src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCSTab_VdwLJ_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_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: CubicSplineTable
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCSTab_VdwLJ_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     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  95     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  98     real             *vftab;
  99     __m128           dummy_mask,cutoff_mask;
 100     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 101     __m128           one     = _mm_set1_ps(1.0);
 102     __m128           two     = _mm_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_ps(fr->ic->epsfac);
 115     charge           = mdatoms->chargeA;
 116     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119
 120     vftab            = kernel_data->table_elec->data;
 121     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 122
 123     /* Avoid stupid compiler warnings */
 124     jnrA = jnrB = jnrC = jnrD = 0;
 125     j_coord_offsetA = 0;
 126     j_coord_offsetB = 0;
 127     j_coord_offsetC = 0;
 128     j_coord_offsetD = 0;
 129
 130     outeriter        = 0;
 131     inneriter        = 0;
 132
 133     for(iidx=0;iidx<4*DIM;iidx++)
 134     {
 135         scratch[iidx] = 0.0;
 136     }
 137
 138     /* Start outer loop over neighborlists */
 139     for(iidx=0; iidx<nri; iidx++)
 140     {
 141         /* Load shift vector for this list */
 142         i_shift_offset   = DIM*shiftidx[iidx];
 143
 144         /* Load limits for loop over neighbors */
 145         j_index_start    = jindex[iidx];
 146         j_index_end      = jindex[iidx+1];
 147
 148         /* Get outer coordinate index */
 149         inr              = iinr[iidx];
 150         i_coord_offset   = DIM*inr;
 151
 152         /* Load i particle coords and add shift vector */
 153         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 154
 155         fix0             = _mm_setzero_ps();
 156         fiy0             = _mm_setzero_ps();
 157         fiz0             = _mm_setzero_ps();
 158
 159         /* Load parameters for i particles */
 160         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 161         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 162
 163         /* Reset potential sums */
 164         velecsum         = _mm_setzero_ps();
 165         vvdwsum          = _mm_setzero_ps();
 166
 167         /* Start inner kernel loop */
 168         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 169         {
 170
 171             /* Get j neighbor index, and coordinate index */
 172             jnrA             = jjnr[jidx];
 173             jnrB             = jjnr[jidx+1];
 174             jnrC             = jjnr[jidx+2];
 175             jnrD             = jjnr[jidx+3];
 176             j_coord_offsetA  = DIM*jnrA;
 177             j_coord_offsetB  = DIM*jnrB;
 178             j_coord_offsetC  = DIM*jnrC;
 179             j_coord_offsetD  = DIM*jnrD;
 180
 181             /* load j atom coordinates */
 182             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 183                                               x+j_coord_offsetC,x+j_coord_offsetD,
 184                                               &jx0,&jy0,&jz0);
 185
 186             /* Calculate displacement vector */
 187             dx00             = _mm_sub_ps(ix0,jx0);
 188             dy00             = _mm_sub_ps(iy0,jy0);
 189             dz00             = _mm_sub_ps(iz0,jz0);
 190
 191             /* Calculate squared distance and things based on it */
 192             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 193
 194             rinv00           = avx128fma_invsqrt_f(rsq00);
 195
 196             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 197
 198             /* Load parameters for j particles */
 199             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 200                                                               charge+jnrC+0,charge+jnrD+0);
 201             vdwjidx0A        = 2*vdwtype[jnrA+0];
 202             vdwjidx0B        = 2*vdwtype[jnrB+0];
 203             vdwjidx0C        = 2*vdwtype[jnrC+0];
 204             vdwjidx0D        = 2*vdwtype[jnrD+0];
 205
 206             /**************************
 207              * CALCULATE INTERACTIONS *
 208              **************************/
 209
 210             r00              = _mm_mul_ps(rsq00,rinv00);
 211
 212             /* Compute parameters for interactions between i and j atoms */
 213             qq00             = _mm_mul_ps(iq0,jq0);
 214             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 215                                          vdwparam+vdwioffset0+vdwjidx0B,
 216                                          vdwparam+vdwioffset0+vdwjidx0C,
 217                                          vdwparam+vdwioffset0+vdwjidx0D,
 218                                          &c6_00,&c12_00);
 219
 220             /* Calculate table index by multiplying r with table scale and truncate to integer */
 221             rt               = _mm_mul_ps(r00,vftabscale);
 222             vfitab           = _mm_cvttps_epi32(rt);
 223 #ifdef __XOP__
 224             vfeps            = _mm_frcz_ps(rt);
 225 #else
 226             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 227 #endif
 228             twovfeps         = _mm_add_ps(vfeps,vfeps);
 229             vfitab           = _mm_slli_epi32(vfitab,2);
 230
 231             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 232             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 233             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 234             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 235             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 236             _MM_TRANSPOSE4_PS(Y,F,G,H);
 237             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 238             VV               = _mm_macc_ps(vfeps,Fp,Y);
 239             velec            = _mm_mul_ps(qq00,VV);
 240             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 241             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 242
 243             /* LENNARD-JONES DISPERSION/REPULSION */
 244
 245             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 246             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 247             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 248             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 249             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 250
 251             /* Update potential sum for this i atom from the interaction with this j atom. */
 252             velecsum         = _mm_add_ps(velecsum,velec);
 253             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 254
 255             fscal            = _mm_add_ps(felec,fvdw);
 256
 257              /* Update vectorial force */
 258             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 259             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 260             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 261
 262             fjptrA             = f+j_coord_offsetA;
 263             fjptrB             = f+j_coord_offsetB;
 264             fjptrC             = f+j_coord_offsetC;
 265             fjptrD             = f+j_coord_offsetD;
 266             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 267                                                    _mm_mul_ps(dx00,fscal),
 268                                                    _mm_mul_ps(dy00,fscal),
 269                                                    _mm_mul_ps(dz00,fscal));
 270
 271             /* Inner loop uses 59 flops */
 272         }
 273
 274         if(jidx<j_index_end)
 275         {
 276
 277             /* Get j neighbor index, and coordinate index */
 278             jnrlistA         = jjnr[jidx];
 279             jnrlistB         = jjnr[jidx+1];
 280             jnrlistC         = jjnr[jidx+2];
 281             jnrlistD         = jjnr[jidx+3];
 282             /* Sign of each element will be negative for non-real atoms.
 283              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 284              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 285              */
 286             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 287             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 288             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 289             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 290             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 291             j_coord_offsetA  = DIM*jnrA;
 292             j_coord_offsetB  = DIM*jnrB;
 293             j_coord_offsetC  = DIM*jnrC;
 294             j_coord_offsetD  = DIM*jnrD;
 295
 296             /* load j atom coordinates */
 297             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 298                                               x+j_coord_offsetC,x+j_coord_offsetD,
 299                                               &jx0,&jy0,&jz0);
 300
 301             /* Calculate displacement vector */
 302             dx00             = _mm_sub_ps(ix0,jx0);
 303             dy00             = _mm_sub_ps(iy0,jy0);
 304             dz00             = _mm_sub_ps(iz0,jz0);
 305
 306             /* Calculate squared distance and things based on it */
 307             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 308
 309             rinv00           = avx128fma_invsqrt_f(rsq00);
 310
 311             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 312
 313             /* Load parameters for j particles */
 314             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 315                                                               charge+jnrC+0,charge+jnrD+0);
 316             vdwjidx0A        = 2*vdwtype[jnrA+0];
 317             vdwjidx0B        = 2*vdwtype[jnrB+0];
 318             vdwjidx0C        = 2*vdwtype[jnrC+0];
 319             vdwjidx0D        = 2*vdwtype[jnrD+0];
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             r00              = _mm_mul_ps(rsq00,rinv00);
 326             r00              = _mm_andnot_ps(dummy_mask,r00);
 327
 328             /* Compute parameters for interactions between i and j atoms */
 329             qq00             = _mm_mul_ps(iq0,jq0);
 330             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 331                                          vdwparam+vdwioffset0+vdwjidx0B,
 332                                          vdwparam+vdwioffset0+vdwjidx0C,
 333                                          vdwparam+vdwioffset0+vdwjidx0D,
 334                                          &c6_00,&c12_00);
 335
 336             /* Calculate table index by multiplying r with table scale and truncate to integer */
 337             rt               = _mm_mul_ps(r00,vftabscale);
 338             vfitab           = _mm_cvttps_epi32(rt);
 339 #ifdef __XOP__
 340             vfeps            = _mm_frcz_ps(rt);
 341 #else
 342             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 343 #endif
 344             twovfeps         = _mm_add_ps(vfeps,vfeps);
 345             vfitab           = _mm_slli_epi32(vfitab,2);
 346
 347             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 348             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 349             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 350             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 351             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 352             _MM_TRANSPOSE4_PS(Y,F,G,H);
 353             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 354             VV               = _mm_macc_ps(vfeps,Fp,Y);
 355             velec            = _mm_mul_ps(qq00,VV);
 356             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 357             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 358
 359             /* LENNARD-JONES DISPERSION/REPULSION */
 360
 361             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 362             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 363             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 364             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 365             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm_andnot_ps(dummy_mask,velec);
 369             velecsum         = _mm_add_ps(velecsum,velec);
 370             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 371             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 372
 373             fscal            = _mm_add_ps(felec,fvdw);
 374
 375             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 376
 377              /* Update vectorial force */
 378             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 379             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 380             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 381
 382             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 383             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 384             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 385             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 386             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 387                                                    _mm_mul_ps(dx00,fscal),
 388                                                    _mm_mul_ps(dy00,fscal),
 389                                                    _mm_mul_ps(dz00,fscal));
 390
 391             /* Inner loop uses 60 flops */
 392         }
 393
 394         /* End of innermost loop */
 395
 396         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 397                                               f+i_coord_offset,fshift+i_shift_offset);
 398
 399         ggid                        = gid[iidx];
 400         /* Update potential energies */
 401         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 402         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 403
 404         /* Increment number of inner iterations */
 405         inneriter                  += j_index_end - j_index_start;
 406
 407         /* Outer loop uses 9 flops */
 408     }
 409
 410     /* Increment number of outer iterations */
 411     outeriter        += nri;
 412
 413     /* Update outer/inner flops */
 414
 415     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*60);
 416 }
 417 /*
 418  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
 419  * Electrostatics interaction: CubicSplineTable
 420  * VdW interaction:            LennardJones
 421  * Geometry:                   Particle-Particle
 422  * Calculate force/pot:        Force
 423  */
 424 void
 425 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
 426                     (t_nblist                    * gmx_restrict       nlist,
 427                      rvec                        * gmx_restrict          xx,
 428                      rvec                        * gmx_restrict          ff,
 429                      struct t_forcerec           * gmx_restrict          fr,
 430                      t_mdatoms                   * gmx_restrict     mdatoms,
 431                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 432                      t_nrnb                      * gmx_restrict        nrnb)
 433 {
 434     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 435      * just 0 for non-waters.
 436      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 437      * jnr indices corresponding to data put in the four positions in the SIMD register.
 438      */
 439     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 440     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 441     int              jnrA,jnrB,jnrC,jnrD;
 442     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 443     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 444     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 445     real             rcutoff_scalar;
 446     real             *shiftvec,*fshift,*x,*f;
 447     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 448     real             scratch[4*DIM];
 449     __m128           fscal,rcutoff,rcutoff2,jidxall;
 450     int              vdwioffset0;
 451     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 452     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 453     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 454     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 455     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 456     real             *charge;
 457     int              nvdwtype;
 458     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 459     int              *vdwtype;
 460     real             *vdwparam;
 461     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 462     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 463     __m128i          vfitab;
 464     __m128i          ifour       = _mm_set1_epi32(4);
 465     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 466     real             *vftab;
 467     __m128           dummy_mask,cutoff_mask;
 468     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 469     __m128           one     = _mm_set1_ps(1.0);
 470     __m128           two     = _mm_set1_ps(2.0);
 471     x                = xx[0];
 472     f                = ff[0];
 473
 474     nri              = nlist->nri;
 475     iinr             = nlist->iinr;
 476     jindex           = nlist->jindex;
 477     jjnr             = nlist->jjnr;
 478     shiftidx         = nlist->shift;
 479     gid              = nlist->gid;
 480     shiftvec         = fr->shift_vec[0];
 481     fshift           = fr->fshift[0];
 482     facel            = _mm_set1_ps(fr->ic->epsfac);
 483     charge           = mdatoms->chargeA;
 484     nvdwtype         = fr->ntype;
 485     vdwparam         = fr->nbfp;
 486     vdwtype          = mdatoms->typeA;
 487
 488     vftab            = kernel_data->table_elec->data;
 489     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 490
 491     /* Avoid stupid compiler warnings */
 492     jnrA = jnrB = jnrC = jnrD = 0;
 493     j_coord_offsetA = 0;
 494     j_coord_offsetB = 0;
 495     j_coord_offsetC = 0;
 496     j_coord_offsetD = 0;
 497
 498     outeriter        = 0;
 499     inneriter        = 0;
 500
 501     for(iidx=0;iidx<4*DIM;iidx++)
 502     {
 503         scratch[iidx] = 0.0;
 504     }
 505
 506     /* Start outer loop over neighborlists */
 507     for(iidx=0; iidx<nri; iidx++)
 508     {
 509         /* Load shift vector for this list */
 510         i_shift_offset   = DIM*shiftidx[iidx];
 511
 512         /* Load limits for loop over neighbors */
 513         j_index_start    = jindex[iidx];
 514         j_index_end      = jindex[iidx+1];
 515
 516         /* Get outer coordinate index */
 517         inr              = iinr[iidx];
 518         i_coord_offset   = DIM*inr;
 519
 520         /* Load i particle coords and add shift vector */
 521         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 522
 523         fix0             = _mm_setzero_ps();
 524         fiy0             = _mm_setzero_ps();
 525         fiz0             = _mm_setzero_ps();
 526
 527         /* Load parameters for i particles */
 528         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 529         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 530
 531         /* Start inner kernel loop */
 532         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 533         {
 534
 535             /* Get j neighbor index, and coordinate index */
 536             jnrA             = jjnr[jidx];
 537             jnrB             = jjnr[jidx+1];
 538             jnrC             = jjnr[jidx+2];
 539             jnrD             = jjnr[jidx+3];
 540             j_coord_offsetA  = DIM*jnrA;
 541             j_coord_offsetB  = DIM*jnrB;
 542             j_coord_offsetC  = DIM*jnrC;
 543             j_coord_offsetD  = DIM*jnrD;
 544
 545             /* load j atom coordinates */
 546             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 547                                               x+j_coord_offsetC,x+j_coord_offsetD,
 548                                               &jx0,&jy0,&jz0);
 549
 550             /* Calculate displacement vector */
 551             dx00             = _mm_sub_ps(ix0,jx0);
 552             dy00             = _mm_sub_ps(iy0,jy0);
 553             dz00             = _mm_sub_ps(iz0,jz0);
 554
 555             /* Calculate squared distance and things based on it */
 556             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 557
 558             rinv00           = avx128fma_invsqrt_f(rsq00);
 559
 560             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 561
 562             /* Load parameters for j particles */
 563             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 564                                                               charge+jnrC+0,charge+jnrD+0);
 565             vdwjidx0A        = 2*vdwtype[jnrA+0];
 566             vdwjidx0B        = 2*vdwtype[jnrB+0];
 567             vdwjidx0C        = 2*vdwtype[jnrC+0];
 568             vdwjidx0D        = 2*vdwtype[jnrD+0];
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             r00              = _mm_mul_ps(rsq00,rinv00);
 575
 576             /* Compute parameters for interactions between i and j atoms */
 577             qq00             = _mm_mul_ps(iq0,jq0);
 578             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 579                                          vdwparam+vdwioffset0+vdwjidx0B,
 580                                          vdwparam+vdwioffset0+vdwjidx0C,
 581                                          vdwparam+vdwioffset0+vdwjidx0D,
 582                                          &c6_00,&c12_00);
 583
 584             /* Calculate table index by multiplying r with table scale and truncate to integer */
 585             rt               = _mm_mul_ps(r00,vftabscale);
 586             vfitab           = _mm_cvttps_epi32(rt);
 587 #ifdef __XOP__
 588             vfeps            = _mm_frcz_ps(rt);
 589 #else
 590             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 591 #endif
 592             twovfeps         = _mm_add_ps(vfeps,vfeps);
 593             vfitab           = _mm_slli_epi32(vfitab,2);
 594
 595             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 596             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 597             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 598             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 599             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 600             _MM_TRANSPOSE4_PS(Y,F,G,H);
 601             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 602             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 603             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 604
 605             /* LENNARD-JONES DISPERSION/REPULSION */
 606
 607             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 608             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 609
 610             fscal            = _mm_add_ps(felec,fvdw);
 611
 612              /* Update vectorial force */
 613             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 614             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 615             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 616
 617             fjptrA             = f+j_coord_offsetA;
 618             fjptrB             = f+j_coord_offsetB;
 619             fjptrC             = f+j_coord_offsetC;
 620             fjptrD             = f+j_coord_offsetD;
 621             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 622                                                    _mm_mul_ps(dx00,fscal),
 623                                                    _mm_mul_ps(dy00,fscal),
 624                                                    _mm_mul_ps(dz00,fscal));
 625
 626             /* Inner loop uses 50 flops */
 627         }
 628
 629         if(jidx<j_index_end)
 630         {
 631
 632             /* Get j neighbor index, and coordinate index */
 633             jnrlistA         = jjnr[jidx];
 634             jnrlistB         = jjnr[jidx+1];
 635             jnrlistC         = jjnr[jidx+2];
 636             jnrlistD         = jjnr[jidx+3];
 637             /* Sign of each element will be negative for non-real atoms.
 638              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 639              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 640              */
 641             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 642             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 643             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 644             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 645             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 646             j_coord_offsetA  = DIM*jnrA;
 647             j_coord_offsetB  = DIM*jnrB;
 648             j_coord_offsetC  = DIM*jnrC;
 649             j_coord_offsetD  = DIM*jnrD;
 650
 651             /* load j atom coordinates */
 652             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 653                                               x+j_coord_offsetC,x+j_coord_offsetD,
 654                                               &jx0,&jy0,&jz0);
 655
 656             /* Calculate displacement vector */
 657             dx00             = _mm_sub_ps(ix0,jx0);
 658             dy00             = _mm_sub_ps(iy0,jy0);
 659             dz00             = _mm_sub_ps(iz0,jz0);
 660
 661             /* Calculate squared distance and things based on it */
 662             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 663
 664             rinv00           = avx128fma_invsqrt_f(rsq00);
 665
 666             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 667
 668             /* Load parameters for j particles */
 669             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 670                                                               charge+jnrC+0,charge+jnrD+0);
 671             vdwjidx0A        = 2*vdwtype[jnrA+0];
 672             vdwjidx0B        = 2*vdwtype[jnrB+0];
 673             vdwjidx0C        = 2*vdwtype[jnrC+0];
 674             vdwjidx0D        = 2*vdwtype[jnrD+0];
 675
 676             /**************************
 677              * CALCULATE INTERACTIONS *
 678              **************************/
 679
 680             r00              = _mm_mul_ps(rsq00,rinv00);
 681             r00              = _mm_andnot_ps(dummy_mask,r00);
 682
 683             /* Compute parameters for interactions between i and j atoms */
 684             qq00             = _mm_mul_ps(iq0,jq0);
 685             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 686                                          vdwparam+vdwioffset0+vdwjidx0B,
 687                                          vdwparam+vdwioffset0+vdwjidx0C,
 688                                          vdwparam+vdwioffset0+vdwjidx0D,
 689                                          &c6_00,&c12_00);
 690
 691             /* Calculate table index by multiplying r with table scale and truncate to integer */
 692             rt               = _mm_mul_ps(r00,vftabscale);
 693             vfitab           = _mm_cvttps_epi32(rt);
 694 #ifdef __XOP__
 695             vfeps            = _mm_frcz_ps(rt);
 696 #else
 697             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 698 #endif
 699             twovfeps         = _mm_add_ps(vfeps,vfeps);
 700             vfitab           = _mm_slli_epi32(vfitab,2);
 701
 702             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 703             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 704             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 705             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 706             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 707             _MM_TRANSPOSE4_PS(Y,F,G,H);
 708             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 709             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 710             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 711
 712             /* LENNARD-JONES DISPERSION/REPULSION */
 713
 714             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 715             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 716
 717             fscal            = _mm_add_ps(felec,fvdw);
 718
 719             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 720
 721              /* Update vectorial force */
 722             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 723             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 724             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 725
 726             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 727             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 728             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 729             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 730             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 731                                                    _mm_mul_ps(dx00,fscal),
 732                                                    _mm_mul_ps(dy00,fscal),
 733                                                    _mm_mul_ps(dz00,fscal));
 734
 735             /* Inner loop uses 51 flops */
 736         }
 737
 738         /* End of innermost loop */
 739
 740         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 741                                               f+i_coord_offset,fshift+i_shift_offset);
 742
 743         /* Increment number of inner iterations */
 744         inneriter                  += j_index_end - j_index_start;
 745
 746         /* Outer loop uses 7 flops */
 747     }
 748
 749     /* Increment number of outer iterations */
 750     outeriter        += nri;
 751
 752     /* Update outer/inner flops */
 753
 754     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*51);
 755 }