src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwLJSh_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: None
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     int              nvdwtype;
  91     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  95     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  96     __m128           dummy_mask,cutoff_mask;
  97     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  98     __m128           one     = _mm_set1_ps(1.0);
  99     __m128           two     = _mm_set1_ps(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     nvdwtype         = fr->ntype;
 112     vdwparam         = fr->nbfp;
 113     vdwtype          = mdatoms->typeA;
 114
 115     rcutoff_scalar   = fr->rvdw;
 116     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 117     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 118
 119     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 120     rvdw             = _mm_set1_ps(fr->rvdw);
 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             rinvsq00         = gmx_mm_inv_ps(rsq00);
 192
 193             /* Load parameters for j particles */
 194             vdwjidx0A        = 2*vdwtype[jnrA+0];
 195             vdwjidx0B        = 2*vdwtype[jnrB+0];
 196             vdwjidx0C        = 2*vdwtype[jnrC+0];
 197             vdwjidx0D        = 2*vdwtype[jnrD+0];
 198
 199             /**************************
 200              * CALCULATE INTERACTIONS *
 201              **************************/
 202
 203             if (gmx_mm_any_lt(rsq00,rcutoff2))
 204             {
 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             /* LENNARD-JONES DISPERSION/REPULSION */
 214
 215             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 216             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 217             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 218             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 219                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 220             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 221
 222             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 223
 224             /* Update potential sum for this i atom from the interaction with this j atom. */
 225             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 226             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 227
 228             fscal            = fvdw;
 229
 230             fscal            = _mm_and_ps(fscal,cutoff_mask);
 231
 232              /* Update vectorial force */
 233             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 234             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 235             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 236
 237             fjptrA             = f+j_coord_offsetA;
 238             fjptrB             = f+j_coord_offsetB;
 239             fjptrC             = f+j_coord_offsetC;
 240             fjptrD             = f+j_coord_offsetD;
 241             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 242                                                    _mm_mul_ps(dx00,fscal),
 243                                                    _mm_mul_ps(dy00,fscal),
 244                                                    _mm_mul_ps(dz00,fscal));
 245
 246             }
 247
 248             /* Inner loop uses 44 flops */
 249         }
 250
 251         if(jidx<j_index_end)
 252         {
 253
 254             /* Get j neighbor index, and coordinate index */
 255             jnrlistA         = jjnr[jidx];
 256             jnrlistB         = jjnr[jidx+1];
 257             jnrlistC         = jjnr[jidx+2];
 258             jnrlistD         = jjnr[jidx+3];
 259             /* Sign of each element will be negative for non-real atoms.
 260              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 261              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 262              */
 263             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 264             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 265             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 266             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 267             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 268             j_coord_offsetA  = DIM*jnrA;
 269             j_coord_offsetB  = DIM*jnrB;
 270             j_coord_offsetC  = DIM*jnrC;
 271             j_coord_offsetD  = DIM*jnrD;
 272
 273             /* load j atom coordinates */
 274             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 275                                               x+j_coord_offsetC,x+j_coord_offsetD,
 276                                               &jx0,&jy0,&jz0);
 277
 278             /* Calculate displacement vector */
 279             dx00             = _mm_sub_ps(ix0,jx0);
 280             dy00             = _mm_sub_ps(iy0,jy0);
 281             dz00             = _mm_sub_ps(iz0,jz0);
 282
 283             /* Calculate squared distance and things based on it */
 284             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 285
 286             rinvsq00         = gmx_mm_inv_ps(rsq00);
 287
 288             /* Load parameters for j particles */
 289             vdwjidx0A        = 2*vdwtype[jnrA+0];
 290             vdwjidx0B        = 2*vdwtype[jnrB+0];
 291             vdwjidx0C        = 2*vdwtype[jnrC+0];
 292             vdwjidx0D        = 2*vdwtype[jnrD+0];
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             if (gmx_mm_any_lt(rsq00,rcutoff2))
 299             {
 300
 301             /* Compute parameters for interactions between i and j atoms */
 302             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 303                                          vdwparam+vdwioffset0+vdwjidx0B,
 304                                          vdwparam+vdwioffset0+vdwjidx0C,
 305                                          vdwparam+vdwioffset0+vdwjidx0D,
 306                                          &c6_00,&c12_00);
 307
 308             /* LENNARD-JONES DISPERSION/REPULSION */
 309
 310             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 311             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 312             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 313             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 314                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 315             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 316
 317             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 318
 319             /* Update potential sum for this i atom from the interaction with this j atom. */
 320             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 321             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 322             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 323
 324             fscal            = fvdw;
 325
 326             fscal            = _mm_and_ps(fscal,cutoff_mask);
 327
 328             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 329
 330              /* Update vectorial force */
 331             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 332             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 333             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 334
 335             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 336             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 337             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 338             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 339             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 340                                                    _mm_mul_ps(dx00,fscal),
 341                                                    _mm_mul_ps(dy00,fscal),
 342                                                    _mm_mul_ps(dz00,fscal));
 343
 344             }
 345
 346             /* Inner loop uses 44 flops */
 347         }
 348
 349         /* End of innermost loop */
 350
 351         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 352                                               f+i_coord_offset,fshift+i_shift_offset);
 353
 354         ggid                        = gid[iidx];
 355         /* Update potential energies */
 356         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 357
 358         /* Increment number of inner iterations */
 359         inneriter                  += j_index_end - j_index_start;
 360
 361         /* Outer loop uses 7 flops */
 362     }
 363
 364     /* Increment number of outer iterations */
 365     outeriter        += nri;
 366
 367     /* Update outer/inner flops */
 368
 369     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*44);
 370 }
 371 /*
 372  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 373  * Electrostatics interaction: None
 374  * VdW interaction:            LennardJones
 375  * Geometry:                   Particle-Particle
 376  * Calculate force/pot:        Force
 377  */
 378 void
 379 nb_kernel_ElecNone_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 380                     (t_nblist                    * gmx_restrict       nlist,
 381                      rvec                        * gmx_restrict          xx,
 382                      rvec                        * gmx_restrict          ff,
 383                      t_forcerec                  * gmx_restrict          fr,
 384                      t_mdatoms                   * gmx_restrict     mdatoms,
 385                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 386                      t_nrnb                      * gmx_restrict        nrnb)
 387 {
 388     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 389      * just 0 for non-waters.
 390      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 391      * jnr indices corresponding to data put in the four positions in the SIMD register.
 392      */
 393     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 394     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 395     int              jnrA,jnrB,jnrC,jnrD;
 396     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 397     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 398     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 399     real             rcutoff_scalar;
 400     real             *shiftvec,*fshift,*x,*f;
 401     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 402     real             scratch[4*DIM];
 403     __m128           fscal,rcutoff,rcutoff2,jidxall;
 404     int              vdwioffset0;
 405     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 406     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 407     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 408     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 409     int              nvdwtype;
 410     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 411     int              *vdwtype;
 412     real             *vdwparam;
 413     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 414     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 415     __m128           dummy_mask,cutoff_mask;
 416     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 417     __m128           one     = _mm_set1_ps(1.0);
 418     __m128           two     = _mm_set1_ps(2.0);
 419     x                = xx[0];
 420     f                = ff[0];
 421
 422     nri              = nlist->nri;
 423     iinr             = nlist->iinr;
 424     jindex           = nlist->jindex;
 425     jjnr             = nlist->jjnr;
 426     shiftidx         = nlist->shift;
 427     gid              = nlist->gid;
 428     shiftvec         = fr->shift_vec[0];
 429     fshift           = fr->fshift[0];
 430     nvdwtype         = fr->ntype;
 431     vdwparam         = fr->nbfp;
 432     vdwtype          = mdatoms->typeA;
 433
 434     rcutoff_scalar   = fr->rvdw;
 435     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 436     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 437
 438     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 439     rvdw             = _mm_set1_ps(fr->rvdw);
 440
 441     /* Avoid stupid compiler warnings */
 442     jnrA = jnrB = jnrC = jnrD = 0;
 443     j_coord_offsetA = 0;
 444     j_coord_offsetB = 0;
 445     j_coord_offsetC = 0;
 446     j_coord_offsetD = 0;
 447
 448     outeriter        = 0;
 449     inneriter        = 0;
 450
 451     for(iidx=0;iidx<4*DIM;iidx++)
 452     {
 453         scratch[iidx] = 0.0;
 454     }
 455
 456     /* Start outer loop over neighborlists */
 457     for(iidx=0; iidx<nri; iidx++)
 458     {
 459         /* Load shift vector for this list */
 460         i_shift_offset   = DIM*shiftidx[iidx];
 461
 462         /* Load limits for loop over neighbors */
 463         j_index_start    = jindex[iidx];
 464         j_index_end      = jindex[iidx+1];
 465
 466         /* Get outer coordinate index */
 467         inr              = iinr[iidx];
 468         i_coord_offset   = DIM*inr;
 469
 470         /* Load i particle coords and add shift vector */
 471         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 472
 473         fix0             = _mm_setzero_ps();
 474         fiy0             = _mm_setzero_ps();
 475         fiz0             = _mm_setzero_ps();
 476
 477         /* Load parameters for i particles */
 478         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 479
 480         /* Start inner kernel loop */
 481         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 482         {
 483
 484             /* Get j neighbor index, and coordinate index */
 485             jnrA             = jjnr[jidx];
 486             jnrB             = jjnr[jidx+1];
 487             jnrC             = jjnr[jidx+2];
 488             jnrD             = jjnr[jidx+3];
 489             j_coord_offsetA  = DIM*jnrA;
 490             j_coord_offsetB  = DIM*jnrB;
 491             j_coord_offsetC  = DIM*jnrC;
 492             j_coord_offsetD  = DIM*jnrD;
 493
 494             /* load j atom coordinates */
 495             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 496                                               x+j_coord_offsetC,x+j_coord_offsetD,
 497                                               &jx0,&jy0,&jz0);
 498
 499             /* Calculate displacement vector */
 500             dx00             = _mm_sub_ps(ix0,jx0);
 501             dy00             = _mm_sub_ps(iy0,jy0);
 502             dz00             = _mm_sub_ps(iz0,jz0);
 503
 504             /* Calculate squared distance and things based on it */
 505             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 506
 507             rinvsq00         = gmx_mm_inv_ps(rsq00);
 508
 509             /* Load parameters for j particles */
 510             vdwjidx0A        = 2*vdwtype[jnrA+0];
 511             vdwjidx0B        = 2*vdwtype[jnrB+0];
 512             vdwjidx0C        = 2*vdwtype[jnrC+0];
 513             vdwjidx0D        = 2*vdwtype[jnrD+0];
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             if (gmx_mm_any_lt(rsq00,rcutoff2))
 520             {
 521
 522             /* Compute parameters for interactions between i and j atoms */
 523             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 524                                          vdwparam+vdwioffset0+vdwjidx0B,
 525                                          vdwparam+vdwioffset0+vdwjidx0C,
 526                                          vdwparam+vdwioffset0+vdwjidx0D,
 527                                          &c6_00,&c12_00);
 528
 529             /* LENNARD-JONES DISPERSION/REPULSION */
 530
 531             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 532             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 533
 534             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 535
 536             fscal            = fvdw;
 537
 538             fscal            = _mm_and_ps(fscal,cutoff_mask);
 539
 540              /* Update vectorial force */
 541             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 542             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 543             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 544
 545             fjptrA             = f+j_coord_offsetA;
 546             fjptrB             = f+j_coord_offsetB;
 547             fjptrC             = f+j_coord_offsetC;
 548             fjptrD             = f+j_coord_offsetD;
 549             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 550                                                    _mm_mul_ps(dx00,fscal),
 551                                                    _mm_mul_ps(dy00,fscal),
 552                                                    _mm_mul_ps(dz00,fscal));
 553
 554             }
 555
 556             /* Inner loop uses 33 flops */
 557         }
 558
 559         if(jidx<j_index_end)
 560         {
 561
 562             /* Get j neighbor index, and coordinate index */
 563             jnrlistA         = jjnr[jidx];
 564             jnrlistB         = jjnr[jidx+1];
 565             jnrlistC         = jjnr[jidx+2];
 566             jnrlistD         = jjnr[jidx+3];
 567             /* Sign of each element will be negative for non-real atoms.
 568              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 569              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 570              */
 571             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 572             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 573             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 574             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 575             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 576             j_coord_offsetA  = DIM*jnrA;
 577             j_coord_offsetB  = DIM*jnrB;
 578             j_coord_offsetC  = DIM*jnrC;
 579             j_coord_offsetD  = DIM*jnrD;
 580
 581             /* load j atom coordinates */
 582             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 583                                               x+j_coord_offsetC,x+j_coord_offsetD,
 584                                               &jx0,&jy0,&jz0);
 585
 586             /* Calculate displacement vector */
 587             dx00             = _mm_sub_ps(ix0,jx0);
 588             dy00             = _mm_sub_ps(iy0,jy0);
 589             dz00             = _mm_sub_ps(iz0,jz0);
 590
 591             /* Calculate squared distance and things based on it */
 592             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 593
 594             rinvsq00         = gmx_mm_inv_ps(rsq00);
 595
 596             /* Load parameters for j particles */
 597             vdwjidx0A        = 2*vdwtype[jnrA+0];
 598             vdwjidx0B        = 2*vdwtype[jnrB+0];
 599             vdwjidx0C        = 2*vdwtype[jnrC+0];
 600             vdwjidx0D        = 2*vdwtype[jnrD+0];
 601
 602             /**************************
 603              * CALCULATE INTERACTIONS *
 604              **************************/
 605
 606             if (gmx_mm_any_lt(rsq00,rcutoff2))
 607             {
 608
 609             /* Compute parameters for interactions between i and j atoms */
 610             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 611                                          vdwparam+vdwioffset0+vdwjidx0B,
 612                                          vdwparam+vdwioffset0+vdwjidx0C,
 613                                          vdwparam+vdwioffset0+vdwjidx0D,
 614                                          &c6_00,&c12_00);
 615
 616             /* LENNARD-JONES DISPERSION/REPULSION */
 617
 618             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 619             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 620
 621             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 622
 623             fscal            = fvdw;
 624
 625             fscal            = _mm_and_ps(fscal,cutoff_mask);
 626
 627             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 628
 629              /* Update vectorial force */
 630             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 631             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 632             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 633
 634             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 635             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 636             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 637             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 638             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 639                                                    _mm_mul_ps(dx00,fscal),
 640                                                    _mm_mul_ps(dy00,fscal),
 641                                                    _mm_mul_ps(dz00,fscal));
 642
 643             }
 644
 645             /* Inner loop uses 33 flops */
 646         }
 647
 648         /* End of innermost loop */
 649
 650         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 651                                               f+i_coord_offset,fshift+i_shift_offset);
 652
 653         /* Increment number of inner iterations */
 654         inneriter                  += j_index_end - j_index_start;
 655
 656         /* Outer loop uses 6 flops */
 657     }
 658
 659     /* Increment number of outer iterations */
 660     outeriter        += nri;
 661
 662     /* Update outer/inner flops */
 663
 664     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*33);
 665 }