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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/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_VdwLJSw_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_VdwLJSw_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           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  97     real             rswitch_scalar,d_scalar;
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     nvdwtype         = fr->ntype;
 114     vdwparam         = fr->nbfp;
 115     vdwtype          = mdatoms->typeA;
 116
 117     rcutoff_scalar   = fr->rvdw;
 118     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 119     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 120
 121     rswitch_scalar   = fr->rvdw_switch;
 122     rswitch          = _mm_set1_ps(rswitch_scalar);
 123     /* Setup switch parameters */
 124     d_scalar         = rcutoff_scalar-rswitch_scalar;
 125     d                = _mm_set1_ps(d_scalar);
 126     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 127     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 128     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 129     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 130     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 131     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 132
 133     /* Avoid stupid compiler warnings */
 134     jnrA = jnrB = jnrC = jnrD = 0;
 135     j_coord_offsetA = 0;
 136     j_coord_offsetB = 0;
 137     j_coord_offsetC = 0;
 138     j_coord_offsetD = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     for(iidx=0;iidx<4*DIM;iidx++)
 144     {
 145         scratch[iidx] = 0.0;
 146     }
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 164
 165         fix0             = _mm_setzero_ps();
 166         fiy0             = _mm_setzero_ps();
 167         fiz0             = _mm_setzero_ps();
 168
 169         /* Load parameters for i particles */
 170         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         vvdwsum          = _mm_setzero_ps();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                               x+j_coord_offsetC,x+j_coord_offsetD,
 192                                               &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm_sub_ps(ix0,jx0);
 196             dy00             = _mm_sub_ps(iy0,jy0);
 197             dz00             = _mm_sub_ps(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 203
 204             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             vdwjidx0A        = 2*vdwtype[jnrA+0];
 208             vdwjidx0B        = 2*vdwtype[jnrB+0];
 209             vdwjidx0C        = 2*vdwtype[jnrC+0];
 210             vdwjidx0D        = 2*vdwtype[jnrD+0];
 211
 212             /**************************
 213              * CALCULATE INTERACTIONS *
 214              **************************/
 215
 216             if (gmx_mm_any_lt(rsq00,rcutoff2))
 217             {
 218
 219             r00              = _mm_mul_ps(rsq00,rinv00);
 220
 221             /* Compute parameters for interactions between i and j atoms */
 222             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 223                                          vdwparam+vdwioffset0+vdwjidx0B,
 224                                          vdwparam+vdwioffset0+vdwjidx0C,
 225                                          vdwparam+vdwioffset0+vdwjidx0D,
 226                                          &c6_00,&c12_00);
 227
 228             /* LENNARD-JONES DISPERSION/REPULSION */
 229
 230             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 231             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 232             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 233             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 234             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 235
 236             d                = _mm_sub_ps(r00,rswitch);
 237             d                = _mm_max_ps(d,_mm_setzero_ps());
 238             d2               = _mm_mul_ps(d,d);
 239             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 240
 241             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 242
 243             /* Evaluate switch function */
 244             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 245             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 246             vvdw             = _mm_mul_ps(vvdw,sw);
 247             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 251             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 252
 253             fscal            = fvdw;
 254
 255             fscal            = _mm_and_ps(fscal,cutoff_mask);
 256
 257              /* Update vectorial force */
 258             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 259             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 260             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 261
 262             fjptrA             = f+j_coord_offsetA;
 263             fjptrB             = f+j_coord_offsetB;
 264             fjptrC             = f+j_coord_offsetC;
 265             fjptrD             = f+j_coord_offsetD;
 266             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 267                                                    _mm_mul_ps(dx00,fscal),
 268                                                    _mm_mul_ps(dy00,fscal),
 269                                                    _mm_mul_ps(dz00,fscal));
 270
 271             }
 272
 273             /* Inner loop uses 62 flops */
 274         }
 275
 276         if(jidx<j_index_end)
 277         {
 278
 279             /* Get j neighbor index, and coordinate index */
 280             jnrlistA         = jjnr[jidx];
 281             jnrlistB         = jjnr[jidx+1];
 282             jnrlistC         = jjnr[jidx+2];
 283             jnrlistD         = jjnr[jidx+3];
 284             /* Sign of each element will be negative for non-real atoms.
 285              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 286              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 287              */
 288             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 289             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 290             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 291             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 292             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 293             j_coord_offsetA  = DIM*jnrA;
 294             j_coord_offsetB  = DIM*jnrB;
 295             j_coord_offsetC  = DIM*jnrC;
 296             j_coord_offsetD  = DIM*jnrD;
 297
 298             /* load j atom coordinates */
 299             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 300                                               x+j_coord_offsetC,x+j_coord_offsetD,
 301                                               &jx0,&jy0,&jz0);
 302
 303             /* Calculate displacement vector */
 304             dx00             = _mm_sub_ps(ix0,jx0);
 305             dy00             = _mm_sub_ps(iy0,jy0);
 306             dz00             = _mm_sub_ps(iz0,jz0);
 307
 308             /* Calculate squared distance and things based on it */
 309             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 310
 311             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 312
 313             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 314
 315             /* Load parameters for j particles */
 316             vdwjidx0A        = 2*vdwtype[jnrA+0];
 317             vdwjidx0B        = 2*vdwtype[jnrB+0];
 318             vdwjidx0C        = 2*vdwtype[jnrC+0];
 319             vdwjidx0D        = 2*vdwtype[jnrD+0];
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             if (gmx_mm_any_lt(rsq00,rcutoff2))
 326             {
 327
 328             r00              = _mm_mul_ps(rsq00,rinv00);
 329             r00              = _mm_andnot_ps(dummy_mask,r00);
 330
 331             /* Compute parameters for interactions between i and j atoms */
 332             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 333                                          vdwparam+vdwioffset0+vdwjidx0B,
 334                                          vdwparam+vdwioffset0+vdwjidx0C,
 335                                          vdwparam+vdwioffset0+vdwjidx0D,
 336                                          &c6_00,&c12_00);
 337
 338             /* LENNARD-JONES DISPERSION/REPULSION */
 339
 340             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 341             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 342             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 343             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 344             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 345
 346             d                = _mm_sub_ps(r00,rswitch);
 347             d                = _mm_max_ps(d,_mm_setzero_ps());
 348             d2               = _mm_mul_ps(d,d);
 349             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 350
 351             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 352
 353             /* Evaluate switch function */
 354             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 355             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 356             vvdw             = _mm_mul_ps(vvdw,sw);
 357             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 361             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 362             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 363
 364             fscal            = fvdw;
 365
 366             fscal            = _mm_and_ps(fscal,cutoff_mask);
 367
 368             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 369
 370              /* Update vectorial force */
 371             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 372             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 373             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 374
 375             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 376             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 377             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 378             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 379             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 380                                                    _mm_mul_ps(dx00,fscal),
 381                                                    _mm_mul_ps(dy00,fscal),
 382                                                    _mm_mul_ps(dz00,fscal));
 383
 384             }
 385
 386             /* Inner loop uses 63 flops */
 387         }
 388
 389         /* End of innermost loop */
 390
 391         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 392                                               f+i_coord_offset,fshift+i_shift_offset);
 393
 394         ggid                        = gid[iidx];
 395         /* Update potential energies */
 396         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 397
 398         /* Increment number of inner iterations */
 399         inneriter                  += j_index_end - j_index_start;
 400
 401         /* Outer loop uses 7 flops */
 402     }
 403
 404     /* Increment number of outer iterations */
 405     outeriter        += nri;
 406
 407     /* Update outer/inner flops */
 408
 409     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*63);
 410 }
 411 /*
 412  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 413  * Electrostatics interaction: None
 414  * VdW interaction:            LennardJones
 415  * Geometry:                   Particle-Particle
 416  * Calculate force/pot:        Force
 417  */
 418 void
 419 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 420                     (t_nblist                    * gmx_restrict       nlist,
 421                      rvec                        * gmx_restrict          xx,
 422                      rvec                        * gmx_restrict          ff,
 423                      t_forcerec                  * gmx_restrict          fr,
 424                      t_mdatoms                   * gmx_restrict     mdatoms,
 425                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 426                      t_nrnb                      * gmx_restrict        nrnb)
 427 {
 428     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 429      * just 0 for non-waters.
 430      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 431      * jnr indices corresponding to data put in the four positions in the SIMD register.
 432      */
 433     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 434     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 435     int              jnrA,jnrB,jnrC,jnrD;
 436     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 437     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 438     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 439     real             rcutoff_scalar;
 440     real             *shiftvec,*fshift,*x,*f;
 441     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 442     real             scratch[4*DIM];
 443     __m128           fscal,rcutoff,rcutoff2,jidxall;
 444     int              vdwioffset0;
 445     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 446     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 447     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 448     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 449     int              nvdwtype;
 450     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 451     int              *vdwtype;
 452     real             *vdwparam;
 453     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 454     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 455     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 456     real             rswitch_scalar,d_scalar;
 457     __m128           dummy_mask,cutoff_mask;
 458     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 459     __m128           one     = _mm_set1_ps(1.0);
 460     __m128           two     = _mm_set1_ps(2.0);
 461     x                = xx[0];
 462     f                = ff[0];
 463
 464     nri              = nlist->nri;
 465     iinr             = nlist->iinr;
 466     jindex           = nlist->jindex;
 467     jjnr             = nlist->jjnr;
 468     shiftidx         = nlist->shift;
 469     gid              = nlist->gid;
 470     shiftvec         = fr->shift_vec[0];
 471     fshift           = fr->fshift[0];
 472     nvdwtype         = fr->ntype;
 473     vdwparam         = fr->nbfp;
 474     vdwtype          = mdatoms->typeA;
 475
 476     rcutoff_scalar   = fr->rvdw;
 477     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 478     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 479
 480     rswitch_scalar   = fr->rvdw_switch;
 481     rswitch          = _mm_set1_ps(rswitch_scalar);
 482     /* Setup switch parameters */
 483     d_scalar         = rcutoff_scalar-rswitch_scalar;
 484     d                = _mm_set1_ps(d_scalar);
 485     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 486     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 487     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 488     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 489     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 490     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 491
 492     /* Avoid stupid compiler warnings */
 493     jnrA = jnrB = jnrC = jnrD = 0;
 494     j_coord_offsetA = 0;
 495     j_coord_offsetB = 0;
 496     j_coord_offsetC = 0;
 497     j_coord_offsetD = 0;
 498
 499     outeriter        = 0;
 500     inneriter        = 0;
 501
 502     for(iidx=0;iidx<4*DIM;iidx++)
 503     {
 504         scratch[iidx] = 0.0;
 505     }
 506
 507     /* Start outer loop over neighborlists */
 508     for(iidx=0; iidx<nri; iidx++)
 509     {
 510         /* Load shift vector for this list */
 511         i_shift_offset   = DIM*shiftidx[iidx];
 512
 513         /* Load limits for loop over neighbors */
 514         j_index_start    = jindex[iidx];
 515         j_index_end      = jindex[iidx+1];
 516
 517         /* Get outer coordinate index */
 518         inr              = iinr[iidx];
 519         i_coord_offset   = DIM*inr;
 520
 521         /* Load i particle coords and add shift vector */
 522         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 523
 524         fix0             = _mm_setzero_ps();
 525         fiy0             = _mm_setzero_ps();
 526         fiz0             = _mm_setzero_ps();
 527
 528         /* Load parameters for i particles */
 529         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 530
 531         /* Start inner kernel loop */
 532         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 533         {
 534
 535             /* Get j neighbor index, and coordinate index */
 536             jnrA             = jjnr[jidx];
 537             jnrB             = jjnr[jidx+1];
 538             jnrC             = jjnr[jidx+2];
 539             jnrD             = jjnr[jidx+3];
 540             j_coord_offsetA  = DIM*jnrA;
 541             j_coord_offsetB  = DIM*jnrB;
 542             j_coord_offsetC  = DIM*jnrC;
 543             j_coord_offsetD  = DIM*jnrD;
 544
 545             /* load j atom coordinates */
 546             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 547                                               x+j_coord_offsetC,x+j_coord_offsetD,
 548                                               &jx0,&jy0,&jz0);
 549
 550             /* Calculate displacement vector */
 551             dx00             = _mm_sub_ps(ix0,jx0);
 552             dy00             = _mm_sub_ps(iy0,jy0);
 553             dz00             = _mm_sub_ps(iz0,jz0);
 554
 555             /* Calculate squared distance and things based on it */
 556             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 557
 558             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 559
 560             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 561
 562             /* Load parameters for j particles */
 563             vdwjidx0A        = 2*vdwtype[jnrA+0];
 564             vdwjidx0B        = 2*vdwtype[jnrB+0];
 565             vdwjidx0C        = 2*vdwtype[jnrC+0];
 566             vdwjidx0D        = 2*vdwtype[jnrD+0];
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             if (gmx_mm_any_lt(rsq00,rcutoff2))
 573             {
 574
 575             r00              = _mm_mul_ps(rsq00,rinv00);
 576
 577             /* Compute parameters for interactions between i and j atoms */
 578             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 579                                          vdwparam+vdwioffset0+vdwjidx0B,
 580                                          vdwparam+vdwioffset0+vdwjidx0C,
 581                                          vdwparam+vdwioffset0+vdwjidx0D,
 582                                          &c6_00,&c12_00);
 583
 584             /* LENNARD-JONES DISPERSION/REPULSION */
 585
 586             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 587             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 588             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 589             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 590             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 591
 592             d                = _mm_sub_ps(r00,rswitch);
 593             d                = _mm_max_ps(d,_mm_setzero_ps());
 594             d2               = _mm_mul_ps(d,d);
 595             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 596
 597             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 598
 599             /* Evaluate switch function */
 600             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 601             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 602             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 603
 604             fscal            = fvdw;
 605
 606             fscal            = _mm_and_ps(fscal,cutoff_mask);
 607
 608              /* Update vectorial force */
 609             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 610             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 611             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 612
 613             fjptrA             = f+j_coord_offsetA;
 614             fjptrB             = f+j_coord_offsetB;
 615             fjptrC             = f+j_coord_offsetC;
 616             fjptrD             = f+j_coord_offsetD;
 617             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 618                                                    _mm_mul_ps(dx00,fscal),
 619                                                    _mm_mul_ps(dy00,fscal),
 620                                                    _mm_mul_ps(dz00,fscal));
 621
 622             }
 623
 624             /* Inner loop uses 59 flops */
 625         }
 626
 627         if(jidx<j_index_end)
 628         {
 629
 630             /* Get j neighbor index, and coordinate index */
 631             jnrlistA         = jjnr[jidx];
 632             jnrlistB         = jjnr[jidx+1];
 633             jnrlistC         = jjnr[jidx+2];
 634             jnrlistD         = jjnr[jidx+3];
 635             /* Sign of each element will be negative for non-real atoms.
 636              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 637              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 638              */
 639             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 640             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 641             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 642             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 643             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 644             j_coord_offsetA  = DIM*jnrA;
 645             j_coord_offsetB  = DIM*jnrB;
 646             j_coord_offsetC  = DIM*jnrC;
 647             j_coord_offsetD  = DIM*jnrD;
 648
 649             /* load j atom coordinates */
 650             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 651                                               x+j_coord_offsetC,x+j_coord_offsetD,
 652                                               &jx0,&jy0,&jz0);
 653
 654             /* Calculate displacement vector */
 655             dx00             = _mm_sub_ps(ix0,jx0);
 656             dy00             = _mm_sub_ps(iy0,jy0);
 657             dz00             = _mm_sub_ps(iz0,jz0);
 658
 659             /* Calculate squared distance and things based on it */
 660             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 661
 662             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 663
 664             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 665
 666             /* Load parameters for j particles */
 667             vdwjidx0A        = 2*vdwtype[jnrA+0];
 668             vdwjidx0B        = 2*vdwtype[jnrB+0];
 669             vdwjidx0C        = 2*vdwtype[jnrC+0];
 670             vdwjidx0D        = 2*vdwtype[jnrD+0];
 671
 672             /**************************
 673              * CALCULATE INTERACTIONS *
 674              **************************/
 675
 676             if (gmx_mm_any_lt(rsq00,rcutoff2))
 677             {
 678
 679             r00              = _mm_mul_ps(rsq00,rinv00);
 680             r00              = _mm_andnot_ps(dummy_mask,r00);
 681
 682             /* Compute parameters for interactions between i and j atoms */
 683             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 684                                          vdwparam+vdwioffset0+vdwjidx0B,
 685                                          vdwparam+vdwioffset0+vdwjidx0C,
 686                                          vdwparam+vdwioffset0+vdwjidx0D,
 687                                          &c6_00,&c12_00);
 688
 689             /* LENNARD-JONES DISPERSION/REPULSION */
 690
 691             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 692             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 693             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 694             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 695             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 696
 697             d                = _mm_sub_ps(r00,rswitch);
 698             d                = _mm_max_ps(d,_mm_setzero_ps());
 699             d2               = _mm_mul_ps(d,d);
 700             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 701
 702             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 703
 704             /* Evaluate switch function */
 705             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 706             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 707             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 708
 709             fscal            = fvdw;
 710
 711             fscal            = _mm_and_ps(fscal,cutoff_mask);
 712
 713             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 714
 715              /* Update vectorial force */
 716             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 717             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 718             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 719
 720             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 721             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 722             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 723             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 724             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 725                                                    _mm_mul_ps(dx00,fscal),
 726                                                    _mm_mul_ps(dy00,fscal),
 727                                                    _mm_mul_ps(dz00,fscal));
 728
 729             }
 730
 731             /* Inner loop uses 60 flops */
 732         }
 733
 734         /* End of innermost loop */
 735
 736         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 737                                               f+i_coord_offset,fshift+i_shift_offset);
 738
 739         /* Increment number of inner iterations */
 740         inneriter                  += j_index_end - j_index_start;
 741
 742         /* Outer loop uses 6 flops */
 743     }
 744
 745     /* Increment number of outer iterations */
 746     outeriter        += nri;
 747
 748     /* Update outer/inner flops */
 749
 750     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*60);
 751 }