src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_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_ElecEwSh_VdwNone_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_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          ewitab;
  90     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  91     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  92     real             *ewtab;
  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     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 112     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 113     beta2            = _mm_mul_ps(beta,beta);
 114     beta3            = _mm_mul_ps(beta,beta2);
 115     ewtab            = fr->ic->tabq_coul_FDV0;
 116     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 117     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 118
 119     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 120     rcutoff_scalar   = fr->ic->rcoulomb;
 121     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 122     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm_setzero_ps();
 157         fiy0             = _mm_setzero_ps();
 158         fiz0             = _mm_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 162
 163         /* Reset potential sums */
 164         velecsum         = _mm_setzero_ps();
 165
 166         /* Start inner kernel loop */
 167         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 168         {
 169
 170             /* Get j neighbor index, and coordinate index */
 171             jnrA             = jjnr[jidx];
 172             jnrB             = jjnr[jidx+1];
 173             jnrC             = jjnr[jidx+2];
 174             jnrD             = jjnr[jidx+3];
 175             j_coord_offsetA  = DIM*jnrA;
 176             j_coord_offsetB  = DIM*jnrB;
 177             j_coord_offsetC  = DIM*jnrC;
 178             j_coord_offsetD  = DIM*jnrD;
 179
 180             /* load j atom coordinates */
 181             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 182                                               x+j_coord_offsetC,x+j_coord_offsetD,
 183                                               &jx0,&jy0,&jz0);
 184
 185             /* Calculate displacement vector */
 186             dx00             = _mm_sub_ps(ix0,jx0);
 187             dy00             = _mm_sub_ps(iy0,jy0);
 188             dz00             = _mm_sub_ps(iz0,jz0);
 189
 190             /* Calculate squared distance and things based on it */
 191             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 192
 193             rinv00           = avx128fma_invsqrt_f(rsq00);
 194
 195             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 196
 197             /* Load parameters for j particles */
 198             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 199                                                               charge+jnrC+0,charge+jnrD+0);
 200
 201             /**************************
 202              * CALCULATE INTERACTIONS *
 203              **************************/
 204
 205             if (gmx_mm_any_lt(rsq00,rcutoff2))
 206             {
 207
 208             r00              = _mm_mul_ps(rsq00,rinv00);
 209
 210             /* Compute parameters for interactions between i and j atoms */
 211             qq00             = _mm_mul_ps(iq0,jq0);
 212
 213             /* EWALD ELECTROSTATICS */
 214
 215             /* Analytical PME correction */
 216             zeta2            = _mm_mul_ps(beta2,rsq00);
 217             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 218             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 219             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 220             felec            = _mm_mul_ps(qq00,felec);
 221             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 222             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 223             velec            = _mm_mul_ps(qq00,velec);
 224
 225             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 226
 227             /* Update potential sum for this i atom from the interaction with this j atom. */
 228             velec            = _mm_and_ps(velec,cutoff_mask);
 229             velecsum         = _mm_add_ps(velecsum,velec);
 230
 231             fscal            = felec;
 232
 233             fscal            = _mm_and_ps(fscal,cutoff_mask);
 234
 235              /* Update vectorial force */
 236             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 237             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 238             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 239
 240             fjptrA             = f+j_coord_offsetA;
 241             fjptrB             = f+j_coord_offsetB;
 242             fjptrC             = f+j_coord_offsetC;
 243             fjptrD             = f+j_coord_offsetD;
 244             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 245                                                    _mm_mul_ps(dx00,fscal),
 246                                                    _mm_mul_ps(dy00,fscal),
 247                                                    _mm_mul_ps(dz00,fscal));
 248
 249             }
 250
 251             /* Inner loop uses 33 flops */
 252         }
 253
 254         if(jidx<j_index_end)
 255         {
 256
 257             /* Get j neighbor index, and coordinate index */
 258             jnrlistA         = jjnr[jidx];
 259             jnrlistB         = jjnr[jidx+1];
 260             jnrlistC         = jjnr[jidx+2];
 261             jnrlistD         = jjnr[jidx+3];
 262             /* Sign of each element will be negative for non-real atoms.
 263              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 264              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 265              */
 266             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 267             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 268             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 269             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 270             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 271             j_coord_offsetA  = DIM*jnrA;
 272             j_coord_offsetB  = DIM*jnrB;
 273             j_coord_offsetC  = DIM*jnrC;
 274             j_coord_offsetD  = DIM*jnrD;
 275
 276             /* load j atom coordinates */
 277             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 278                                               x+j_coord_offsetC,x+j_coord_offsetD,
 279                                               &jx0,&jy0,&jz0);
 280
 281             /* Calculate displacement vector */
 282             dx00             = _mm_sub_ps(ix0,jx0);
 283             dy00             = _mm_sub_ps(iy0,jy0);
 284             dz00             = _mm_sub_ps(iz0,jz0);
 285
 286             /* Calculate squared distance and things based on it */
 287             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 288
 289             rinv00           = avx128fma_invsqrt_f(rsq00);
 290
 291             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 292
 293             /* Load parameters for j particles */
 294             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 295                                                               charge+jnrC+0,charge+jnrD+0);
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             if (gmx_mm_any_lt(rsq00,rcutoff2))
 302             {
 303
 304             r00              = _mm_mul_ps(rsq00,rinv00);
 305             r00              = _mm_andnot_ps(dummy_mask,r00);
 306
 307             /* Compute parameters for interactions between i and j atoms */
 308             qq00             = _mm_mul_ps(iq0,jq0);
 309
 310             /* EWALD ELECTROSTATICS */
 311
 312             /* Analytical PME correction */
 313             zeta2            = _mm_mul_ps(beta2,rsq00);
 314             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 315             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 316             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 317             felec            = _mm_mul_ps(qq00,felec);
 318             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 319             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 320             velec            = _mm_mul_ps(qq00,velec);
 321
 322             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 323
 324             /* Update potential sum for this i atom from the interaction with this j atom. */
 325             velec            = _mm_and_ps(velec,cutoff_mask);
 326             velec            = _mm_andnot_ps(dummy_mask,velec);
 327             velecsum         = _mm_add_ps(velecsum,velec);
 328
 329             fscal            = felec;
 330
 331             fscal            = _mm_and_ps(fscal,cutoff_mask);
 332
 333             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 334
 335              /* Update vectorial force */
 336             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 337             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 338             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 339
 340             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 341             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 342             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 343             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 344             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 345                                                    _mm_mul_ps(dx00,fscal),
 346                                                    _mm_mul_ps(dy00,fscal),
 347                                                    _mm_mul_ps(dz00,fscal));
 348
 349             }
 350
 351             /* Inner loop uses 34 flops */
 352         }
 353
 354         /* End of innermost loop */
 355
 356         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 357                                               f+i_coord_offset,fshift+i_shift_offset);
 358
 359         ggid                        = gid[iidx];
 360         /* Update potential energies */
 361         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 362
 363         /* Increment number of inner iterations */
 364         inneriter                  += j_index_end - j_index_start;
 365
 366         /* Outer loop uses 8 flops */
 367     }
 368
 369     /* Increment number of outer iterations */
 370     outeriter        += nri;
 371
 372     /* Update outer/inner flops */
 373
 374     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*34);
 375 }
 376 /*
 377  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_128_fma_single
 378  * Electrostatics interaction: Ewald
 379  * VdW interaction:            None
 380  * Geometry:                   Particle-Particle
 381  * Calculate force/pot:        Force
 382  */
 383 void
 384 nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_128_fma_single
 385                     (t_nblist                    * gmx_restrict       nlist,
 386                      rvec                        * gmx_restrict          xx,
 387                      rvec                        * gmx_restrict          ff,
 388                      struct t_forcerec           * gmx_restrict          fr,
 389                      t_mdatoms                   * gmx_restrict     mdatoms,
 390                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 391                      t_nrnb                      * gmx_restrict        nrnb)
 392 {
 393     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 394      * just 0 for non-waters.
 395      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 396      * jnr indices corresponding to data put in the four positions in the SIMD register.
 397      */
 398     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 399     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 400     int              jnrA,jnrB,jnrC,jnrD;
 401     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 402     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 403     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 404     real             rcutoff_scalar;
 405     real             *shiftvec,*fshift,*x,*f;
 406     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 407     real             scratch[4*DIM];
 408     __m128           fscal,rcutoff,rcutoff2,jidxall;
 409     int              vdwioffset0;
 410     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 411     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 412     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 413     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 414     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 415     real             *charge;
 416     __m128i          ewitab;
 417     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 418     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 419     real             *ewtab;
 420     __m128           dummy_mask,cutoff_mask;
 421     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 422     __m128           one     = _mm_set1_ps(1.0);
 423     __m128           two     = _mm_set1_ps(2.0);
 424     x                = xx[0];
 425     f                = ff[0];
 426
 427     nri              = nlist->nri;
 428     iinr             = nlist->iinr;
 429     jindex           = nlist->jindex;
 430     jjnr             = nlist->jjnr;
 431     shiftidx         = nlist->shift;
 432     gid              = nlist->gid;
 433     shiftvec         = fr->shift_vec[0];
 434     fshift           = fr->fshift[0];
 435     facel            = _mm_set1_ps(fr->ic->epsfac);
 436     charge           = mdatoms->chargeA;
 437
 438     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 439     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 440     beta2            = _mm_mul_ps(beta,beta);
 441     beta3            = _mm_mul_ps(beta,beta2);
 442     ewtab            = fr->ic->tabq_coul_F;
 443     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 444     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 445
 446     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 447     rcutoff_scalar   = fr->ic->rcoulomb;
 448     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 449     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 450
 451     /* Avoid stupid compiler warnings */
 452     jnrA = jnrB = jnrC = jnrD = 0;
 453     j_coord_offsetA = 0;
 454     j_coord_offsetB = 0;
 455     j_coord_offsetC = 0;
 456     j_coord_offsetD = 0;
 457
 458     outeriter        = 0;
 459     inneriter        = 0;
 460
 461     for(iidx=0;iidx<4*DIM;iidx++)
 462     {
 463         scratch[iidx] = 0.0;
 464     }
 465
 466     /* Start outer loop over neighborlists */
 467     for(iidx=0; iidx<nri; iidx++)
 468     {
 469         /* Load shift vector for this list */
 470         i_shift_offset   = DIM*shiftidx[iidx];
 471
 472         /* Load limits for loop over neighbors */
 473         j_index_start    = jindex[iidx];
 474         j_index_end      = jindex[iidx+1];
 475
 476         /* Get outer coordinate index */
 477         inr              = iinr[iidx];
 478         i_coord_offset   = DIM*inr;
 479
 480         /* Load i particle coords and add shift vector */
 481         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 482
 483         fix0             = _mm_setzero_ps();
 484         fiy0             = _mm_setzero_ps();
 485         fiz0             = _mm_setzero_ps();
 486
 487         /* Load parameters for i particles */
 488         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 489
 490         /* Start inner kernel loop */
 491         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 492         {
 493
 494             /* Get j neighbor index, and coordinate index */
 495             jnrA             = jjnr[jidx];
 496             jnrB             = jjnr[jidx+1];
 497             jnrC             = jjnr[jidx+2];
 498             jnrD             = jjnr[jidx+3];
 499             j_coord_offsetA  = DIM*jnrA;
 500             j_coord_offsetB  = DIM*jnrB;
 501             j_coord_offsetC  = DIM*jnrC;
 502             j_coord_offsetD  = DIM*jnrD;
 503
 504             /* load j atom coordinates */
 505             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 506                                               x+j_coord_offsetC,x+j_coord_offsetD,
 507                                               &jx0,&jy0,&jz0);
 508
 509             /* Calculate displacement vector */
 510             dx00             = _mm_sub_ps(ix0,jx0);
 511             dy00             = _mm_sub_ps(iy0,jy0);
 512             dz00             = _mm_sub_ps(iz0,jz0);
 513
 514             /* Calculate squared distance and things based on it */
 515             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 516
 517             rinv00           = avx128fma_invsqrt_f(rsq00);
 518
 519             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 520
 521             /* Load parameters for j particles */
 522             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 523                                                               charge+jnrC+0,charge+jnrD+0);
 524
 525             /**************************
 526              * CALCULATE INTERACTIONS *
 527              **************************/
 528
 529             if (gmx_mm_any_lt(rsq00,rcutoff2))
 530             {
 531
 532             r00              = _mm_mul_ps(rsq00,rinv00);
 533
 534             /* Compute parameters for interactions between i and j atoms */
 535             qq00             = _mm_mul_ps(iq0,jq0);
 536
 537             /* EWALD ELECTROSTATICS */
 538
 539             /* Analytical PME correction */
 540             zeta2            = _mm_mul_ps(beta2,rsq00);
 541             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 542             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 543             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 544             felec            = _mm_mul_ps(qq00,felec);
 545
 546             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 547
 548             fscal            = felec;
 549
 550             fscal            = _mm_and_ps(fscal,cutoff_mask);
 551
 552              /* Update vectorial force */
 553             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 554             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 555             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 556
 557             fjptrA             = f+j_coord_offsetA;
 558             fjptrB             = f+j_coord_offsetB;
 559             fjptrC             = f+j_coord_offsetC;
 560             fjptrD             = f+j_coord_offsetD;
 561             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 562                                                    _mm_mul_ps(dx00,fscal),
 563                                                    _mm_mul_ps(dy00,fscal),
 564                                                    _mm_mul_ps(dz00,fscal));
 565
 566             }
 567
 568             /* Inner loop uses 31 flops */
 569         }
 570
 571         if(jidx<j_index_end)
 572         {
 573
 574             /* Get j neighbor index, and coordinate index */
 575             jnrlistA         = jjnr[jidx];
 576             jnrlistB         = jjnr[jidx+1];
 577             jnrlistC         = jjnr[jidx+2];
 578             jnrlistD         = jjnr[jidx+3];
 579             /* Sign of each element will be negative for non-real atoms.
 580              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 581              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 582              */
 583             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 584             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 585             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 586             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 587             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 588             j_coord_offsetA  = DIM*jnrA;
 589             j_coord_offsetB  = DIM*jnrB;
 590             j_coord_offsetC  = DIM*jnrC;
 591             j_coord_offsetD  = DIM*jnrD;
 592
 593             /* load j atom coordinates */
 594             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 595                                               x+j_coord_offsetC,x+j_coord_offsetD,
 596                                               &jx0,&jy0,&jz0);
 597
 598             /* Calculate displacement vector */
 599             dx00             = _mm_sub_ps(ix0,jx0);
 600             dy00             = _mm_sub_ps(iy0,jy0);
 601             dz00             = _mm_sub_ps(iz0,jz0);
 602
 603             /* Calculate squared distance and things based on it */
 604             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 605
 606             rinv00           = avx128fma_invsqrt_f(rsq00);
 607
 608             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 609
 610             /* Load parameters for j particles */
 611             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 612                                                               charge+jnrC+0,charge+jnrD+0);
 613
 614             /**************************
 615              * CALCULATE INTERACTIONS *
 616              **************************/
 617
 618             if (gmx_mm_any_lt(rsq00,rcutoff2))
 619             {
 620
 621             r00              = _mm_mul_ps(rsq00,rinv00);
 622             r00              = _mm_andnot_ps(dummy_mask,r00);
 623
 624             /* Compute parameters for interactions between i and j atoms */
 625             qq00             = _mm_mul_ps(iq0,jq0);
 626
 627             /* EWALD ELECTROSTATICS */
 628
 629             /* Analytical PME correction */
 630             zeta2            = _mm_mul_ps(beta2,rsq00);
 631             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 632             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 633             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 634             felec            = _mm_mul_ps(qq00,felec);
 635
 636             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 637
 638             fscal            = felec;
 639
 640             fscal            = _mm_and_ps(fscal,cutoff_mask);
 641
 642             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 643
 644              /* Update vectorial force */
 645             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 646             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 647             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 648
 649             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 650             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 651             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 652             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 653             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 654                                                    _mm_mul_ps(dx00,fscal),
 655                                                    _mm_mul_ps(dy00,fscal),
 656                                                    _mm_mul_ps(dz00,fscal));
 657
 658             }
 659
 660             /* Inner loop uses 32 flops */
 661         }
 662
 663         /* End of innermost loop */
 664
 665         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 666                                               f+i_coord_offset,fshift+i_shift_offset);
 667
 668         /* Increment number of inner iterations */
 669         inneriter                  += j_index_end - j_index_start;
 670
 671         /* Outer loop uses 7 flops */
 672     }
 673
 674     /* Increment number of outer iterations */
 675     outeriter        += nri;
 676
 677     /* Update outer/inner flops */
 678
 679     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*32);
 680 }