src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwLJEw_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, 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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: None
  53  * VdW interaction:            LJEwald
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwLJEw_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     int              nvdwtype;
  89     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  90     int              *vdwtype;
  91     real             *vdwparam;
  92     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  93     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  94     __m128           c6grid_00;
  95     real             *vdwgridparam;
  96     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  97     __m128           one_half = _mm_set1_ps(0.5);
  98     __m128           minus_one = _mm_set1_ps(-1.0);
  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     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117     vdwgridparam     = fr->ljpme_c6grid;
 118     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 119     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 120     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm_setzero_ps();
 155         fiy0             = _mm_setzero_ps();
 156         fiz0             = _mm_setzero_ps();
 157
 158         /* Load parameters for i particles */
 159         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         vvdwsum          = _mm_setzero_ps();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             jnrC             = jjnr[jidx+2];
 172             jnrD             = jjnr[jidx+3];
 173             j_coord_offsetA  = DIM*jnrA;
 174             j_coord_offsetB  = DIM*jnrB;
 175             j_coord_offsetC  = DIM*jnrC;
 176             j_coord_offsetD  = DIM*jnrD;
 177
 178             /* load j atom coordinates */
 179             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                               x+j_coord_offsetC,x+j_coord_offsetD,
 181                                               &jx0,&jy0,&jz0);
 182
 183             /* Calculate displacement vector */
 184             dx00             = _mm_sub_ps(ix0,jx0);
 185             dy00             = _mm_sub_ps(iy0,jy0);
 186             dz00             = _mm_sub_ps(iz0,jz0);
 187
 188             /* Calculate squared distance and things based on it */
 189             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 190
 191             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 192
 193             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 194
 195             /* Load parameters for j particles */
 196             vdwjidx0A        = 2*vdwtype[jnrA+0];
 197             vdwjidx0B        = 2*vdwtype[jnrB+0];
 198             vdwjidx0C        = 2*vdwtype[jnrC+0];
 199             vdwjidx0D        = 2*vdwtype[jnrD+0];
 200
 201             /**************************
 202              * CALCULATE INTERACTIONS *
 203              **************************/
 204
 205             r00              = _mm_mul_ps(rsq00,rinv00);
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 209                                          vdwparam+vdwioffset0+vdwjidx0B,
 210                                          vdwparam+vdwioffset0+vdwjidx0C,
 211                                          vdwparam+vdwioffset0+vdwjidx0D,
 212                                          &c6_00,&c12_00);
 213
 214             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 215                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 216                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 217                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 218
 219             /* Analytical LJ-PME */
 220             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 221             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 222             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 223             exponent         = gmx_simd_exp_r(ewcljrsq);
 224             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 225             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 226             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 227             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 228             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 229             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 230             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 231             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);
 232
 233             /* Update potential sum for this i atom from the interaction with this j atom. */
 234             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 235
 236             fscal            = fvdw;
 237
 238              /* Update vectorial force */
 239             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 240             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 241             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 242
 243             fjptrA             = f+j_coord_offsetA;
 244             fjptrB             = f+j_coord_offsetB;
 245             fjptrC             = f+j_coord_offsetC;
 246             fjptrD             = f+j_coord_offsetD;
 247             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 248                                                    _mm_mul_ps(dx00,fscal),
 249                                                    _mm_mul_ps(dy00,fscal),
 250                                                    _mm_mul_ps(dz00,fscal));
 251
 252             /* Inner loop uses 50 flops */
 253         }
 254
 255         if(jidx<j_index_end)
 256         {
 257
 258             /* Get j neighbor index, and coordinate index */
 259             jnrlistA         = jjnr[jidx];
 260             jnrlistB         = jjnr[jidx+1];
 261             jnrlistC         = jjnr[jidx+2];
 262             jnrlistD         = jjnr[jidx+3];
 263             /* Sign of each element will be negative for non-real atoms.
 264              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 265              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 266              */
 267             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 268             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 269             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 270             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 271             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 272             j_coord_offsetA  = DIM*jnrA;
 273             j_coord_offsetB  = DIM*jnrB;
 274             j_coord_offsetC  = DIM*jnrC;
 275             j_coord_offsetD  = DIM*jnrD;
 276
 277             /* load j atom coordinates */
 278             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 279                                               x+j_coord_offsetC,x+j_coord_offsetD,
 280                                               &jx0,&jy0,&jz0);
 281
 282             /* Calculate displacement vector */
 283             dx00             = _mm_sub_ps(ix0,jx0);
 284             dy00             = _mm_sub_ps(iy0,jy0);
 285             dz00             = _mm_sub_ps(iz0,jz0);
 286
 287             /* Calculate squared distance and things based on it */
 288             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 289
 290             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 291
 292             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 293
 294             /* Load parameters for j particles */
 295             vdwjidx0A        = 2*vdwtype[jnrA+0];
 296             vdwjidx0B        = 2*vdwtype[jnrB+0];
 297             vdwjidx0C        = 2*vdwtype[jnrC+0];
 298             vdwjidx0D        = 2*vdwtype[jnrD+0];
 299
 300             /**************************
 301              * CALCULATE INTERACTIONS *
 302              **************************/
 303
 304             r00              = _mm_mul_ps(rsq00,rinv00);
 305             r00              = _mm_andnot_ps(dummy_mask,r00);
 306
 307             /* Compute parameters for interactions between i and j atoms */
 308             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 309                                          vdwparam+vdwioffset0+vdwjidx0B,
 310                                          vdwparam+vdwioffset0+vdwjidx0C,
 311                                          vdwparam+vdwioffset0+vdwjidx0D,
 312                                          &c6_00,&c12_00);
 313
 314             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 315                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 316                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 317                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 318
 319             /* Analytical LJ-PME */
 320             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 321             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 322             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 323             exponent         = gmx_simd_exp_r(ewcljrsq);
 324             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 325             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 326             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 327             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 328             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 329             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 330             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 331             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);
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 335             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 336
 337             fscal            = fvdw;
 338
 339             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 340
 341              /* Update vectorial force */
 342             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 343             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 344             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 345
 346             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 347             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 348             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 349             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 350             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 351                                                    _mm_mul_ps(dx00,fscal),
 352                                                    _mm_mul_ps(dy00,fscal),
 353                                                    _mm_mul_ps(dz00,fscal));
 354
 355             /* Inner loop uses 51 flops */
 356         }
 357
 358         /* End of innermost loop */
 359
 360         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 361                                               f+i_coord_offset,fshift+i_shift_offset);
 362
 363         ggid                        = gid[iidx];
 364         /* Update potential energies */
 365         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 366
 367         /* Increment number of inner iterations */
 368         inneriter                  += j_index_end - j_index_start;
 369
 370         /* Outer loop uses 7 flops */
 371     }
 372
 373     /* Increment number of outer iterations */
 374     outeriter        += nri;
 375
 376     /* Update outer/inner flops */
 377
 378     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*51);
 379 }
 380 /*
 381  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 382  * Electrostatics interaction: None
 383  * VdW interaction:            LJEwald
 384  * Geometry:                   Particle-Particle
 385  * Calculate force/pot:        Force
 386  */
 387 void
 388 nb_kernel_ElecNone_VdwLJEw_GeomP1P1_F_avx_128_fma_single
 389                     (t_nblist                    * gmx_restrict       nlist,
 390                      rvec                        * gmx_restrict          xx,
 391                      rvec                        * gmx_restrict          ff,
 392                      t_forcerec                  * gmx_restrict          fr,
 393                      t_mdatoms                   * gmx_restrict     mdatoms,
 394                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 395                      t_nrnb                      * gmx_restrict        nrnb)
 396 {
 397     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 398      * just 0 for non-waters.
 399      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 400      * jnr indices corresponding to data put in the four positions in the SIMD register.
 401      */
 402     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 403     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 404     int              jnrA,jnrB,jnrC,jnrD;
 405     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 406     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 407     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 408     real             rcutoff_scalar;
 409     real             *shiftvec,*fshift,*x,*f;
 410     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 411     real             scratch[4*DIM];
 412     __m128           fscal,rcutoff,rcutoff2,jidxall;
 413     int              vdwioffset0;
 414     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 415     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 416     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 417     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 418     int              nvdwtype;
 419     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 420     int              *vdwtype;
 421     real             *vdwparam;
 422     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 423     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 424     __m128           c6grid_00;
 425     real             *vdwgridparam;
 426     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 427     __m128           one_half = _mm_set1_ps(0.5);
 428     __m128           minus_one = _mm_set1_ps(-1.0);
 429     __m128           dummy_mask,cutoff_mask;
 430     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 431     __m128           one     = _mm_set1_ps(1.0);
 432     __m128           two     = _mm_set1_ps(2.0);
 433     x                = xx[0];
 434     f                = ff[0];
 435
 436     nri              = nlist->nri;
 437     iinr             = nlist->iinr;
 438     jindex           = nlist->jindex;
 439     jjnr             = nlist->jjnr;
 440     shiftidx         = nlist->shift;
 441     gid              = nlist->gid;
 442     shiftvec         = fr->shift_vec[0];
 443     fshift           = fr->fshift[0];
 444     nvdwtype         = fr->ntype;
 445     vdwparam         = fr->nbfp;
 446     vdwtype          = mdatoms->typeA;
 447     vdwgridparam     = fr->ljpme_c6grid;
 448     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 449     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 450     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 451
 452     /* Avoid stupid compiler warnings */
 453     jnrA = jnrB = jnrC = jnrD = 0;
 454     j_coord_offsetA = 0;
 455     j_coord_offsetB = 0;
 456     j_coord_offsetC = 0;
 457     j_coord_offsetD = 0;
 458
 459     outeriter        = 0;
 460     inneriter        = 0;
 461
 462     for(iidx=0;iidx<4*DIM;iidx++)
 463     {
 464         scratch[iidx] = 0.0;
 465     }
 466
 467     /* Start outer loop over neighborlists */
 468     for(iidx=0; iidx<nri; iidx++)
 469     {
 470         /* Load shift vector for this list */
 471         i_shift_offset   = DIM*shiftidx[iidx];
 472
 473         /* Load limits for loop over neighbors */
 474         j_index_start    = jindex[iidx];
 475         j_index_end      = jindex[iidx+1];
 476
 477         /* Get outer coordinate index */
 478         inr              = iinr[iidx];
 479         i_coord_offset   = DIM*inr;
 480
 481         /* Load i particle coords and add shift vector */
 482         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 483
 484         fix0             = _mm_setzero_ps();
 485         fiy0             = _mm_setzero_ps();
 486         fiz0             = _mm_setzero_ps();
 487
 488         /* Load parameters for i particles */
 489         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 490
 491         /* Start inner kernel loop */
 492         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 493         {
 494
 495             /* Get j neighbor index, and coordinate index */
 496             jnrA             = jjnr[jidx];
 497             jnrB             = jjnr[jidx+1];
 498             jnrC             = jjnr[jidx+2];
 499             jnrD             = jjnr[jidx+3];
 500             j_coord_offsetA  = DIM*jnrA;
 501             j_coord_offsetB  = DIM*jnrB;
 502             j_coord_offsetC  = DIM*jnrC;
 503             j_coord_offsetD  = DIM*jnrD;
 504
 505             /* load j atom coordinates */
 506             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 507                                               x+j_coord_offsetC,x+j_coord_offsetD,
 508                                               &jx0,&jy0,&jz0);
 509
 510             /* Calculate displacement vector */
 511             dx00             = _mm_sub_ps(ix0,jx0);
 512             dy00             = _mm_sub_ps(iy0,jy0);
 513             dz00             = _mm_sub_ps(iz0,jz0);
 514
 515             /* Calculate squared distance and things based on it */
 516             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 517
 518             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 519
 520             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 521
 522             /* Load parameters for j particles */
 523             vdwjidx0A        = 2*vdwtype[jnrA+0];
 524             vdwjidx0B        = 2*vdwtype[jnrB+0];
 525             vdwjidx0C        = 2*vdwtype[jnrC+0];
 526             vdwjidx0D        = 2*vdwtype[jnrD+0];
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             r00              = _mm_mul_ps(rsq00,rinv00);
 533
 534             /* Compute parameters for interactions between i and j atoms */
 535             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 536                                          vdwparam+vdwioffset0+vdwjidx0B,
 537                                          vdwparam+vdwioffset0+vdwjidx0C,
 538                                          vdwparam+vdwioffset0+vdwjidx0D,
 539                                          &c6_00,&c12_00);
 540
 541             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 542                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 543                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 544                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 545
 546             /* Analytical LJ-PME */
 547             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 548             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 549             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 550             exponent         = gmx_simd_exp_r(ewcljrsq);
 551             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 552             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 553             /* f6A = 6 * C6grid * (1 - poly) */
 554             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 555             /* f6B = C6grid * exponent * beta^6 */
 556             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 557             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 558             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 559
 560             fscal            = fvdw;
 561
 562              /* Update vectorial force */
 563             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 564             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 565             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 566
 567             fjptrA             = f+j_coord_offsetA;
 568             fjptrB             = f+j_coord_offsetB;
 569             fjptrC             = f+j_coord_offsetC;
 570             fjptrD             = f+j_coord_offsetD;
 571             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 572                                                    _mm_mul_ps(dx00,fscal),
 573                                                    _mm_mul_ps(dy00,fscal),
 574                                                    _mm_mul_ps(dz00,fscal));
 575
 576             /* Inner loop uses 47 flops */
 577         }
 578
 579         if(jidx<j_index_end)
 580         {
 581
 582             /* Get j neighbor index, and coordinate index */
 583             jnrlistA         = jjnr[jidx];
 584             jnrlistB         = jjnr[jidx+1];
 585             jnrlistC         = jjnr[jidx+2];
 586             jnrlistD         = jjnr[jidx+3];
 587             /* Sign of each element will be negative for non-real atoms.
 588              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 589              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 590              */
 591             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 592             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 593             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 594             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 595             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 596             j_coord_offsetA  = DIM*jnrA;
 597             j_coord_offsetB  = DIM*jnrB;
 598             j_coord_offsetC  = DIM*jnrC;
 599             j_coord_offsetD  = DIM*jnrD;
 600
 601             /* load j atom coordinates */
 602             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 603                                               x+j_coord_offsetC,x+j_coord_offsetD,
 604                                               &jx0,&jy0,&jz0);
 605
 606             /* Calculate displacement vector */
 607             dx00             = _mm_sub_ps(ix0,jx0);
 608             dy00             = _mm_sub_ps(iy0,jy0);
 609             dz00             = _mm_sub_ps(iz0,jz0);
 610
 611             /* Calculate squared distance and things based on it */
 612             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 613
 614             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 615
 616             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 617
 618             /* Load parameters for j particles */
 619             vdwjidx0A        = 2*vdwtype[jnrA+0];
 620             vdwjidx0B        = 2*vdwtype[jnrB+0];
 621             vdwjidx0C        = 2*vdwtype[jnrC+0];
 622             vdwjidx0D        = 2*vdwtype[jnrD+0];
 623
 624             /**************************
 625              * CALCULATE INTERACTIONS *
 626              **************************/
 627
 628             r00              = _mm_mul_ps(rsq00,rinv00);
 629             r00              = _mm_andnot_ps(dummy_mask,r00);
 630
 631             /* Compute parameters for interactions between i and j atoms */
 632             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 633                                          vdwparam+vdwioffset0+vdwjidx0B,
 634                                          vdwparam+vdwioffset0+vdwjidx0C,
 635                                          vdwparam+vdwioffset0+vdwjidx0D,
 636                                          &c6_00,&c12_00);
 637
 638             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 639                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 640                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 641                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 642
 643             /* Analytical LJ-PME */
 644             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 645             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 646             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 647             exponent         = gmx_simd_exp_r(ewcljrsq);
 648             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 649             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 650             /* f6A = 6 * C6grid * (1 - poly) */
 651             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 652             /* f6B = C6grid * exponent * beta^6 */
 653             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 654             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 655             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 656
 657             fscal            = fvdw;
 658
 659             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 660
 661              /* Update vectorial force */
 662             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 663             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 664             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 665
 666             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 667             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 668             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 669             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 670             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 671                                                    _mm_mul_ps(dx00,fscal),
 672                                                    _mm_mul_ps(dy00,fscal),
 673                                                    _mm_mul_ps(dz00,fscal));
 674
 675             /* Inner loop uses 48 flops */
 676         }
 677
 678         /* End of innermost loop */
 679
 680         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 681                                               f+i_coord_offset,fshift+i_shift_offset);
 682
 683         /* Increment number of inner iterations */
 684         inneriter                  += j_index_end - j_index_start;
 685
 686         /* Outer loop uses 6 flops */
 687     }
 688
 689     /* Increment number of outer iterations */
 690     outeriter        += nri;
 691
 692     /* Update outer/inner flops */
 693
 694     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*48);
 695 }