz_ps(rt);
 203 #else
 204             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 205 #endif
 206             twovfeps         = _mm_add_ps(vfeps,vfeps);
 207             vfitab           = _mm_slli_epi32(vfitab,2);
 208
 209             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 210             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 211             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 212             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 213             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 214             _MM_TRANSPOSE4_PS(Y,F,G,H);
 215             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 216             VV               = _mm_macc_ps(vfeps,Fp,Y);
 217             velec            = _mm_mul_ps(qq00,VV);
 218             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 219             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 220
 221             /* Update potential sum for this i atom from the interaction with this j atom. */
 222             velecsum         = _mm_add_ps(velecsum,velec);
 223
 224             fscal            = felec;
 225
 226              /* Update vectorial force */
 227             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 228             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 229             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 230
 231             fjptrA             = f+j_coord_offsetA;
 232             fjptrB             = f+j_coord_offsetB;
 233             fjptrC             = f+j_coord_offsetC;
 234             fjptrD             = f+j_coord_offsetD;
 235             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 236                                                    _mm_mul_ps(dx00,fscal),
 237                                                    _mm_mul_ps(dy00,fscal),
 238                                                    _mm_mul_ps(dz00,fscal));
 239
 240             /* Inner loop uses 46 flops */
 241         }
 242
 243         if(jidx<j_index_end)
 244         {
 245
 246             /* Get j neighbor index, and coordinate index */
 247             jnrlistA         = jjnr[jidx];
 248             jnrlistB         = jjnr[jidx+1];
 249             jnrlistC         = jjnr[jidx+2];
 250             jnrlistD         = jjnr[jidx+3];
 251             /* Sign of each element will be negative for non-real atoms.
 252              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 253              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 254              */
 255             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 256             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 257             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 258             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 259             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 260             j_coord_offsetA  = DIM*jnrA;
 261             j_coord_offsetB  = DIM*jnrB;
 262             j_coord_offsetC  = DIM*jnrC;
 263             j_coord_offsetD  = DIM*jnrD;
 264
 265             /* load j atom coordinates */
 266             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 267                                               x+j_coord_offsetC,x+j_coord_offsetD,
 268                                               &jx0,&jy0,&jz0);
 269
 270             /* Calculate displacement vector */
 271             dx00             = _mm_sub_ps(ix0,jx0);
 272             dy00             = _mm_sub_ps(iy0,jy0);
 273             dz00             = _mm_sub_ps(iz0,jz0);
 274
 275             /* Calculate squared distance and things based on it */
 276             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 277
 278             rinv00           = avx128fma_invsqrt_f(rsq00);
 279
 280             /* Load parameters for j particles */
 281             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 282                                                               charge+jnrC+0,charge+jnrD+0);
 283
 284             /**************************
 285              * CALCULATE INTERACTIONS *
 286              **************************/
 287
 288             r00              = _mm_mul_ps(rsq00,rinv00);
 289             r00              = _mm_andnot_ps(dummy_mask,r00);
 290
 291             /* Compute parameters for interactions between i and j atoms */
 292             qq00             = _mm_mul_ps(iq0,jq0);
 293
 294             /* Calculate table index by multiplying r with table scale and truncate to integer */
 295             rt               = _mm_mul_ps(r00,vftabscale);
 296             vfitab           = _mm_cvttps_epi32(rt);
 297 #ifdef __XOP__
 298             vfeps            = _mm_frcz_ps(rt);
 299 #else
 300             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 301 #endif
 302             twovfeps         = _mm_add_ps(vfeps,vfeps);
 303             vfitab           = _mm_slli_epi32(vfitab,2);
 304
 305             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 306             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 307             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 308             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 309             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 310             _MM_TRANSPOSE4_PS(Y,F,G,H);
 311             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 312             VV               = _mm_macc_ps(vfeps,Fp,Y);
 313             velec            = _mm_mul_ps(qq00,VV);
 314             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 315             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 316
 317             /* Update potential sum for this i atom from the interaction with this j atom. */
 318             velec            = _mm_andnot_ps(dummy_mask,velec);
 319             velecsum         = _mm_add_ps(velecsum,velec);
 320
 321             fscal            = felec;
 322
 323             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 324
 325              /* Update vectorial force */
 326             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 327             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 328             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 329
 330             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 331             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 332             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 333             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 334             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 335                                                    _mm_mul_ps(dx00,fscal),
 336                                                    _mm_mul_ps(dy00,fscal),
 337                                                    _mm_mul_ps(dz00,fscal));
 338
 339             /* Inner loop uses 47 flops */
 340         }
 341
 342         /* End of innermost loop */
 343
 344         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 345                                               f+i_coord_offset,fshift+i_shift_offset);
 346
 347         ggid                        = gid[iidx];
 348         /* Update potential energies */
 349         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 350
 351         /* Increment number of inner iterations */
 352         inneriter                  += j_index_end - j_index_start;
 353
 354         /* Outer loop uses 8 flops */
 355     }
 356
 357     /* Increment number of outer iterations */
 358     outeriter        += nri;
 359
 360     /* Update outer/inner flops */
 361
 362     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*47);
 363 }
 364 /*
 365  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_128_fma_single
 366  * Electrostatics interaction: CubicSplineTable
 367  * VdW interaction:            None
 368  * Geometry:                   Particle-Particle
 369  * Calculate force/pot:        Force
 370  */
 371 void
 372 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_128_fma_single
 373                     (t_nblist                    * gmx_restrict       nlist,
 374                      rvec                        * gmx_restrict          xx,
 375                      rvec                        * gmx_restrict          ff,
 376                      struct t_forcerec           * gmx_restrict          fr,
 377                      t_mdatoms                   * gmx_restrict     mdatoms,
 378                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 379                      t_nrnb                      * gmx_restrict        nrnb)
 380 {
 381     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 382      * just 0 for non-waters.
 383      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 384      * jnr indices corresponding to data put in the four positions in the SIMD register.
 385      */
 386     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 387     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 388     int              jnrA,jnrB,jnrC,jnrD;
 389     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 390     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 391     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 392     real             rcutoff_scalar;
 393     real             *shiftvec,*fshift,*x,*f;
 394     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 395     real             scratch[4*DIM];
 396     __m128           fscal,rcutoff,rcutoff2,jidxall;
 397     int              vdwioffset0;
 398     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 399     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 400     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 401     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 402     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 403     real             *charge;
 404     __m128i          vfitab;
 405     __m128i          ifour       = _mm_set1_epi32(4);
 406     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 407     real             *vftab;
 408     __m128           dummy_mask,cutoff_mask;
 409     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 410     __m128           one     = _mm_set1_ps(1.0);
 411     __m128           two     = _mm_set1_ps(2.0);
 412     x                = xx[0];
 413     f                = ff[0];
 414
 415     nri              = nlist->nri;
 416     iinr             = nlist->iinr;
 417     jindex           = nlist->jindex;
 418     jjnr             = nlist->jjnr;
 419     shiftidx         = nlist->shift;
 420     gid              = nlist->gid;
 421     shiftvec         = fr->shift_vec[0];
 422     fshift           = fr->fshift[0];
 423     facel            = _mm_set1_ps(fr->ic->epsfac);
 424     charge           = mdatoms->chargeA;
 425
 426     vftab            = kernel_data->table_elec->data;
 427     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 428
 429     /* Avoid stupid compiler warnings */
 430     jnrA = jnrB = jnrC = jnrD = 0;
 431     j_coord_offsetA = 0;
 432     j_coord_offsetB = 0;
 433     j_coord_offsetC = 0;
 434     j_coord_offsetD = 0;
 435
 436     outeriter        = 0;
 437     inneriter        = 0;
 438
 439     for(iidx=0;iidx<4*DIM;iidx++)
 440     {
 441         scratch[iidx] = 0.0;
 442     }
 443
 444     /* Start outer loop over neighborlists */
 445     for(iidx=0; iidx<nri; iidx++)
 446     {
 447         /* Load shift vector for this list */
 448         i_shift_offset   = DIM*shiftidx[iidx];
 449
 450         /* Load limits for loop over neighbors */
 451         j_index_start    = jindex[iidx];
 452         j_index_end      = jindex[iidx+1];
 453
 454         /* Get outer coordinate index */
 455         inr              = iinr[iidx];
 456         i_coord_offset   = DIM*inr;
 457
 458         /* Load i particle coords and add shift vector */
 459         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 460
 461         fix0             = _mm_setzero_ps();
 462         fiy0             = _mm_setzero_ps();
 463         fiz0             = _mm_setzero_ps();
 464
 465         /* Load parameters for i particles */
 466         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 467
 468         /* Start inner kernel loop */
 469         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 470         {
 471
 472             /* Get j neighbor index, and coordinate index */
 473             jnrA             = jjnr[jidx];
 474             jnrB             = jjnr[jidx+1];
 475             jnrC             = jjnr[jidx+2];
 476             jnrD             = jjnr[jidx+3];
 477             j_coord_offsetA  = DIM*jnrA;
 478             j_coord_offsetB  = DIM*jnrB;
 479             j_coord_offsetC  = DIM*jnrC;
 480             j_coord_offsetD  = DIM*jnrD;
 481
 482             /* load j atom coordinates */
 483             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 484                                               x+j_coord_offsetC,x+j_coord_offsetD,
 485                                               &jx0,&jy0,&jz0);
 486
 487             /* Calculate displacement vector */
 488             dx00             = _mm_sub_ps(ix0,jx0);
 489             dy00             = _mm_sub_ps(iy0,jy0);
 490             dz00             = _mm_sub_ps(iz0,jz0);
 491
 492             /* Calculate squared distance and things based on it */
 493             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 494
 495             rinv00           = avx128fma_invsqrt_f(rsq00);
 496
 497             /* Load parameters for j particles */
 498             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 499                                                               charge+jnrC+0,charge+jnrD+0);
 500
 501             /**************************
 502              * CALCULATE INTERACTIONS *
 503              **************************/
 504
 505             r00              = _mm_mul_ps(rsq00,rinv00);
 506
 507             /* Compute parameters for interactions between i and j atoms */
 508             qq00             = _mm_mul_ps(iq0,jq0);
 509
 510             /* Calculate table index by multiplying r with table scale and truncate to integer */
 511             rt               = _mm_mul_ps(r00,vftabscale);
 512             vfitab           = _mm_cvttps_epi32(rt);
 513 #ifdef __XOP__
 514             vfeps            = _mm_frcz_ps(rt);
 515 #else
 516             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 517 #endif
 518             twovfeps         = _mm_add_ps(vfeps,vfeps);
 519             vfitab           = _mm_slli_epi32(vfitab,2);
 520
 521             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 522             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 523             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 524             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 525             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 526             _MM_TRANSPOSE4_PS(Y,F,G,H);
 527             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 528             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 529             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 530
 531             fscal            = felec;
 532
 533              /* Update vectorial force */
 534             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 535             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 536             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 537
 538             fjptrA             = f+j_coord_offsetA;
 539             fjptrB             = f+j_coord_offsetB;
 540             fjptrC             = f+j_coord_offsetC;
 541             fjptrD             = f+j_coord_offsetD;
 542             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 543                                                    _mm_mul_ps(dx00,fscal),
 544                                                    _mm_mul_ps(dy00,fscal),
 545                                                    _mm_mul_ps(dz00,fscal));
 546
 547             /* Inner loop uses 42 flops */
 548         }
 549
 550         if(jidx<j_index_end)
 551         {
 552
 553             /* Get j neighbor index, and coordinate index */
 554             jnrlistA         = jjnr[jidx];
 555             jnrlistB         = jjnr[jidx+1];
 556             jnrlistC         = jjnr[jidx+2];
 557             jnrlistD         = jjnr[jidx+3];
 558             /* Sign of each element will be negative for non-real atoms.
 559              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 560              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 561              */
 562             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 563             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 564             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 565             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 566             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 567             j_coord_offsetA  = DIM*jnrA;
 568             j_coord_offsetB  = DIM*jnrB;
 569             j_coord_offsetC  = DIM*jnrC;
 570             j_coord_offsetD  = DIM*jnrD;
 571
 572             /* load j atom coordinates */
 573             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 574                                               x+j_coord_offsetC,x+j_coord_offsetD,
 575                                               &jx0,&jy0,&jz0);
 576
 577             /* Calculate displacement vector */
 578             dx00             = _mm_sub_ps(ix0,jx0);
 579             dy00             = _mm_sub_ps(iy0,jy0);
 580             dz00             = _mm_sub_ps(iz0,jz0);
 581
 582             /* Calculate squared distance and things based on it */
 583             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 584
 585             rinv00           = avx128fma_invsqrt_f(rsq00);
 586
 587             /* Load parameters for j particles */
 588             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 589                                                               charge+jnrC+0,charge+jnrD+0);
 590
 591             /**************************
 592              * CALCULATE INTERACTIONS *
 593              **************************/
 594
 595             r00              = _mm_mul_ps(rsq00,rinv00);
 596             r00              = _mm_andnot_ps(dummy_mask,r00);
 597
 598             /* Compute parameters for interactions between i and j atoms */
 599             qq00             = _mm_mul_ps(iq0,jq0);
 600
 601             /* Calculate table index by multiplying r with table scale and truncate to integer */
 602             rt               = _mm_mul_ps(r00,vftabscale);
 603             vfitab           = _mm_cvttps_epi32(rt);
 604 #ifdef __XOP__
 605             vfeps            = _mm_frcz_ps(rt);
 606 #else
 607             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 608 #endif
 609             twovfeps         = _mm_add_ps(vfeps,vfeps);
 610             vfitab           = _mm_slli_epi32(vfitab,2);
 611
 612             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 613             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 614             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 615             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 616             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 617             _MM_TRANSPOSE4_PS(Y,F,G,H);
 618             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 619             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 620             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 621
 622             fscal            = felec;
 623
 624             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 625
 626              /* Update vectorial force */
 627             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 628             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 629             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 630
 631             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 632             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 633             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 634             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 635             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 636                                                    _mm_mul_ps(dx00,fscal),
 637                                                    _mm_mul_ps(dy00,fscal),
 638                                                    _mm_mul_ps(dz00,fscal));
 639
 640             /* Inner loop uses 43 flops */
 641         }
 642
 643         /* End of innermost loop */
 644
 645         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 646                                               f+i_coord_offset,fshift+i_shift_offset);
 647
 648         /* Increment number of inner iterations */
 649         inneriter                  += j_index_end - j_index_start;
 650
 651         /* Outer loop uses 7 flops */
 652     }
 653
 654     /* Increment number of outer iterations */
 655     outeriter        += nri;
 656
 657     /* Update outer/inner flops */
 658
 659     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*43);
 660 }