src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_avx_128_fma_single.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_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_ElecEwSw_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_ElecEwSw_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           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  94     real             rswitch_scalar,d_scalar;
  95     __m128           dummy_mask,cutoff_mask;
  96     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  97     __m128           one     = _mm_set1_ps(1.0);
  98     __m128           two     = _mm_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_ps(fr->ic->epsfac);
 111     charge           = mdatoms->chargeA;
 112
 113     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 114     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 115     beta2            = _mm_mul_ps(beta,beta);
 116     beta3            = _mm_mul_ps(beta,beta2);
 117     ewtab            = fr->ic->tabq_coul_FDV0;
 118     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 119     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 120
 121     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 122     rcutoff_scalar   = fr->ic->rcoulomb;
 123     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 124     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 125
 126     rswitch_scalar   = fr->ic->rcoulomb_switch;
 127     rswitch          = _mm_set1_ps(rswitch_scalar);
 128     /* Setup switch parameters */
 129     d_scalar         = rcutoff_scalar-rswitch_scalar;
 130     d                = _mm_set1_ps(d_scalar);
 131     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 132     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 133     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 134     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 135     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 136     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = jnrC = jnrD = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142     j_coord_offsetC = 0;
 143     j_coord_offsetD = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     for(iidx=0;iidx<4*DIM;iidx++)
 149     {
 150         scratch[iidx] = 0.0;
 151     }
 152
 153     /* Start outer loop over neighborlists */
 154     for(iidx=0; iidx<nri; iidx++)
 155     {
 156         /* Load shift vector for this list */
 157         i_shift_offset   = DIM*shiftidx[iidx];
 158
 159         /* Load limits for loop over neighbors */
 160         j_index_start    = jindex[iidx];
 161         j_index_end      = jindex[iidx+1];
 162
 163         /* Get outer coordinate index */
 164         inr              = iinr[iidx];
 165         i_coord_offset   = DIM*inr;
 166
 167         /* Load i particle coords and add shift vector */
 168         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 169
 170         fix0             = _mm_setzero_ps();
 171         fiy0             = _mm_setzero_ps();
 172         fiz0             = _mm_setzero_ps();
 173
 174         /* Load parameters for i particles */
 175         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 176
 177         /* Reset potential sums */
 178         velecsum         = _mm_setzero_ps();
 179
 180         /* Start inner kernel loop */
 181         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 182         {
 183
 184             /* Get j neighbor index, and coordinate index */
 185             jnrA             = jjnr[jidx];
 186             jnrB             = jjnr[jidx+1];
 187             jnrC             = jjnr[jidx+2];
 188             jnrD             = jjnr[jidx+3];
 189             j_coord_offsetA  = DIM*jnrA;
 190             j_coord_offsetB  = DIM*jnrB;
 191             j_coord_offsetC  = DIM*jnrC;
 192             j_coord_offsetD  = DIM*jnrD;
 193
 194             /* load j atom coordinates */
 195             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 196                                               x+j_coord_offsetC,x+j_coord_offsetD,
 197                                               &jx0,&jy0,&jz0);
 198
 199             /* Calculate displacement vector */
 200             dx00             = _mm_sub_ps(ix0,jx0);
 201             dy00             = _mm_sub_ps(iy0,jy0);
 202             dz00             = _mm_sub_ps(iz0,jz0);
 203
 204             /* Calculate squared distance and things based on it */
 205             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 206
 207             rinv00           = avx128fma_invsqrt_f(rsq00);
 208
 209             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 210
 211             /* Load parameters for j particles */
 212             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 213                                                               charge+jnrC+0,charge+jnrD+0);
 214
 215             /**************************
 216              * CALCULATE INTERACTIONS *
 217              **************************/
 218
 219             if (gmx_mm_any_lt(rsq00,rcutoff2))
 220             {
 221
 222             r00              = _mm_mul_ps(rsq00,rinv00);
 223
 224             /* Compute parameters for interactions between i and j atoms */
 225             qq00             = _mm_mul_ps(iq0,jq0);
 226
 227             /* EWALD ELECTROSTATICS */
 228
 229             /* Analytical PME correction */
 230             zeta2            = _mm_mul_ps(beta2,rsq00);
 231             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 232             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 233             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 234             felec            = _mm_mul_ps(qq00,felec);
 235             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 236             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 237             velec            = _mm_mul_ps(qq00,velec);
 238
 239             d                = _mm_sub_ps(r00,rswitch);
 240             d                = _mm_max_ps(d,_mm_setzero_ps());
 241             d2               = _mm_mul_ps(d,d);
 242             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 243
 244             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 245
 246             /* Evaluate switch function */
 247             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 248             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 249             velec            = _mm_mul_ps(velec,sw);
 250             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velec            = _mm_and_ps(velec,cutoff_mask);
 254             velecsum         = _mm_add_ps(velecsum,velec);
 255
 256             fscal            = felec;
 257
 258             fscal            = _mm_and_ps(fscal,cutoff_mask);
 259
 260              /* Update vectorial force */
 261             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 262             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 263             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 264
 265             fjptrA             = f+j_coord_offsetA;
 266             fjptrB             = f+j_coord_offsetB;
 267             fjptrC             = f+j_coord_offsetC;
 268             fjptrD             = f+j_coord_offsetD;
 269             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 270                                                    _mm_mul_ps(dx00,fscal),
 271                                                    _mm_mul_ps(dy00,fscal),
 272                                                    _mm_mul_ps(dz00,fscal));
 273
 274             }
 275
 276             /* Inner loop uses 53 flops */
 277         }
 278
 279         if(jidx<j_index_end)
 280         {
 281
 282             /* Get j neighbor index, and coordinate index */
 283             jnrlistA         = jjnr[jidx];
 284             jnrlistB         = jjnr[jidx+1];
 285             jnrlistC         = jjnr[jidx+2];
 286             jnrlistD         = jjnr[jidx+3];
 287             /* Sign of each element will be negative for non-real atoms.
 288              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 289              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 290              */
 291             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 292             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 293             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 294             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 295             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 296             j_coord_offsetA  = DIM*jnrA;
 297             j_coord_offsetB  = DIM*jnrB;
 298             j_coord_offsetC  = DIM*jnrC;
 299             j_coord_offsetD  = DIM*jnrD;
 300
 301             /* load j atom coordinates */
 302             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 303                                               x+j_coord_offsetC,x+j_coord_offsetD,
 304                                               &jx0,&jy0,&jz0);
 305
 306             /* Calculate displacement vector */
 307             dx00             = _mm_sub_ps(ix0,jx0);
 308             dy00             = _mm_sub_ps(iy0,jy0);
 309             dz00             = _mm_sub_ps(iz0,jz0);
 310
 311             /* Calculate squared distance and things based on it */
 312             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 313
 314             rinv00           = avx128fma_invsqrt_f(rsq00);
 315
 316             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 317
 318             /* Load parameters for j particles */
 319             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 320                                                               charge+jnrC+0,charge+jnrD+0);
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             if (gmx_mm_any_lt(rsq00,rcutoff2))
 327             {
 328
 329             r00              = _mm_mul_ps(rsq00,rinv00);
 330             r00              = _mm_andnot_ps(dummy_mask,r00);
 331
 332             /* Compute parameters for interactions between i and j atoms */
 333             qq00             = _mm_mul_ps(iq0,jq0);
 334
 335             /* EWALD ELECTROSTATICS */
 336
 337             /* Analytical PME correction */
 338             zeta2            = _mm_mul_ps(beta2,rsq00);
 339             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 340             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 341             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 342             felec            = _mm_mul_ps(qq00,felec);
 343             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 344             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 345             velec            = _mm_mul_ps(qq00,velec);
 346
 347             d                = _mm_sub_ps(r00,rswitch);
 348             d                = _mm_max_ps(d,_mm_setzero_ps());
 349             d2               = _mm_mul_ps(d,d);
 350             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 351
 352             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 353
 354             /* Evaluate switch function */
 355             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 356             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 357             velec            = _mm_mul_ps(velec,sw);
 358             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             velec            = _mm_and_ps(velec,cutoff_mask);
 362             velec            = _mm_andnot_ps(dummy_mask,velec);
 363             velecsum         = _mm_add_ps(velecsum,velec);
 364
 365             fscal            = felec;
 366
 367             fscal            = _mm_and_ps(fscal,cutoff_mask);
 368
 369             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 370
 371              /* Update vectorial force */
 372             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 373             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 374             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 375
 376             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 377             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 378             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 379             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 380             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 381                                                    _mm_mul_ps(dx00,fscal),
 382                                                    _mm_mul_ps(dy00,fscal),
 383                                                    _mm_mul_ps(dz00,fscal));
 384
 385             }
 386
 387             /* Inner loop uses 54 flops */
 388         }
 389
 390         /* End of innermost loop */
 391
 392         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 393                                               f+i_coord_offset,fshift+i_shift_offset);
 394
 395         ggid                        = gid[iidx];
 396         /* Update potential energies */
 397         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 398
 399         /* Increment number of inner iterations */
 400         inneriter                  += j_index_end - j_index_start;
 401
 402         /* Outer loop uses 8 flops */
 403     }
 404
 405     /* Increment number of outer iterations */
 406     outeriter        += nri;
 407
 408     /* Update outer/inner flops */
 409
 410     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*54);
 411 }
 412 /*
 413  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 414  * Electrostatics interaction: Ewald
 415  * VdW interaction:            None
 416  * Geometry:                   Particle-Particle
 417  * Calculate force/pot:        Force
 418  */
 419 void
 420 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 421                     (t_nblist                    * gmx_restrict       nlist,
 422                      rvec                        * gmx_restrict          xx,
 423                      rvec                        * gmx_restrict          ff,
 424                      struct t_forcerec           * gmx_restrict          fr,
 425                      t_mdatoms                   * gmx_restrict     mdatoms,
 426                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 427                      t_nrnb                      * gmx_restrict        nrnb)
 428 {
 429     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 430      * just 0 for non-waters.
 431      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 432      * jnr indices corresponding to data put in the four positions in the SIMD register.
 433      */
 434     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 435     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 436     int              jnrA,jnrB,jnrC,jnrD;
 437     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 438     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 439     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 440     real             rcutoff_scalar;
 441     real             *shiftvec,*fshift,*x,*f;
 442     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 443     real             scratch[4*DIM];
 444     __m128           fscal,rcutoff,rcutoff2,jidxall;
 445     int              vdwioffset0;
 446     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 447     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 448     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 449     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 450     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 451     real             *charge;
 452     __m128i          ewitab;
 453     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 454     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 455     real             *ewtab;
 456     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 457     real             rswitch_scalar,d_scalar;
 458     __m128           dummy_mask,cutoff_mask;
 459     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 460     __m128           one     = _mm_set1_ps(1.0);
 461     __m128           two     = _mm_set1_ps(2.0);
 462     x                = xx[0];
 463     f                = ff[0];
 464
 465     nri              = nlist->nri;
 466     iinr             = nlist->iinr;
 467     jindex           = nlist->jindex;
 468     jjnr             = nlist->jjnr;
 469     shiftidx         = nlist->shift;
 470     gid              = nlist->gid;
 471     shiftvec         = fr->shift_vec[0];
 472     fshift           = fr->fshift[0];
 473     facel            = _mm_set1_ps(fr->ic->epsfac);
 474     charge           = mdatoms->chargeA;
 475
 476     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 477     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 478     beta2            = _mm_mul_ps(beta,beta);
 479     beta3            = _mm_mul_ps(beta,beta2);
 480     ewtab            = fr->ic->tabq_coul_FDV0;
 481     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 482     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 483
 484     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 485     rcutoff_scalar   = fr->ic->rcoulomb;
 486     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 487     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 488
 489     rswitch_scalar   = fr->ic->rcoulomb_switch;
 490     rswitch          = _mm_set1_ps(rswitch_scalar);
 491     /* Setup switch parameters */
 492     d_scalar         = rcutoff_scalar-rswitch_scalar;
 493     d                = _mm_set1_ps(d_scalar);
 494     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 495     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 496     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 497     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 498     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 499     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 500
 501     /* Avoid stupid compiler warnings */
 502     jnrA = jnrB = jnrC = jnrD = 0;
 503     j_coord_offsetA = 0;
 504     j_coord_offsetB = 0;
 505     j_coord_offsetC = 0;
 506     j_coord_offsetD = 0;
 507
 508     outeriter        = 0;
 509     inneriter        = 0;
 510
 511     for(iidx=0;iidx<4*DIM;iidx++)
 512     {
 513         scratch[iidx] = 0.0;
 514     }
 515
 516     /* Start outer loop over neighborlists */
 517     for(iidx=0; iidx<nri; iidx++)
 518     {
 519         /* Load shift vector for this list */
 520         i_shift_offset   = DIM*shiftidx[iidx];
 521
 522         /* Load limits for loop over neighbors */
 523         j_index_start    = jindex[iidx];
 524         j_index_end      = jindex[iidx+1];
 525
 526         /* Get outer coordinate index */
 527         inr              = iinr[iidx];
 528         i_coord_offset   = DIM*inr;
 529
 530         /* Load i particle coords and add shift vector */
 531         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 532
 533         fix0             = _mm_setzero_ps();
 534         fiy0             = _mm_setzero_ps();
 535         fiz0             = _mm_setzero_ps();
 536
 537         /* Load parameters for i particles */
 538         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 539
 540         /* Start inner kernel loop */
 541         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 542         {
 543
 544             /* Get j neighbor index, and coordinate index */
 545             jnrA             = jjnr[jidx];
 546             jnrB             = jjnr[jidx+1];
 547             jnrC             = jjnr[jidx+2];
 548             jnrD             = jjnr[jidx+3];
 549             j_coord_offsetA  = DIM*jnrA;
 550             j_coord_offsetB  = DIM*jnrB;
 551             j_coord_offsetC  = DIM*jnrC;
 552             j_coord_offsetD  = DIM*jnrD;
 553
 554             /* load j atom coordinates */
 555             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 556                                               x+j_coord_offsetC,x+j_coord_offsetD,
 557                                               &jx0,&jy0,&jz0);
 558
 559             /* Calculate displacement vector */
 560             dx00             = _mm_sub_ps(ix0,jx0);
 561             dy00             = _mm_sub_ps(iy0,jy0);
 562             dz00             = _mm_sub_ps(iz0,jz0);
 563
 564             /* Calculate squared distance and things based on it */
 565             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 566
 567             rinv00           = avx128fma_invsqrt_f(rsq00);
 568
 569             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 570
 571             /* Load parameters for j particles */
 572             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 573                                                               charge+jnrC+0,charge+jnrD+0);
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             if (gmx_mm_any_lt(rsq00,rcutoff2))
 580             {
 581
 582             r00              = _mm_mul_ps(rsq00,rinv00);
 583
 584             /* Compute parameters for interactions between i and j atoms */
 585             qq00             = _mm_mul_ps(iq0,jq0);
 586
 587             /* EWALD ELECTROSTATICS */
 588
 589             /* Analytical PME correction */
 590             zeta2            = _mm_mul_ps(beta2,rsq00);
 591             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 592             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 593             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 594             felec            = _mm_mul_ps(qq00,felec);
 595             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 596             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 597             velec            = _mm_mul_ps(qq00,velec);
 598
 599             d                = _mm_sub_ps(r00,rswitch);
 600             d                = _mm_max_ps(d,_mm_setzero_ps());
 601             d2               = _mm_mul_ps(d,d);
 602             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 603
 604             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 605
 606             /* Evaluate switch function */
 607             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 608             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 609             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 610
 611             fscal            = felec;
 612
 613             fscal            = _mm_and_ps(fscal,cutoff_mask);
 614
 615              /* Update vectorial force */
 616             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 617             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 618             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 619
 620             fjptrA             = f+j_coord_offsetA;
 621             fjptrB             = f+j_coord_offsetB;
 622             fjptrC             = f+j_coord_offsetC;
 623             fjptrD             = f+j_coord_offsetD;
 624             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 625                                                    _mm_mul_ps(dx00,fscal),
 626                                                    _mm_mul_ps(dy00,fscal),
 627                                                    _mm_mul_ps(dz00,fscal));
 628
 629             }
 630
 631             /* Inner loop uses 50 flops */
 632         }
 633
 634         if(jidx<j_index_end)
 635         {
 636
 637             /* Get j neighbor index, and coordinate index */
 638             jnrlistA         = jjnr[jidx];
 639             jnrlistB         = jjnr[jidx+1];
 640             jnrlistC         = jjnr[jidx+2];
 641             jnrlistD         = jjnr[jidx+3];
 642             /* Sign of each element will be negative for non-real atoms.
 643              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 644              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 645              */
 646             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 647             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 648             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 649             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 650             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 651             j_coord_offsetA  = DIM*jnrA;
 652             j_coord_offsetB  = DIM*jnrB;
 653             j_coord_offsetC  = DIM*jnrC;
 654             j_coord_offsetD  = DIM*jnrD;
 655
 656             /* load j atom coordinates */
 657             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 658                                               x+j_coord_offsetC,x+j_coord_offsetD,
 659                                               &jx0,&jy0,&jz0);
 660
 661             /* Calculate displacement vector */
 662             dx00             = _mm_sub_ps(ix0,jx0);
 663             dy00             = _mm_sub_ps(iy0,jy0);
 664             dz00             = _mm_sub_ps(iz0,jz0);
 665
 666             /* Calculate squared distance and things based on it */
 667             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 668
 669             rinv00           = avx128fma_invsqrt_f(rsq00);
 670
 671             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 672
 673             /* Load parameters for j particles */
 674             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 675                                                               charge+jnrC+0,charge+jnrD+0);
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             if (gmx_mm_any_lt(rsq00,rcutoff2))
 682             {
 683
 684             r00              = _mm_mul_ps(rsq00,rinv00);
 685             r00              = _mm_andnot_ps(dummy_mask,r00);
 686
 687             /* Compute parameters for interactions between i and j atoms */
 688             qq00             = _mm_mul_ps(iq0,jq0);
 689
 690             /* EWALD ELECTROSTATICS */
 691
 692             /* Analytical PME correction */
 693             zeta2            = _mm_mul_ps(beta2,rsq00);
 694             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 695             pmecorrF         = avx128fma_pmecorrF_f(zeta2);
 696             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 697             felec            = _mm_mul_ps(qq00,felec);
 698             pmecorrV         = avx128fma_pmecorrV_f(zeta2);
 699             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 700             velec            = _mm_mul_ps(qq00,velec);
 701
 702             d                = _mm_sub_ps(r00,rswitch);
 703             d                = _mm_max_ps(d,_mm_setzero_ps());
 704             d2               = _mm_mul_ps(d,d);
 705             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 706
 707             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 708
 709             /* Evaluate switch function */
 710             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 711             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 712             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 713
 714             fscal            = felec;
 715
 716             fscal            = _mm_and_ps(fscal,cutoff_mask);
 717
 718             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 719
 720              /* Update vectorial force */
 721             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 722             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 723             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 724
 725             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 726             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 727             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 728             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 729             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 730                                                    _mm_mul_ps(dx00,fscal),
 731                                                    _mm_mul_ps(dy00,fscal),
 732                                                    _mm_mul_ps(dz00,fscal));
 733
 734             }
 735
 736             /* Inner loop uses 51 flops */
 737         }
 738
 739         /* End of innermost loop */
 740
 741         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 742                                               f+i_coord_offset,fshift+i_shift_offset);
 743
 744         /* Increment number of inner iterations */
 745         inneriter                  += j_index_end - j_index_start;
 746
 747         /* Outer loop uses 7 flops */
 748     }
 749
 750     /* Increment number of outer iterations */
 751     outeriter        += nri;
 752
 753     /* Update outer/inner flops */
 754
 755     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*51);
 756 }