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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128i          ewitab;
  93     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  94     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  95     real             *ewtab;
  96     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  97     real             rswitch_scalar,d_scalar;
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm_set1_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 117     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 118     beta2            = _mm_mul_ps(beta,beta);
 119     beta3            = _mm_mul_ps(beta,beta2);
 120     ewtab            = fr->ic->tabq_coul_FDV0;
 121     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 122     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 123
 124     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 125     rcutoff_scalar   = fr->rcoulomb;
 126     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 127     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 128
 129     rswitch_scalar   = fr->rcoulomb_switch;
 130     rswitch          = _mm_set1_ps(rswitch_scalar);
 131     /* Setup switch parameters */
 132     d_scalar         = rcutoff_scalar-rswitch_scalar;
 133     d                = _mm_set1_ps(d_scalar);
 134     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 135     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 136     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 137     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 138     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = jnrC = jnrD = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145     j_coord_offsetC = 0;
 146     j_coord_offsetD = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     for(iidx=0;iidx<4*DIM;iidx++)
 152     {
 153         scratch[iidx] = 0.0;
 154     }
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 172
 173         fix0             = _mm_setzero_ps();
 174         fiy0             = _mm_setzero_ps();
 175         fiz0             = _mm_setzero_ps();
 176
 177         /* Load parameters for i particles */
 178         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 179
 180         /* Reset potential sums */
 181         velecsum         = _mm_setzero_ps();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               x+j_coord_offsetC,x+j_coord_offsetD,
 200                                               &jx0,&jy0,&jz0);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm_sub_ps(ix0,jx0);
 204             dy00             = _mm_sub_ps(iy0,jy0);
 205             dz00             = _mm_sub_ps(iz0,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 209
 210             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 211
 212             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 213
 214             /* Load parameters for j particles */
 215             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 216                                                               charge+jnrC+0,charge+jnrD+0);
 217
 218             /**************************
 219              * CALCULATE INTERACTIONS *
 220              **************************/
 221
 222             if (gmx_mm_any_lt(rsq00,rcutoff2))
 223             {
 224
 225             r00              = _mm_mul_ps(rsq00,rinv00);
 226
 227             /* Compute parameters for interactions between i and j atoms */
 228             qq00             = _mm_mul_ps(iq0,jq0);
 229
 230             /* EWALD ELECTROSTATICS */
 231
 232             /* Analytical PME correction */
 233             zeta2            = _mm_mul_ps(beta2,rsq00);
 234             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 235             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 236             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 237             felec            = _mm_mul_ps(qq00,felec);
 238             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 239             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 240             velec            = _mm_mul_ps(qq00,velec);
 241
 242             d                = _mm_sub_ps(r00,rswitch);
 243             d                = _mm_max_ps(d,_mm_setzero_ps());
 244             d2               = _mm_mul_ps(d,d);
 245             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 246
 247             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 248
 249             /* Evaluate switch function */
 250             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 251             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 252             velec            = _mm_mul_ps(velec,sw);
 253             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 254
 255             /* Update potential sum for this i atom from the interaction with this j atom. */
 256             velec            = _mm_and_ps(velec,cutoff_mask);
 257             velecsum         = _mm_add_ps(velecsum,velec);
 258
 259             fscal            = felec;
 260
 261             fscal            = _mm_and_ps(fscal,cutoff_mask);
 262
 263              /* Update vectorial force */
 264             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 265             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 266             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 267
 268             fjptrA             = f+j_coord_offsetA;
 269             fjptrB             = f+j_coord_offsetB;
 270             fjptrC             = f+j_coord_offsetC;
 271             fjptrD             = f+j_coord_offsetD;
 272             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 273                                                    _mm_mul_ps(dx00,fscal),
 274                                                    _mm_mul_ps(dy00,fscal),
 275                                                    _mm_mul_ps(dz00,fscal));
 276
 277             }
 278
 279             /* Inner loop uses 53 flops */
 280         }
 281
 282         if(jidx<j_index_end)
 283         {
 284
 285             /* Get j neighbor index, and coordinate index */
 286             jnrlistA         = jjnr[jidx];
 287             jnrlistB         = jjnr[jidx+1];
 288             jnrlistC         = jjnr[jidx+2];
 289             jnrlistD         = jjnr[jidx+3];
 290             /* Sign of each element will be negative for non-real atoms.
 291              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 292              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 293              */
 294             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 295             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 296             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 297             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 298             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 299             j_coord_offsetA  = DIM*jnrA;
 300             j_coord_offsetB  = DIM*jnrB;
 301             j_coord_offsetC  = DIM*jnrC;
 302             j_coord_offsetD  = DIM*jnrD;
 303
 304             /* load j atom coordinates */
 305             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 306                                               x+j_coord_offsetC,x+j_coord_offsetD,
 307                                               &jx0,&jy0,&jz0);
 308
 309             /* Calculate displacement vector */
 310             dx00             = _mm_sub_ps(ix0,jx0);
 311             dy00             = _mm_sub_ps(iy0,jy0);
 312             dz00             = _mm_sub_ps(iz0,jz0);
 313
 314             /* Calculate squared distance and things based on it */
 315             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 316
 317             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 318
 319             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 320
 321             /* Load parameters for j particles */
 322             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 323                                                               charge+jnrC+0,charge+jnrD+0);
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             if (gmx_mm_any_lt(rsq00,rcutoff2))
 330             {
 331
 332             r00              = _mm_mul_ps(rsq00,rinv00);
 333             r00              = _mm_andnot_ps(dummy_mask,r00);
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq00             = _mm_mul_ps(iq0,jq0);
 337
 338             /* EWALD ELECTROSTATICS */
 339
 340             /* Analytical PME correction */
 341             zeta2            = _mm_mul_ps(beta2,rsq00);
 342             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 343             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 344             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 345             felec            = _mm_mul_ps(qq00,felec);
 346             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 347             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 348             velec            = _mm_mul_ps(qq00,velec);
 349
 350             d                = _mm_sub_ps(r00,rswitch);
 351             d                = _mm_max_ps(d,_mm_setzero_ps());
 352             d2               = _mm_mul_ps(d,d);
 353             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 354
 355             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 356
 357             /* Evaluate switch function */
 358             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 359             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 360             velec            = _mm_mul_ps(velec,sw);
 361             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velec            = _mm_and_ps(velec,cutoff_mask);
 365             velec            = _mm_andnot_ps(dummy_mask,velec);
 366             velecsum         = _mm_add_ps(velecsum,velec);
 367
 368             fscal            = felec;
 369
 370             fscal            = _mm_and_ps(fscal,cutoff_mask);
 371
 372             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 373
 374              /* Update vectorial force */
 375             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 376             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 377             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 378
 379             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 380             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 381             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 382             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 383             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 384                                                    _mm_mul_ps(dx00,fscal),
 385                                                    _mm_mul_ps(dy00,fscal),
 386                                                    _mm_mul_ps(dz00,fscal));
 387
 388             }
 389
 390             /* Inner loop uses 54 flops */
 391         }
 392
 393         /* End of innermost loop */
 394
 395         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 396                                               f+i_coord_offset,fshift+i_shift_offset);
 397
 398         ggid                        = gid[iidx];
 399         /* Update potential energies */
 400         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 401
 402         /* Increment number of inner iterations */
 403         inneriter                  += j_index_end - j_index_start;
 404
 405         /* Outer loop uses 8 flops */
 406     }
 407
 408     /* Increment number of outer iterations */
 409     outeriter        += nri;
 410
 411     /* Update outer/inner flops */
 412
 413     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*54);
 414 }
 415 /*
 416  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 417  * Electrostatics interaction: Ewald
 418  * VdW interaction:            None
 419  * Geometry:                   Particle-Particle
 420  * Calculate force/pot:        Force
 421  */
 422 void
 423 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_128_fma_single
 424                     (t_nblist                    * gmx_restrict       nlist,
 425                      rvec                        * gmx_restrict          xx,
 426                      rvec                        * gmx_restrict          ff,
 427                      t_forcerec                  * gmx_restrict          fr,
 428                      t_mdatoms                   * gmx_restrict     mdatoms,
 429                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 430                      t_nrnb                      * gmx_restrict        nrnb)
 431 {
 432     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 433      * just 0 for non-waters.
 434      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 435      * jnr indices corresponding to data put in the four positions in the SIMD register.
 436      */
 437     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 438     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 439     int              jnrA,jnrB,jnrC,jnrD;
 440     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 441     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 442     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 443     real             rcutoff_scalar;
 444     real             *shiftvec,*fshift,*x,*f;
 445     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 446     real             scratch[4*DIM];
 447     __m128           fscal,rcutoff,rcutoff2,jidxall;
 448     int              vdwioffset0;
 449     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 450     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 451     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 452     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 453     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 454     real             *charge;
 455     __m128i          ewitab;
 456     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 457     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 458     real             *ewtab;
 459     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 460     real             rswitch_scalar,d_scalar;
 461     __m128           dummy_mask,cutoff_mask;
 462     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 463     __m128           one     = _mm_set1_ps(1.0);
 464     __m128           two     = _mm_set1_ps(2.0);
 465     x                = xx[0];
 466     f                = ff[0];
 467
 468     nri              = nlist->nri;
 469     iinr             = nlist->iinr;
 470     jindex           = nlist->jindex;
 471     jjnr             = nlist->jjnr;
 472     shiftidx         = nlist->shift;
 473     gid              = nlist->gid;
 474     shiftvec         = fr->shift_vec[0];
 475     fshift           = fr->fshift[0];
 476     facel            = _mm_set1_ps(fr->epsfac);
 477     charge           = mdatoms->chargeA;
 478
 479     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 480     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 481     beta2            = _mm_mul_ps(beta,beta);
 482     beta3            = _mm_mul_ps(beta,beta2);
 483     ewtab            = fr->ic->tabq_coul_FDV0;
 484     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 485     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 486
 487     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 488     rcutoff_scalar   = fr->rcoulomb;
 489     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 490     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 491
 492     rswitch_scalar   = fr->rcoulomb_switch;
 493     rswitch          = _mm_set1_ps(rswitch_scalar);
 494     /* Setup switch parameters */
 495     d_scalar         = rcutoff_scalar-rswitch_scalar;
 496     d                = _mm_set1_ps(d_scalar);
 497     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 498     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 499     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 500     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 501     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 502     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 503
 504     /* Avoid stupid compiler warnings */
 505     jnrA = jnrB = jnrC = jnrD = 0;
 506     j_coord_offsetA = 0;
 507     j_coord_offsetB = 0;
 508     j_coord_offsetC = 0;
 509     j_coord_offsetD = 0;
 510
 511     outeriter        = 0;
 512     inneriter        = 0;
 513
 514     for(iidx=0;iidx<4*DIM;iidx++)
 515     {
 516         scratch[iidx] = 0.0;
 517     }
 518
 519     /* Start outer loop over neighborlists */
 520     for(iidx=0; iidx<nri; iidx++)
 521     {
 522         /* Load shift vector for this list */
 523         i_shift_offset   = DIM*shiftidx[iidx];
 524
 525         /* Load limits for loop over neighbors */
 526         j_index_start    = jindex[iidx];
 527         j_index_end      = jindex[iidx+1];
 528
 529         /* Get outer coordinate index */
 530         inr              = iinr[iidx];
 531         i_coord_offset   = DIM*inr;
 532
 533         /* Load i particle coords and add shift vector */
 534         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 535
 536         fix0             = _mm_setzero_ps();
 537         fiy0             = _mm_setzero_ps();
 538         fiz0             = _mm_setzero_ps();
 539
 540         /* Load parameters for i particles */
 541         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 542
 543         /* Start inner kernel loop */
 544         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 545         {
 546
 547             /* Get j neighbor index, and coordinate index */
 548             jnrA             = jjnr[jidx];
 549             jnrB             = jjnr[jidx+1];
 550             jnrC             = jjnr[jidx+2];
 551             jnrD             = jjnr[jidx+3];
 552             j_coord_offsetA  = DIM*jnrA;
 553             j_coord_offsetB  = DIM*jnrB;
 554             j_coord_offsetC  = DIM*jnrC;
 555             j_coord_offsetD  = DIM*jnrD;
 556
 557             /* load j atom coordinates */
 558             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 559                                               x+j_coord_offsetC,x+j_coord_offsetD,
 560                                               &jx0,&jy0,&jz0);
 561
 562             /* Calculate displacement vector */
 563             dx00             = _mm_sub_ps(ix0,jx0);
 564             dy00             = _mm_sub_ps(iy0,jy0);
 565             dz00             = _mm_sub_ps(iz0,jz0);
 566
 567             /* Calculate squared distance and things based on it */
 568             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 569
 570             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 571
 572             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 573
 574             /* Load parameters for j particles */
 575             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 576                                                               charge+jnrC+0,charge+jnrD+0);
 577
 578             /**************************
 579              * CALCULATE INTERACTIONS *
 580              **************************/
 581
 582             if (gmx_mm_any_lt(rsq00,rcutoff2))
 583             {
 584
 585             r00              = _mm_mul_ps(rsq00,rinv00);
 586
 587             /* Compute parameters for interactions between i and j atoms */
 588             qq00             = _mm_mul_ps(iq0,jq0);
 589
 590             /* EWALD ELECTROSTATICS */
 591
 592             /* Analytical PME correction */
 593             zeta2            = _mm_mul_ps(beta2,rsq00);
 594             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 595             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 596             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 597             felec            = _mm_mul_ps(qq00,felec);
 598             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 599             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 600             velec            = _mm_mul_ps(qq00,velec);
 601
 602             d                = _mm_sub_ps(r00,rswitch);
 603             d                = _mm_max_ps(d,_mm_setzero_ps());
 604             d2               = _mm_mul_ps(d,d);
 605             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 606
 607             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 608
 609             /* Evaluate switch function */
 610             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 611             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 612             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 613
 614             fscal            = felec;
 615
 616             fscal            = _mm_and_ps(fscal,cutoff_mask);
 617
 618              /* Update vectorial force */
 619             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 620             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 621             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 622
 623             fjptrA             = f+j_coord_offsetA;
 624             fjptrB             = f+j_coord_offsetB;
 625             fjptrC             = f+j_coord_offsetC;
 626             fjptrD             = f+j_coord_offsetD;
 627             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 628                                                    _mm_mul_ps(dx00,fscal),
 629                                                    _mm_mul_ps(dy00,fscal),
 630                                                    _mm_mul_ps(dz00,fscal));
 631
 632             }
 633
 634             /* Inner loop uses 50 flops */
 635         }
 636
 637         if(jidx<j_index_end)
 638         {
 639
 640             /* Get j neighbor index, and coordinate index */
 641             jnrlistA         = jjnr[jidx];
 642             jnrlistB         = jjnr[jidx+1];
 643             jnrlistC         = jjnr[jidx+2];
 644             jnrlistD         = jjnr[jidx+3];
 645             /* Sign of each element will be negative for non-real atoms.
 646              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 647              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 648              */
 649             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 650             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 651             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 652             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 653             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 654             j_coord_offsetA  = DIM*jnrA;
 655             j_coord_offsetB  = DIM*jnrB;
 656             j_coord_offsetC  = DIM*jnrC;
 657             j_coord_offsetD  = DIM*jnrD;
 658
 659             /* load j atom coordinates */
 660             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 661                                               x+j_coord_offsetC,x+j_coord_offsetD,
 662                                               &jx0,&jy0,&jz0);
 663
 664             /* Calculate displacement vector */
 665             dx00             = _mm_sub_ps(ix0,jx0);
 666             dy00             = _mm_sub_ps(iy0,jy0);
 667             dz00             = _mm_sub_ps(iz0,jz0);
 668
 669             /* Calculate squared distance and things based on it */
 670             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 671
 672             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 673
 674             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 675
 676             /* Load parameters for j particles */
 677             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 678                                                               charge+jnrC+0,charge+jnrD+0);
 679
 680             /**************************
 681              * CALCULATE INTERACTIONS *
 682              **************************/
 683
 684             if (gmx_mm_any_lt(rsq00,rcutoff2))
 685             {
 686
 687             r00              = _mm_mul_ps(rsq00,rinv00);
 688             r00              = _mm_andnot_ps(dummy_mask,r00);
 689
 690             /* Compute parameters for interactions between i and j atoms */
 691             qq00             = _mm_mul_ps(iq0,jq0);
 692
 693             /* EWALD ELECTROSTATICS */
 694
 695             /* Analytical PME correction */
 696             zeta2            = _mm_mul_ps(beta2,rsq00);
 697             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 698             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 699             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 700             felec            = _mm_mul_ps(qq00,felec);
 701             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 702             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 703             velec            = _mm_mul_ps(qq00,velec);
 704
 705             d                = _mm_sub_ps(r00,rswitch);
 706             d                = _mm_max_ps(d,_mm_setzero_ps());
 707             d2               = _mm_mul_ps(d,d);
 708             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 709
 710             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 711
 712             /* Evaluate switch function */
 713             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 714             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 715             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 716
 717             fscal            = felec;
 718
 719             fscal            = _mm_and_ps(fscal,cutoff_mask);
 720
 721             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 722
 723              /* Update vectorial force */
 724             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 725             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 726             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 727
 728             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 729             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 730             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 731             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 732             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 733                                                    _mm_mul_ps(dx00,fscal),
 734                                                    _mm_mul_ps(dy00,fscal),
 735                                                    _mm_mul_ps(dz00,fscal));
 736
 737             }
 738
 739             /* Inner loop uses 51 flops */
 740         }
 741
 742         /* End of innermost loop */
 743
 744         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 745                                               f+i_coord_offset,fshift+i_shift_offset);
 746
 747         /* Increment number of inner iterations */
 748         inneriter                  += j_index_end - j_index_start;
 749
 750         /* Outer loop uses 7 flops */
 751     }
 752
 753     /* Increment number of outer iterations */
 754     outeriter        += nri;
 755
 756     /* Update outer/inner flops */
 757
 758     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*51);
 759 }