src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwCSTab_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_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: None
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecNone_VdwCSTab_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     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  99     real             *vftab;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     vftab            = kernel_data->table_vdw->data;
 120     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 121
 122     /* Avoid stupid compiler warnings */
 123     jnrA = jnrB = jnrC = jnrD = 0;
 124     j_coord_offsetA = 0;
 125     j_coord_offsetB = 0;
 126     j_coord_offsetC = 0;
 127     j_coord_offsetD = 0;
 128
 129     outeriter        = 0;
 130     inneriter        = 0;
 131
 132     for(iidx=0;iidx<4*DIM;iidx++)
 133     {
 134         scratch[iidx] = 0.0;
 135     }
 136
 137     /* Start outer loop over neighborlists */
 138     for(iidx=0; iidx<nri; iidx++)
 139     {
 140         /* Load shift vector for this list */
 141         i_shift_offset   = DIM*shiftidx[iidx];
 142
 143         /* Load limits for loop over neighbors */
 144         j_index_start    = jindex[iidx];
 145         j_index_end      = jindex[iidx+1];
 146
 147         /* Get outer coordinate index */
 148         inr              = iinr[iidx];
 149         i_coord_offset   = DIM*inr;
 150
 151         /* Load i particle coords and add shift vector */
 152         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 153
 154         fix0             = _mm_setzero_ps();
 155         fiy0             = _mm_setzero_ps();
 156         fiz0             = _mm_setzero_ps();
 157
 158         /* Load parameters for i particles */
 159         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         vvdwsum          = _mm_setzero_ps();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             jnrC             = jjnr[jidx+2];
 172             jnrD             = jjnr[jidx+3];
 173             j_coord_offsetA  = DIM*jnrA;
 174             j_coord_offsetB  = DIM*jnrB;
 175             j_coord_offsetC  = DIM*jnrC;
 176             j_coord_offsetD  = DIM*jnrD;
 177
 178             /* load j atom coordinates */
 179             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                               x+j_coord_offsetC,x+j_coord_offsetD,
 181                                               &jx0,&jy0,&jz0);
 182
 183             /* Calculate displacement vector */
 184             dx00             = _mm_sub_ps(ix0,jx0);
 185             dy00             = _mm_sub_ps(iy0,jy0);
 186             dz00             = _mm_sub_ps(iz0,jz0);
 187
 188             /* Calculate squared distance and things based on it */
 189             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 190
 191             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 192
 193             /* 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             r00              = _mm_mul_ps(rsq00,rinv00);
 204
 205             /* Compute parameters for interactions between i and j atoms */
 206             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 207                                          vdwparam+vdwioffset0+vdwjidx0B,
 208                                          vdwparam+vdwioffset0+vdwjidx0C,
 209                                          vdwparam+vdwioffset0+vdwjidx0D,
 210                                          &c6_00,&c12_00);
 211
 212             /* Calculate table index by multiplying r with table scale and truncate to integer */
 213             rt               = _mm_mul_ps(r00,vftabscale);
 214             vfitab           = _mm_cvttps_epi32(rt);
 215 #ifdef __XOP__
 216             vfeps            = _mm_frcz_ps(rt);
 217 #else
 218             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 219 #endif
 220             twovfeps         = _mm_add_ps(vfeps,vfeps);
 221             vfitab           = _mm_slli_epi32(vfitab,3);
 222
 223             /* CUBIC SPLINE TABLE DISPERSION */
 224             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 225             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 226             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 227             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 228             _MM_TRANSPOSE4_PS(Y,F,G,H);
 229             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 230             VV               = _mm_macc_ps(vfeps,Fp,Y);
 231             vvdw6            = _mm_mul_ps(c6_00,VV);
 232             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 233             fvdw6            = _mm_mul_ps(c6_00,FF);
 234
 235             /* CUBIC SPLINE TABLE REPULSION */
 236             vfitab           = _mm_add_epi32(vfitab,ifour);
 237             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 238             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 239             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 240             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 241             _MM_TRANSPOSE4_PS(Y,F,G,H);
 242             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 243             VV               = _mm_macc_ps(vfeps,Fp,Y);
 244             vvdw12           = _mm_mul_ps(c12_00,VV);
 245             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 246             fvdw12           = _mm_mul_ps(c12_00,FF);
 247             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 248             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 249
 250             /* Update potential sum for this i atom from the interaction with this j atom. */
 251             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 252
 253             fscal            = fvdw;
 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             /* Inner loop uses 59 flops */
 270         }
 271
 272         if(jidx<j_index_end)
 273         {
 274
 275             /* Get j neighbor index, and coordinate index */
 276             jnrlistA         = jjnr[jidx];
 277             jnrlistB         = jjnr[jidx+1];
 278             jnrlistC         = jjnr[jidx+2];
 279             jnrlistD         = jjnr[jidx+3];
 280             /* Sign of each element will be negative for non-real atoms.
 281              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 282              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 283              */
 284             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 285             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 286             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 287             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 288             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 289             j_coord_offsetA  = DIM*jnrA;
 290             j_coord_offsetB  = DIM*jnrB;
 291             j_coord_offsetC  = DIM*jnrC;
 292             j_coord_offsetD  = DIM*jnrD;
 293
 294             /* load j atom coordinates */
 295             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 296                                               x+j_coord_offsetC,x+j_coord_offsetD,
 297                                               &jx0,&jy0,&jz0);
 298
 299             /* Calculate displacement vector */
 300             dx00             = _mm_sub_ps(ix0,jx0);
 301             dy00             = _mm_sub_ps(iy0,jy0);
 302             dz00             = _mm_sub_ps(iz0,jz0);
 303
 304             /* Calculate squared distance and things based on it */
 305             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 306
 307             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 308
 309             /* Load parameters for j particles */
 310             vdwjidx0A        = 2*vdwtype[jnrA+0];
 311             vdwjidx0B        = 2*vdwtype[jnrB+0];
 312             vdwjidx0C        = 2*vdwtype[jnrC+0];
 313             vdwjidx0D        = 2*vdwtype[jnrD+0];
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             r00              = _mm_mul_ps(rsq00,rinv00);
 320             r00              = _mm_andnot_ps(dummy_mask,r00);
 321
 322             /* Compute parameters for interactions between i and j atoms */
 323             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 324                                          vdwparam+vdwioffset0+vdwjidx0B,
 325                                          vdwparam+vdwioffset0+vdwjidx0C,
 326                                          vdwparam+vdwioffset0+vdwjidx0D,
 327                                          &c6_00,&c12_00);
 328
 329             /* Calculate table index by multiplying r with table scale and truncate to integer */
 330             rt               = _mm_mul_ps(r00,vftabscale);
 331             vfitab           = _mm_cvttps_epi32(rt);
 332 #ifdef __XOP__
 333             vfeps            = _mm_frcz_ps(rt);
 334 #else
 335             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 336 #endif
 337             twovfeps         = _mm_add_ps(vfeps,vfeps);
 338             vfitab           = _mm_slli_epi32(vfitab,3);
 339
 340             /* CUBIC SPLINE TABLE DISPERSION */
 341             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 342             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 343             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 344             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 345             _MM_TRANSPOSE4_PS(Y,F,G,H);
 346             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 347             VV               = _mm_macc_ps(vfeps,Fp,Y);
 348             vvdw6            = _mm_mul_ps(c6_00,VV);
 349             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 350             fvdw6            = _mm_mul_ps(c6_00,FF);
 351
 352             /* CUBIC SPLINE TABLE REPULSION */
 353             vfitab           = _mm_add_epi32(vfitab,ifour);
 354             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 355             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 356             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 357             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 358             _MM_TRANSPOSE4_PS(Y,F,G,H);
 359             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 360             VV               = _mm_macc_ps(vfeps,Fp,Y);
 361             vvdw12           = _mm_mul_ps(c12_00,VV);
 362             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 363             fvdw12           = _mm_mul_ps(c12_00,FF);
 364             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 365             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 369             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 370
 371             fscal            = fvdw;
 372
 373             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 374
 375              /* Update vectorial force */
 376             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 377             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 378             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 379
 380             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 381             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 382             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 383             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 384             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 385                                                    _mm_mul_ps(dx00,fscal),
 386                                                    _mm_mul_ps(dy00,fscal),
 387                                                    _mm_mul_ps(dz00,fscal));
 388
 389             /* Inner loop uses 60 flops */
 390         }
 391
 392         /* End of innermost loop */
 393
 394         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 395                                               f+i_coord_offset,fshift+i_shift_offset);
 396
 397         ggid                        = gid[iidx];
 398         /* Update potential energies */
 399         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 400
 401         /* Increment number of inner iterations */
 402         inneriter                  += j_index_end - j_index_start;
 403
 404         /* Outer loop uses 7 flops */
 405     }
 406
 407     /* Increment number of outer iterations */
 408     outeriter        += nri;
 409
 410     /* Update outer/inner flops */
 411
 412     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 413 }
 414 /*
 415  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 416  * Electrostatics interaction: None
 417  * VdW interaction:            CubicSplineTable
 418  * Geometry:                   Particle-Particle
 419  * Calculate force/pot:        Force
 420  */
 421 void
 422 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 423                     (t_nblist                    * gmx_restrict       nlist,
 424                      rvec                        * gmx_restrict          xx,
 425                      rvec                        * gmx_restrict          ff,
 426                      t_forcerec                  * gmx_restrict          fr,
 427                      t_mdatoms                   * gmx_restrict     mdatoms,
 428                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 429                      t_nrnb                      * gmx_restrict        nrnb)
 430 {
 431     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 432      * just 0 for non-waters.
 433      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 434      * jnr indices corresponding to data put in the four positions in the SIMD register.
 435      */
 436     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 437     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 438     int              jnrA,jnrB,jnrC,jnrD;
 439     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 440     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 441     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 442     real             rcutoff_scalar;
 443     real             *shiftvec,*fshift,*x,*f;
 444     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 445     real             scratch[4*DIM];
 446     __m128           fscal,rcutoff,rcutoff2,jidxall;
 447     int              vdwioffset0;
 448     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 449     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 450     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 451     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 452     int              nvdwtype;
 453     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 454     int              *vdwtype;
 455     real             *vdwparam;
 456     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 457     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 458     __m128i          vfitab;
 459     __m128i          ifour       = _mm_set1_epi32(4);
 460     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 461     real             *vftab;
 462     __m128           dummy_mask,cutoff_mask;
 463     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 464     __m128           one     = _mm_set1_ps(1.0);
 465     __m128           two     = _mm_set1_ps(2.0);
 466     x                = xx[0];
 467     f                = ff[0];
 468
 469     nri              = nlist->nri;
 470     iinr             = nlist->iinr;
 471     jindex           = nlist->jindex;
 472     jjnr             = nlist->jjnr;
 473     shiftidx         = nlist->shift;
 474     gid              = nlist->gid;
 475     shiftvec         = fr->shift_vec[0];
 476     fshift           = fr->fshift[0];
 477     nvdwtype         = fr->ntype;
 478     vdwparam         = fr->nbfp;
 479     vdwtype          = mdatoms->typeA;
 480
 481     vftab            = kernel_data->table_vdw->data;
 482     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 483
 484     /* Avoid stupid compiler warnings */
 485     jnrA = jnrB = jnrC = jnrD = 0;
 486     j_coord_offsetA = 0;
 487     j_coord_offsetB = 0;
 488     j_coord_offsetC = 0;
 489     j_coord_offsetD = 0;
 490
 491     outeriter        = 0;
 492     inneriter        = 0;
 493
 494     for(iidx=0;iidx<4*DIM;iidx++)
 495     {
 496         scratch[iidx] = 0.0;
 497     }
 498
 499     /* Start outer loop over neighborlists */
 500     for(iidx=0; iidx<nri; iidx++)
 501     {
 502         /* Load shift vector for this list */
 503         i_shift_offset   = DIM*shiftidx[iidx];
 504
 505         /* Load limits for loop over neighbors */
 506         j_index_start    = jindex[iidx];
 507         j_index_end      = jindex[iidx+1];
 508
 509         /* Get outer coordinate index */
 510         inr              = iinr[iidx];
 511         i_coord_offset   = DIM*inr;
 512
 513         /* Load i particle coords and add shift vector */
 514         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 515
 516         fix0             = _mm_setzero_ps();
 517         fiy0             = _mm_setzero_ps();
 518         fiz0             = _mm_setzero_ps();
 519
 520         /* Load parameters for i particles */
 521         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 522
 523         /* Start inner kernel loop */
 524         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 525         {
 526
 527             /* Get j neighbor index, and coordinate index */
 528             jnrA             = jjnr[jidx];
 529             jnrB             = jjnr[jidx+1];
 530             jnrC             = jjnr[jidx+2];
 531             jnrD             = jjnr[jidx+3];
 532             j_coord_offsetA  = DIM*jnrA;
 533             j_coord_offsetB  = DIM*jnrB;
 534             j_coord_offsetC  = DIM*jnrC;
 535             j_coord_offsetD  = DIM*jnrD;
 536
 537             /* load j atom coordinates */
 538             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 539                                               x+j_coord_offsetC,x+j_coord_offsetD,
 540                                               &jx0,&jy0,&jz0);
 541
 542             /* Calculate displacement vector */
 543             dx00             = _mm_sub_ps(ix0,jx0);
 544             dy00             = _mm_sub_ps(iy0,jy0);
 545             dz00             = _mm_sub_ps(iz0,jz0);
 546
 547             /* Calculate squared distance and things based on it */
 548             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 549
 550             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 551
 552             /* Load parameters for j particles */
 553             vdwjidx0A        = 2*vdwtype[jnrA+0];
 554             vdwjidx0B        = 2*vdwtype[jnrB+0];
 555             vdwjidx0C        = 2*vdwtype[jnrC+0];
 556             vdwjidx0D        = 2*vdwtype[jnrD+0];
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             r00              = _mm_mul_ps(rsq00,rinv00);
 563
 564             /* Compute parameters for interactions between i and j atoms */
 565             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 566                                          vdwparam+vdwioffset0+vdwjidx0B,
 567                                          vdwparam+vdwioffset0+vdwjidx0C,
 568                                          vdwparam+vdwioffset0+vdwjidx0D,
 569                                          &c6_00,&c12_00);
 570
 571             /* Calculate table index by multiplying r with table scale and truncate to integer */
 572             rt               = _mm_mul_ps(r00,vftabscale);
 573             vfitab           = _mm_cvttps_epi32(rt);
 574 #ifdef __XOP__
 575             vfeps            = _mm_frcz_ps(rt);
 576 #else
 577             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 578 #endif
 579             twovfeps         = _mm_add_ps(vfeps,vfeps);
 580             vfitab           = _mm_slli_epi32(vfitab,3);
 581
 582             /* CUBIC SPLINE TABLE DISPERSION */
 583             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 584             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 585             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 586             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 587             _MM_TRANSPOSE4_PS(Y,F,G,H);
 588             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 589             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 590             fvdw6            = _mm_mul_ps(c6_00,FF);
 591
 592             /* CUBIC SPLINE TABLE REPULSION */
 593             vfitab           = _mm_add_epi32(vfitab,ifour);
 594             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 595             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 596             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 597             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 598             _MM_TRANSPOSE4_PS(Y,F,G,H);
 599             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 600             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 601             fvdw12           = _mm_mul_ps(c12_00,FF);
 602             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 603
 604             fscal            = fvdw;
 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             /* Inner loop uses 51 flops */
 621         }
 622
 623         if(jidx<j_index_end)
 624         {
 625
 626             /* Get j neighbor index, and coordinate index */
 627             jnrlistA         = jjnr[jidx];
 628             jnrlistB         = jjnr[jidx+1];
 629             jnrlistC         = jjnr[jidx+2];
 630             jnrlistD         = jjnr[jidx+3];
 631             /* Sign of each element will be negative for non-real atoms.
 632              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 633              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 634              */
 635             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 636             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 637             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 638             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 639             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 640             j_coord_offsetA  = DIM*jnrA;
 641             j_coord_offsetB  = DIM*jnrB;
 642             j_coord_offsetC  = DIM*jnrC;
 643             j_coord_offsetD  = DIM*jnrD;
 644
 645             /* load j atom coordinates */
 646             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 647                                               x+j_coord_offsetC,x+j_coord_offsetD,
 648                                               &jx0,&jy0,&jz0);
 649
 650             /* Calculate displacement vector */
 651             dx00             = _mm_sub_ps(ix0,jx0);
 652             dy00             = _mm_sub_ps(iy0,jy0);
 653             dz00             = _mm_sub_ps(iz0,jz0);
 654
 655             /* Calculate squared distance and things based on it */
 656             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 657
 658             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 659
 660             /* Load parameters for j particles */
 661             vdwjidx0A        = 2*vdwtype[jnrA+0];
 662             vdwjidx0B        = 2*vdwtype[jnrB+0];
 663             vdwjidx0C        = 2*vdwtype[jnrC+0];
 664             vdwjidx0D        = 2*vdwtype[jnrD+0];
 665
 666             /**************************
 667              * CALCULATE INTERACTIONS *
 668              **************************/
 669
 670             r00              = _mm_mul_ps(rsq00,rinv00);
 671             r00              = _mm_andnot_ps(dummy_mask,r00);
 672
 673             /* Compute parameters for interactions between i and j atoms */
 674             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 675                                          vdwparam+vdwioffset0+vdwjidx0B,
 676                                          vdwparam+vdwioffset0+vdwjidx0C,
 677                                          vdwparam+vdwioffset0+vdwjidx0D,
 678                                          &c6_00,&c12_00);
 679
 680             /* Calculate table index by multiplying r with table scale and truncate to integer */
 681             rt               = _mm_mul_ps(r00,vftabscale);
 682             vfitab           = _mm_cvttps_epi32(rt);
 683 #ifdef __XOP__
 684             vfeps            = _mm_frcz_ps(rt);
 685 #else
 686             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 687 #endif
 688             twovfeps         = _mm_add_ps(vfeps,vfeps);
 689             vfitab           = _mm_slli_epi32(vfitab,3);
 690
 691             /* CUBIC SPLINE TABLE DISPERSION */
 692             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 693             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 694             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 695             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 696             _MM_TRANSPOSE4_PS(Y,F,G,H);
 697             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 698             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 699             fvdw6            = _mm_mul_ps(c6_00,FF);
 700
 701             /* CUBIC SPLINE TABLE REPULSION */
 702             vfitab           = _mm_add_epi32(vfitab,ifour);
 703             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 704             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 705             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 706             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 707             _MM_TRANSPOSE4_PS(Y,F,G,H);
 708             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 709             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 710             fvdw12           = _mm_mul_ps(c12_00,FF);
 711             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 712
 713             fscal            = fvdw;
 714
 715             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 716
 717              /* Update vectorial force */
 718             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 719             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 720             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 721
 722             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 723             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 724             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 725             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 726             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 727                                                    _mm_mul_ps(dx00,fscal),
 728                                                    _mm_mul_ps(dy00,fscal),
 729                                                    _mm_mul_ps(dz00,fscal));
 730
 731             /* Inner loop uses 52 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*52);
 751 }