src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwLJSw_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_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: None
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwLJSw_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           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  95     real             rswitch_scalar,d_scalar;
  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     rswitch_scalar   = fr->rvdw_switch;
 120     rswitch          = _mm_set1_ps(rswitch_scalar);
 121     /* Setup switch parameters */
 122     d_scalar         = rcutoff_scalar-rswitch_scalar;
 123     d                = _mm_set1_ps(d_scalar);
 124     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 125     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 126     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 127     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 128     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 129     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 130
 131     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = jnrC = jnrD = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135     j_coord_offsetC = 0;
 136     j_coord_offsetD = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm_setzero_ps();
 164         fiy0             = _mm_setzero_ps();
 165         fiz0             = _mm_setzero_ps();
 166
 167         /* Load parameters for i particles */
 168         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 169
 170         /* Reset potential sums */
 171         vvdwsum          = _mm_setzero_ps();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             jnrC             = jjnr[jidx+2];
 181             jnrD             = jjnr[jidx+3];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184             j_coord_offsetC  = DIM*jnrC;
 185             j_coord_offsetD  = DIM*jnrD;
 186
 187             /* load j atom coordinates */
 188             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                               x+j_coord_offsetC,x+j_coord_offsetD,
 190                                               &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_ps(ix0,jx0);
 194             dy00             = _mm_sub_ps(iy0,jy0);
 195             dz00             = _mm_sub_ps(iz0,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 199
 200             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 201
 202             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 203
 204             /* Load parameters for j particles */
 205             vdwjidx0A        = 2*vdwtype[jnrA+0];
 206             vdwjidx0B        = 2*vdwtype[jnrB+0];
 207             vdwjidx0C        = 2*vdwtype[jnrC+0];
 208             vdwjidx0D        = 2*vdwtype[jnrD+0];
 209
 210             /**************************
 211              * CALCULATE INTERACTIONS *
 212              **************************/
 213
 214             if (gmx_mm_any_lt(rsq00,rcutoff2))
 215             {
 216
 217             r00              = _mm_mul_ps(rsq00,rinv00);
 218
 219             /* Compute parameters for interactions between i and j atoms */
 220             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 221                                          vdwparam+vdwioffset0+vdwjidx0B,
 222                                          vdwparam+vdwioffset0+vdwjidx0C,
 223                                          vdwparam+vdwioffset0+vdwjidx0D,
 224                                          &c6_00,&c12_00);
 225
 226             /* LENNARD-JONES DISPERSION/REPULSION */
 227
 228             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 229             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 230             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 231             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 232             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 233
 234             d                = _mm_sub_ps(r00,rswitch);
 235             d                = _mm_max_ps(d,_mm_setzero_ps());
 236             d2               = _mm_mul_ps(d,d);
 237             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 238
 239             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 240
 241             /* Evaluate switch function */
 242             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 243             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 244             vvdw             = _mm_mul_ps(vvdw,sw);
 245             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 249             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 250
 251             fscal            = fvdw;
 252
 253             fscal            = _mm_and_ps(fscal,cutoff_mask);
 254
 255              /* Update vectorial force */
 256             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 257             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 258             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 259
 260             fjptrA             = f+j_coord_offsetA;
 261             fjptrB             = f+j_coord_offsetB;
 262             fjptrC             = f+j_coord_offsetC;
 263             fjptrD             = f+j_coord_offsetD;
 264             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 265                                                    _mm_mul_ps(dx00,fscal),
 266                                                    _mm_mul_ps(dy00,fscal),
 267                                                    _mm_mul_ps(dz00,fscal));
 268
 269             }
 270
 271             /* Inner loop uses 62 flops */
 272         }
 273
 274         if(jidx<j_index_end)
 275         {
 276
 277             /* Get j neighbor index, and coordinate index */
 278             jnrlistA         = jjnr[jidx];
 279             jnrlistB         = jjnr[jidx+1];
 280             jnrlistC         = jjnr[jidx+2];
 281             jnrlistD         = jjnr[jidx+3];
 282             /* Sign of each element will be negative for non-real atoms.
 283              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 284              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 285              */
 286             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 287             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 288             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 289             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 290             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 291             j_coord_offsetA  = DIM*jnrA;
 292             j_coord_offsetB  = DIM*jnrB;
 293             j_coord_offsetC  = DIM*jnrC;
 294             j_coord_offsetD  = DIM*jnrD;
 295
 296             /* load j atom coordinates */
 297             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 298                                               x+j_coord_offsetC,x+j_coord_offsetD,
 299                                               &jx0,&jy0,&jz0);
 300
 301             /* Calculate displacement vector */
 302             dx00             = _mm_sub_ps(ix0,jx0);
 303             dy00             = _mm_sub_ps(iy0,jy0);
 304             dz00             = _mm_sub_ps(iz0,jz0);
 305
 306             /* Calculate squared distance and things based on it */
 307             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 308
 309             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 310
 311             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 312
 313             /* Load parameters for j particles */
 314             vdwjidx0A        = 2*vdwtype[jnrA+0];
 315             vdwjidx0B        = 2*vdwtype[jnrB+0];
 316             vdwjidx0C        = 2*vdwtype[jnrC+0];
 317             vdwjidx0D        = 2*vdwtype[jnrD+0];
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             if (gmx_mm_any_lt(rsq00,rcutoff2))
 324             {
 325
 326             r00              = _mm_mul_ps(rsq00,rinv00);
 327             r00              = _mm_andnot_ps(dummy_mask,r00);
 328
 329             /* Compute parameters for interactions between i and j atoms */
 330             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 331                                          vdwparam+vdwioffset0+vdwjidx0B,
 332                                          vdwparam+vdwioffset0+vdwjidx0C,
 333                                          vdwparam+vdwioffset0+vdwjidx0D,
 334                                          &c6_00,&c12_00);
 335
 336             /* LENNARD-JONES DISPERSION/REPULSION */
 337
 338             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 339             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 340             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 341             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 342             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 343
 344             d                = _mm_sub_ps(r00,rswitch);
 345             d                = _mm_max_ps(d,_mm_setzero_ps());
 346             d2               = _mm_mul_ps(d,d);
 347             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 348
 349             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 350
 351             /* Evaluate switch function */
 352             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 353             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 354             vvdw             = _mm_mul_ps(vvdw,sw);
 355             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 359             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 360             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 361
 362             fscal            = fvdw;
 363
 364             fscal            = _mm_and_ps(fscal,cutoff_mask);
 365
 366             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 367
 368              /* Update vectorial force */
 369             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 370             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 371             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 372
 373             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 374             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 375             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 376             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 377             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 378                                                    _mm_mul_ps(dx00,fscal),
 379                                                    _mm_mul_ps(dy00,fscal),
 380                                                    _mm_mul_ps(dz00,fscal));
 381
 382             }
 383
 384             /* Inner loop uses 63 flops */
 385         }
 386
 387         /* End of innermost loop */
 388
 389         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 390                                               f+i_coord_offset,fshift+i_shift_offset);
 391
 392         ggid                        = gid[iidx];
 393         /* Update potential energies */
 394         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 395
 396         /* Increment number of inner iterations */
 397         inneriter                  += j_index_end - j_index_start;
 398
 399         /* Outer loop uses 7 flops */
 400     }
 401
 402     /* Increment number of outer iterations */
 403     outeriter        += nri;
 404
 405     /* Update outer/inner flops */
 406
 407     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*63);
 408 }
 409 /*
 410  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 411  * Electrostatics interaction: None
 412  * VdW interaction:            LennardJones
 413  * Geometry:                   Particle-Particle
 414  * Calculate force/pot:        Force
 415  */
 416 void
 417 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 418                     (t_nblist                    * gmx_restrict       nlist,
 419                      rvec                        * gmx_restrict          xx,
 420                      rvec                        * gmx_restrict          ff,
 421                      t_forcerec                  * gmx_restrict          fr,
 422                      t_mdatoms                   * gmx_restrict     mdatoms,
 423                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 424                      t_nrnb                      * gmx_restrict        nrnb)
 425 {
 426     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 427      * just 0 for non-waters.
 428      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 429      * jnr indices corresponding to data put in the four positions in the SIMD register.
 430      */
 431     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 432     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 433     int              jnrA,jnrB,jnrC,jnrD;
 434     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 435     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 436     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 437     real             rcutoff_scalar;
 438     real             *shiftvec,*fshift,*x,*f;
 439     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 440     real             scratch[4*DIM];
 441     __m128           fscal,rcutoff,rcutoff2,jidxall;
 442     int              vdwioffset0;
 443     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 444     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 445     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 446     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 447     int              nvdwtype;
 448     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 449     int              *vdwtype;
 450     real             *vdwparam;
 451     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 452     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 453     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 454     real             rswitch_scalar,d_scalar;
 455     __m128           dummy_mask,cutoff_mask;
 456     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 457     __m128           one     = _mm_set1_ps(1.0);
 458     __m128           two     = _mm_set1_ps(2.0);
 459     x                = xx[0];
 460     f                = ff[0];
 461
 462     nri              = nlist->nri;
 463     iinr             = nlist->iinr;
 464     jindex           = nlist->jindex;
 465     jjnr             = nlist->jjnr;
 466     shiftidx         = nlist->shift;
 467     gid              = nlist->gid;
 468     shiftvec         = fr->shift_vec[0];
 469     fshift           = fr->fshift[0];
 470     nvdwtype         = fr->ntype;
 471     vdwparam         = fr->nbfp;
 472     vdwtype          = mdatoms->typeA;
 473
 474     rcutoff_scalar   = fr->rvdw;
 475     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 476     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 477
 478     rswitch_scalar   = fr->rvdw_switch;
 479     rswitch          = _mm_set1_ps(rswitch_scalar);
 480     /* Setup switch parameters */
 481     d_scalar         = rcutoff_scalar-rswitch_scalar;
 482     d                = _mm_set1_ps(d_scalar);
 483     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 484     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 485     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 486     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 487     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 488     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 489
 490     /* Avoid stupid compiler warnings */
 491     jnrA = jnrB = jnrC = jnrD = 0;
 492     j_coord_offsetA = 0;
 493     j_coord_offsetB = 0;
 494     j_coord_offsetC = 0;
 495     j_coord_offsetD = 0;
 496
 497     outeriter        = 0;
 498     inneriter        = 0;
 499
 500     for(iidx=0;iidx<4*DIM;iidx++)
 501     {
 502         scratch[iidx] = 0.0;
 503     }
 504
 505     /* Start outer loop over neighborlists */
 506     for(iidx=0; iidx<nri; iidx++)
 507     {
 508         /* Load shift vector for this list */
 509         i_shift_offset   = DIM*shiftidx[iidx];
 510
 511         /* Load limits for loop over neighbors */
 512         j_index_start    = jindex[iidx];
 513         j_index_end      = jindex[iidx+1];
 514
 515         /* Get outer coordinate index */
 516         inr              = iinr[iidx];
 517         i_coord_offset   = DIM*inr;
 518
 519         /* Load i particle coords and add shift vector */
 520         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 521
 522         fix0             = _mm_setzero_ps();
 523         fiy0             = _mm_setzero_ps();
 524         fiz0             = _mm_setzero_ps();
 525
 526         /* Load parameters for i particles */
 527         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 528
 529         /* Start inner kernel loop */
 530         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 531         {
 532
 533             /* Get j neighbor index, and coordinate index */
 534             jnrA             = jjnr[jidx];
 535             jnrB             = jjnr[jidx+1];
 536             jnrC             = jjnr[jidx+2];
 537             jnrD             = jjnr[jidx+3];
 538             j_coord_offsetA  = DIM*jnrA;
 539             j_coord_offsetB  = DIM*jnrB;
 540             j_coord_offsetC  = DIM*jnrC;
 541             j_coord_offsetD  = DIM*jnrD;
 542
 543             /* load j atom coordinates */
 544             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 545                                               x+j_coord_offsetC,x+j_coord_offsetD,
 546                                               &jx0,&jy0,&jz0);
 547
 548             /* Calculate displacement vector */
 549             dx00             = _mm_sub_ps(ix0,jx0);
 550             dy00             = _mm_sub_ps(iy0,jy0);
 551             dz00             = _mm_sub_ps(iz0,jz0);
 552
 553             /* Calculate squared distance and things based on it */
 554             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 555
 556             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 557
 558             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 559
 560             /* Load parameters for j particles */
 561             vdwjidx0A        = 2*vdwtype[jnrA+0];
 562             vdwjidx0B        = 2*vdwtype[jnrB+0];
 563             vdwjidx0C        = 2*vdwtype[jnrC+0];
 564             vdwjidx0D        = 2*vdwtype[jnrD+0];
 565
 566             /**************************
 567              * CALCULATE INTERACTIONS *
 568              **************************/
 569
 570             if (gmx_mm_any_lt(rsq00,rcutoff2))
 571             {
 572
 573             r00              = _mm_mul_ps(rsq00,rinv00);
 574
 575             /* Compute parameters for interactions between i and j atoms */
 576             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 577                                          vdwparam+vdwioffset0+vdwjidx0B,
 578                                          vdwparam+vdwioffset0+vdwjidx0C,
 579                                          vdwparam+vdwioffset0+vdwjidx0D,
 580                                          &c6_00,&c12_00);
 581
 582             /* LENNARD-JONES DISPERSION/REPULSION */
 583
 584             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 585             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 586             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 587             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 588             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 589
 590             d                = _mm_sub_ps(r00,rswitch);
 591             d                = _mm_max_ps(d,_mm_setzero_ps());
 592             d2               = _mm_mul_ps(d,d);
 593             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 594
 595             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 596
 597             /* Evaluate switch function */
 598             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 599             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 600             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 601
 602             fscal            = fvdw;
 603
 604             fscal            = _mm_and_ps(fscal,cutoff_mask);
 605
 606              /* Update vectorial force */
 607             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 608             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 609             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 610
 611             fjptrA             = f+j_coord_offsetA;
 612             fjptrB             = f+j_coord_offsetB;
 613             fjptrC             = f+j_coord_offsetC;
 614             fjptrD             = f+j_coord_offsetD;
 615             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 616                                                    _mm_mul_ps(dx00,fscal),
 617                                                    _mm_mul_ps(dy00,fscal),
 618                                                    _mm_mul_ps(dz00,fscal));
 619
 620             }
 621
 622             /* Inner loop uses 59 flops */
 623         }
 624
 625         if(jidx<j_index_end)
 626         {
 627
 628             /* Get j neighbor index, and coordinate index */
 629             jnrlistA         = jjnr[jidx];
 630             jnrlistB         = jjnr[jidx+1];
 631             jnrlistC         = jjnr[jidx+2];
 632             jnrlistD         = jjnr[jidx+3];
 633             /* Sign of each element will be negative for non-real atoms.
 634              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 635              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 636              */
 637             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 638             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 639             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 640             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 641             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 642             j_coord_offsetA  = DIM*jnrA;
 643             j_coord_offsetB  = DIM*jnrB;
 644             j_coord_offsetC  = DIM*jnrC;
 645             j_coord_offsetD  = DIM*jnrD;
 646
 647             /* load j atom coordinates */
 648             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 649                                               x+j_coord_offsetC,x+j_coord_offsetD,
 650                                               &jx0,&jy0,&jz0);
 651
 652             /* Calculate displacement vector */
 653             dx00             = _mm_sub_ps(ix0,jx0);
 654             dy00             = _mm_sub_ps(iy0,jy0);
 655             dz00             = _mm_sub_ps(iz0,jz0);
 656
 657             /* Calculate squared distance and things based on it */
 658             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 659
 660             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 661
 662             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 663
 664             /* Load parameters for j particles */
 665             vdwjidx0A        = 2*vdwtype[jnrA+0];
 666             vdwjidx0B        = 2*vdwtype[jnrB+0];
 667             vdwjidx0C        = 2*vdwtype[jnrC+0];
 668             vdwjidx0D        = 2*vdwtype[jnrD+0];
 669
 670             /**************************
 671              * CALCULATE INTERACTIONS *
 672              **************************/
 673
 674             if (gmx_mm_any_lt(rsq00,rcutoff2))
 675             {
 676
 677             r00              = _mm_mul_ps(rsq00,rinv00);
 678             r00              = _mm_andnot_ps(dummy_mask,r00);
 679
 680             /* Compute parameters for interactions between i and j atoms */
 681             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 682                                          vdwparam+vdwioffset0+vdwjidx0B,
 683                                          vdwparam+vdwioffset0+vdwjidx0C,
 684                                          vdwparam+vdwioffset0+vdwjidx0D,
 685                                          &c6_00,&c12_00);
 686
 687             /* LENNARD-JONES DISPERSION/REPULSION */
 688
 689             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 690             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 691             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 692             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 693             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 694
 695             d                = _mm_sub_ps(r00,rswitch);
 696             d                = _mm_max_ps(d,_mm_setzero_ps());
 697             d2               = _mm_mul_ps(d,d);
 698             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 699
 700             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 701
 702             /* Evaluate switch function */
 703             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 704             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 705             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 706
 707             fscal            = fvdw;
 708
 709             fscal            = _mm_and_ps(fscal,cutoff_mask);
 710
 711             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 712
 713              /* Update vectorial force */
 714             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 715             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 716             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 717
 718             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 719             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 720             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 721             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 722             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 723                                                    _mm_mul_ps(dx00,fscal),
 724                                                    _mm_mul_ps(dy00,fscal),
 725                                                    _mm_mul_ps(dz00,fscal));
 726
 727             }
 728
 729             /* Inner loop uses 60 flops */
 730         }
 731
 732         /* End of innermost loop */
 733
 734         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 735                                               f+i_coord_offset,fshift+i_shift_offset);
 736
 737         /* Increment number of inner iterations */
 738         inneriter                  += j_index_end - j_index_start;
 739
 740         /* Outer loop uses 6 flops */
 741     }
 742
 743     /* Increment number of outer iterations */
 744     outeriter        += nri;
 745
 746     /* Update outer/inner flops */
 747
 748     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*60);
 749 }