src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwLJEw_GeomP1P1_avx_128_fma_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_VF_avx_128_fma_single
  51  * Electrostatics interaction: None
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecNone_VdwLJEw_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     int              nvdwtype;
  88     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  89     int              *vdwtype;
  90     real             *vdwparam;
  91     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  92     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  93     __m128           c6grid_00;
  94     real             *vdwgridparam;
  95     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  96     __m128           one_half = _mm_set1_ps(0.5);
  97     __m128           minus_one = _mm_set1_ps(-1.0);
  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     nvdwtype         = fr->ntype;
 114     vdwparam         = fr->nbfp;
 115     vdwtype          = mdatoms->typeA;
 116     vdwgridparam     = fr->ljpme_c6grid;
 117     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 118     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 119     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 120
 121     /* Avoid stupid compiler warnings */
 122     jnrA = jnrB = jnrC = jnrD = 0;
 123     j_coord_offsetA = 0;
 124     j_coord_offsetB = 0;
 125     j_coord_offsetC = 0;
 126     j_coord_offsetD = 0;
 127
 128     outeriter        = 0;
 129     inneriter        = 0;
 130
 131     for(iidx=0;iidx<4*DIM;iidx++)
 132     {
 133         scratch[iidx] = 0.0;
 134     }
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141
 142         /* Load limits for loop over neighbors */
 143         j_index_start    = jindex[iidx];
 144         j_index_end      = jindex[iidx+1];
 145
 146         /* Get outer coordinate index */
 147         inr              = iinr[iidx];
 148         i_coord_offset   = DIM*inr;
 149
 150         /* Load i particle coords and add shift vector */
 151         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 152
 153         fix0             = _mm_setzero_ps();
 154         fiy0             = _mm_setzero_ps();
 155         fiz0             = _mm_setzero_ps();
 156
 157         /* Load parameters for i particles */
 158         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 159
 160         /* Reset potential sums */
 161         vvdwsum          = _mm_setzero_ps();
 162
 163         /* Start inner kernel loop */
 164         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 165         {
 166
 167             /* Get j neighbor index, and coordinate index */
 168             jnrA             = jjnr[jidx];
 169             jnrB             = jjnr[jidx+1];
 170             jnrC             = jjnr[jidx+2];
 171             jnrD             = jjnr[jidx+3];
 172             j_coord_offsetA  = DIM*jnrA;
 173             j_coord_offsetB  = DIM*jnrB;
 174             j_coord_offsetC  = DIM*jnrC;
 175             j_coord_offsetD  = DIM*jnrD;
 176
 177             /* load j atom coordinates */
 178             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 179                                               x+j_coord_offsetC,x+j_coord_offsetD,
 180                                               &jx0,&jy0,&jz0);
 181
 182             /* Calculate displacement vector */
 183             dx00             = _mm_sub_ps(ix0,jx0);
 184             dy00             = _mm_sub_ps(iy0,jy0);
 185             dz00             = _mm_sub_ps(iz0,jz0);
 186
 187             /* Calculate squared distance and things based on it */
 188             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 189
 190             rinv00           = avx128fma_invsqrt_f(rsq00);
 191
 192             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 193
 194             /* Load parameters for j particles */
 195             vdwjidx0A        = 2*vdwtype[jnrA+0];
 196             vdwjidx0B        = 2*vdwtype[jnrB+0];
 197             vdwjidx0C        = 2*vdwtype[jnrC+0];
 198             vdwjidx0D        = 2*vdwtype[jnrD+0];
 199
 200             /**************************
 201              * CALCULATE INTERACTIONS *
 202              **************************/
 203
 204             r00              = _mm_mul_ps(rsq00,rinv00);
 205
 206             /* Compute parameters for interactions between i and j atoms */
 207             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 208                                          vdwparam+vdwioffset0+vdwjidx0B,
 209                                          vdwparam+vdwioffset0+vdwjidx0C,
 210                                          vdwparam+vdwioffset0+vdwjidx0D,
 211                                          &c6_00,&c12_00);
 212
 213             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 214                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 215                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 216                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 217
 218             /* Analytical LJ-PME */
 219             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 220             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 221             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 222             exponent         = avx128fma_exp_f(ewcljrsq);
 223             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 224             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 225             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 226             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 227             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 228             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 229             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 230             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 231
 232             /* Update potential sum for this i atom from the interaction with this j atom. */
 233             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 234
 235             fscal            = fvdw;
 236
 237              /* Update vectorial force */
 238             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 239             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 240             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 241
 242             fjptrA             = f+j_coord_offsetA;
 243             fjptrB             = f+j_coord_offsetB;
 244             fjptrC             = f+j_coord_offsetC;
 245             fjptrD             = f+j_coord_offsetD;
 246             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 247                                                    _mm_mul_ps(dx00,fscal),
 248                                                    _mm_mul_ps(dy00,fscal),
 249                                                    _mm_mul_ps(dz00,fscal));
 250
 251             /* Inner loop uses 50 flops */
 252         }
 253
 254         if(jidx<j_index_end)
 255         {
 256
 257             /* Get j neighbor index, and coordinate index */
 258             jnrlistA         = jjnr[jidx];
 259             jnrlistB         = jjnr[jidx+1];
 260             jnrlistC         = jjnr[jidx+2];
 261             jnrlistD         = jjnr[jidx+3];
 262             /* Sign of each element will be negative for non-real atoms.
 263              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 264              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 265              */
 266             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 267             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 268             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 269             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 270             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 271             j_coord_offsetA  = DIM*jnrA;
 272             j_coord_offsetB  = DIM*jnrB;
 273             j_coord_offsetC  = DIM*jnrC;
 274             j_coord_offsetD  = DIM*jnrD;
 275
 276             /* load j atom coordinates */
 277             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 278                                               x+j_coord_offsetC,x+j_coord_offsetD,
 279                                               &jx0,&jy0,&jz0);
 280
 281             /* Calculate displacement vector */
 282             dx00             = _mm_sub_ps(ix0,jx0);
 283             dy00             = _mm_sub_ps(iy0,jy0);
 284             dz00             = _mm_sub_ps(iz0,jz0);
 285
 286             /* Calculate squared distance and things based on it */
 287             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 288
 289             rinv00           = avx128fma_invsqrt_f(rsq00);
 290
 291             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 292
 293             /* Load parameters for j particles */
 294             vdwjidx0A        = 2*vdwtype[jnrA+0];
 295             vdwjidx0B        = 2*vdwtype[jnrB+0];
 296             vdwjidx0C        = 2*vdwtype[jnrC+0];
 297             vdwjidx0D        = 2*vdwtype[jnrD+0];
 298
 299             /**************************
 300              * CALCULATE INTERACTIONS *
 301              **************************/
 302
 303             r00              = _mm_mul_ps(rsq00,rinv00);
 304             r00              = _mm_andnot_ps(dummy_mask,r00);
 305
 306             /* Compute parameters for interactions between i and j atoms */
 307             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 308                                          vdwparam+vdwioffset0+vdwjidx0B,
 309                                          vdwparam+vdwioffset0+vdwjidx0C,
 310                                          vdwparam+vdwioffset0+vdwjidx0D,
 311                                          &c6_00,&c12_00);
 312
 313             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 314                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 315                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 316                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 317
 318             /* Analytical LJ-PME */
 319             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 320             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 321             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 322             exponent         = avx128fma_exp_f(ewcljrsq);
 323             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 324             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 325             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 326             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 327             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 328             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 329             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 330             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 331
 332             /* Update potential sum for this i atom from the interaction with this j atom. */
 333             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 334             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 335
 336             fscal            = fvdw;
 337
 338             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 339
 340              /* Update vectorial force */
 341             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 342             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 343             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 344
 345             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 346             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 347             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 348             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 349             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 350                                                    _mm_mul_ps(dx00,fscal),
 351                                                    _mm_mul_ps(dy00,fscal),
 352                                                    _mm_mul_ps(dz00,fscal));
 353
 354             /* Inner loop uses 51 flops */
 355         }
 356
 357         /* End of innermost loop */
 358
 359         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 360                                               f+i_coord_offset,fshift+i_shift_offset);
 361
 362         ggid                        = gid[iidx];
 363         /* Update potential energies */
 364         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 365
 366         /* Increment number of inner iterations */
 367         inneriter                  += j_index_end - j_index_start;
 368
 369         /* Outer loop uses 7 flops */
 370     }
 371
 372     /* Increment number of outer iterations */
 373     outeriter        += nri;
 374
 375     /* Update outer/inner flops */
 376
 377     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*51);
 378 }
 379 /*
 380  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 381  * Electrostatics interaction: None
 382  * VdW interaction:            LJEwald
 383  * Geometry:                   Particle-Particle
 384  * Calculate force/pot:        Force
 385  */
 386 void
 387 nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 388                     (t_nblist                    * gmx_restrict       nlist,
 389                      rvec                        * gmx_restrict          xx,
 390                      rvec                        * gmx_restrict          ff,
 391                      struct t_forcerec           * gmx_restrict          fr,
 392                      t_mdatoms                   * gmx_restrict     mdatoms,
 393                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 394                      t_nrnb                      * gmx_restrict        nrnb)
 395 {
 396     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 397      * just 0 for non-waters.
 398      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 399      * jnr indices corresponding to data put in the four positions in the SIMD register.
 400      */
 401     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 402     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 403     int              jnrA,jnrB,jnrC,jnrD;
 404     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 405     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 406     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 407     real             rcutoff_scalar;
 408     real             *shiftvec,*fshift,*x,*f;
 409     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 410     real             scratch[4*DIM];
 411     __m128           fscal,rcutoff,rcutoff2,jidxall;
 412     int              vdwioffset0;
 413     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 414     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 415     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 416     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 417     int              nvdwtype;
 418     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 419     int              *vdwtype;
 420     real             *vdwparam;
 421     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 422     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 423     __m128           c6grid_00;
 424     real             *vdwgridparam;
 425     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 426     __m128           one_half = _mm_set1_ps(0.5);
 427     __m128           minus_one = _mm_set1_ps(-1.0);
 428     __m128           dummy_mask,cutoff_mask;
 429     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 430     __m128           one     = _mm_set1_ps(1.0);
 431     __m128           two     = _mm_set1_ps(2.0);
 432     x                = xx[0];
 433     f                = ff[0];
 434
 435     nri              = nlist->nri;
 436     iinr             = nlist->iinr;
 437     jindex           = nlist->jindex;
 438     jjnr             = nlist->jjnr;
 439     shiftidx         = nlist->shift;
 440     gid              = nlist->gid;
 441     shiftvec         = fr->shift_vec[0];
 442     fshift           = fr->fshift[0];
 443     nvdwtype         = fr->ntype;
 444     vdwparam         = fr->nbfp;
 445     vdwtype          = mdatoms->typeA;
 446     vdwgridparam     = fr->ljpme_c6grid;
 447     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 448     ewclj            = _mm_set1_ps(fr->ic->ewaldcoeff_lj);
 449     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 450
 451     /* Avoid stupid compiler warnings */
 452     jnrA = jnrB = jnrC = jnrD = 0;
 453     j_coord_offsetA = 0;
 454     j_coord_offsetB = 0;
 455     j_coord_offsetC = 0;
 456     j_coord_offsetD = 0;
 457
 458     outeriter        = 0;
 459     inneriter        = 0;
 460
 461     for(iidx=0;iidx<4*DIM;iidx++)
 462     {
 463         scratch[iidx] = 0.0;
 464     }
 465
 466     /* Start outer loop over neighborlists */
 467     for(iidx=0; iidx<nri; iidx++)
 468     {
 469         /* Load shift vector for this list */
 470         i_shift_offset   = DIM*shiftidx[iidx];
 471
 472         /* Load limits for loop over neighbors */
 473         j_index_start    = jindex[iidx];
 474         j_index_end      = jindex[iidx+1];
 475
 476         /* Get outer coordinate index */
 477         inr              = iinr[iidx];
 478         i_coord_offset   = DIM*inr;
 479
 480         /* Load i particle coords and add shift vector */
 481         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 482
 483         fix0             = _mm_setzero_ps();
 484         fiy0             = _mm_setzero_ps();
 485         fiz0             = _mm_setzero_ps();
 486
 487         /* Load parameters for i particles */
 488         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 489
 490         /* Start inner kernel loop */
 491         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 492         {
 493
 494             /* Get j neighbor index, and coordinate index */
 495             jnrA             = jjnr[jidx];
 496             jnrB             = jjnr[jidx+1];
 497             jnrC             = jjnr[jidx+2];
 498             jnrD             = jjnr[jidx+3];
 499             j_coord_offsetA  = DIM*jnrA;
 500             j_coord_offsetB  = DIM*jnrB;
 501             j_coord_offsetC  = DIM*jnrC;
 502             j_coord_offsetD  = DIM*jnrD;
 503
 504             /* load j atom coordinates */
 505             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 506                                               x+j_coord_offsetC,x+j_coord_offsetD,
 507                                               &jx0,&jy0,&jz0);
 508
 509             /* Calculate displacement vector */
 510             dx00             = _mm_sub_ps(ix0,jx0);
 511             dy00             = _mm_sub_ps(iy0,jy0);
 512             dz00             = _mm_sub_ps(iz0,jz0);
 513
 514             /* Calculate squared distance and things based on it */
 515             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 516
 517             rinv00           = avx128fma_invsqrt_f(rsq00);
 518
 519             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 520
 521             /* Load parameters for j particles */
 522             vdwjidx0A        = 2*vdwtype[jnrA+0];
 523             vdwjidx0B        = 2*vdwtype[jnrB+0];
 524             vdwjidx0C        = 2*vdwtype[jnrC+0];
 525             vdwjidx0D        = 2*vdwtype[jnrD+0];
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             r00              = _mm_mul_ps(rsq00,rinv00);
 532
 533             /* Compute parameters for interactions between i and j atoms */
 534             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 535                                          vdwparam+vdwioffset0+vdwjidx0B,
 536                                          vdwparam+vdwioffset0+vdwjidx0C,
 537                                          vdwparam+vdwioffset0+vdwjidx0D,
 538                                          &c6_00,&c12_00);
 539
 540             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 541                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 542                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 543                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 544
 545             /* Analytical LJ-PME */
 546             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 547             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 548             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 549             exponent         = avx128fma_exp_f(ewcljrsq);
 550             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 551             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 552             /* f6A = 6 * C6grid * (1 - poly) */
 553             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 554             /* f6B = C6grid * exponent * beta^6 */
 555             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 556             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 557             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 558
 559             fscal            = fvdw;
 560
 561              /* Update vectorial force */
 562             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 563             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 564             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 565
 566             fjptrA             = f+j_coord_offsetA;
 567             fjptrB             = f+j_coord_offsetB;
 568             fjptrC             = f+j_coord_offsetC;
 569             fjptrD             = f+j_coord_offsetD;
 570             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 571                                                    _mm_mul_ps(dx00,fscal),
 572                                                    _mm_mul_ps(dy00,fscal),
 573                                                    _mm_mul_ps(dz00,fscal));
 574
 575             /* Inner loop uses 47 flops */
 576         }
 577
 578         if(jidx<j_index_end)
 579         {
 580
 581             /* Get j neighbor index, and coordinate index */
 582             jnrlistA         = jjnr[jidx];
 583             jnrlistB         = jjnr[jidx+1];
 584             jnrlistC         = jjnr[jidx+2];
 585             jnrlistD         = jjnr[jidx+3];
 586             /* Sign of each element will be negative for non-real atoms.
 587              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 588              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 589              */
 590             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 591             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 592             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 593             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 594             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 595             j_coord_offsetA  = DIM*jnrA;
 596             j_coord_offsetB  = DIM*jnrB;
 597             j_coord_offsetC  = DIM*jnrC;
 598             j_coord_offsetD  = DIM*jnrD;
 599
 600             /* load j atom coordinates */
 601             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 602                                               x+j_coord_offsetC,x+j_coord_offsetD,
 603                                               &jx0,&jy0,&jz0);
 604
 605             /* Calculate displacement vector */
 606             dx00             = _mm_sub_ps(ix0,jx0);
 607             dy00             = _mm_sub_ps(iy0,jy0);
 608             dz00             = _mm_sub_ps(iz0,jz0);
 609
 610             /* Calculate squared distance and things based on it */
 611             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 612
 613             rinv00           = avx128fma_invsqrt_f(rsq00);
 614
 615             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 616
 617             /* Load parameters for j particles */
 618             vdwjidx0A        = 2*vdwtype[jnrA+0];
 619             vdwjidx0B        = 2*vdwtype[jnrB+0];
 620             vdwjidx0C        = 2*vdwtype[jnrC+0];
 621             vdwjidx0D        = 2*vdwtype[jnrD+0];
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             r00              = _mm_mul_ps(rsq00,rinv00);
 628             r00              = _mm_andnot_ps(dummy_mask,r00);
 629
 630             /* Compute parameters for interactions between i and j atoms */
 631             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 632                                          vdwparam+vdwioffset0+vdwjidx0B,
 633                                          vdwparam+vdwioffset0+vdwjidx0C,
 634                                          vdwparam+vdwioffset0+vdwjidx0D,
 635                                          &c6_00,&c12_00);
 636
 637             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 638                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 639                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 640                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 641
 642             /* Analytical LJ-PME */
 643             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 644             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 645             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 646             exponent         = avx128fma_exp_f(ewcljrsq);
 647             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 648             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 649             /* f6A = 6 * C6grid * (1 - poly) */
 650             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 651             /* f6B = C6grid * exponent * beta^6 */
 652             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 653             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 654             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 655
 656             fscal            = fvdw;
 657
 658             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 659
 660              /* Update vectorial force */
 661             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 662             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 663             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 664
 665             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 666             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 667             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 668             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 669             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 670                                                    _mm_mul_ps(dx00,fscal),
 671                                                    _mm_mul_ps(dy00,fscal),
 672                                                    _mm_mul_ps(dz00,fscal));
 673
 674             /* Inner loop uses 48 flops */
 675         }
 676
 677         /* End of innermost loop */
 678
 679         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 680                                               f+i_coord_offset,fshift+i_shift_offset);
 681
 682         /* Increment number of inner iterations */
 683         inneriter                  += j_index_end - j_index_start;
 684
 685         /* Outer loop uses 6 flops */
 686     }
 687
 688     /* Increment number of outer iterations */
 689     outeriter        += nri;
 690
 691     /* Update outer/inner flops */
 692
 693     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*48);
 694 }