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 "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_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: None
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwCSTab_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     __m128i          vfitab;
  95     __m128i          ifour       = _mm_set1_epi32(4);
  96     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  97     real             *vftab;
  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     vftab            = kernel_data->table_vdw->data;
 118     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 119
 120     /* Avoid stupid compiler warnings */
 121     jnrA = jnrB = jnrC = jnrD = 0;
 122     j_coord_offsetA = 0;
 123     j_coord_offsetB = 0;
 124     j_coord_offsetC = 0;
 125     j_coord_offsetD = 0;
 126
 127     outeriter        = 0;
 128     inneriter        = 0;
 129
 130     for(iidx=0;iidx<4*DIM;iidx++)
 131     {
 132         scratch[iidx] = 0.0;
 133     }
 134
 135     /* Start outer loop over neighborlists */
 136     for(iidx=0; iidx<nri; iidx++)
 137     {
 138         /* Load shift vector for this list */
 139         i_shift_offset   = DIM*shiftidx[iidx];
 140
 141         /* Load limits for loop over neighbors */
 142         j_index_start    = jindex[iidx];
 143         j_index_end      = jindex[iidx+1];
 144
 145         /* Get outer coordinate index */
 146         inr              = iinr[iidx];
 147         i_coord_offset   = DIM*inr;
 148
 149         /* Load i particle coords and add shift vector */
 150         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 151
 152         fix0             = _mm_setzero_ps();
 153         fiy0             = _mm_setzero_ps();
 154         fiz0             = _mm_setzero_ps();
 155
 156         /* Load parameters for i particles */
 157         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 158
 159         /* Reset potential sums */
 160         vvdwsum          = _mm_setzero_ps();
 161
 162         /* Start inner kernel loop */
 163         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 164         {
 165
 166             /* Get j neighbor index, and coordinate index */
 167             jnrA             = jjnr[jidx];
 168             jnrB             = jjnr[jidx+1];
 169             jnrC             = jjnr[jidx+2];
 170             jnrD             = jjnr[jidx+3];
 171             j_coord_offsetA  = DIM*jnrA;
 172             j_coord_offsetB  = DIM*jnrB;
 173             j_coord_offsetC  = DIM*jnrC;
 174             j_coord_offsetD  = DIM*jnrD;
 175
 176             /* load j atom coordinates */
 177             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 178                                               x+j_coord_offsetC,x+j_coord_offsetD,
 179                                               &jx0,&jy0,&jz0);
 180
 181             /* Calculate displacement vector */
 182             dx00             = _mm_sub_ps(ix0,jx0);
 183             dy00             = _mm_sub_ps(iy0,jy0);
 184             dz00             = _mm_sub_ps(iz0,jz0);
 185
 186             /* Calculate squared distance and things based on it */
 187             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 188
 189             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 190
 191             /* Load parameters for j particles */
 192             vdwjidx0A        = 2*vdwtype[jnrA+0];
 193             vdwjidx0B        = 2*vdwtype[jnrB+0];
 194             vdwjidx0C        = 2*vdwtype[jnrC+0];
 195             vdwjidx0D        = 2*vdwtype[jnrD+0];
 196
 197             /**************************
 198              * CALCULATE INTERACTIONS *
 199              **************************/
 200
 201             r00              = _mm_mul_ps(rsq00,rinv00);
 202
 203             /* Compute parameters for interactions between i and j atoms */
 204             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 205                                          vdwparam+vdwioffset0+vdwjidx0B,
 206                                          vdwparam+vdwioffset0+vdwjidx0C,
 207                                          vdwparam+vdwioffset0+vdwjidx0D,
 208                                          &c6_00,&c12_00);
 209
 210             /* Calculate table index by multiplying r with table scale and truncate to integer */
 211             rt               = _mm_mul_ps(r00,vftabscale);
 212             vfitab           = _mm_cvttps_epi32(rt);
 213 #ifdef __XOP__
 214             vfeps            = _mm_frcz_ps(rt);
 215 #else
 216             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 217 #endif
 218             twovfeps         = _mm_add_ps(vfeps,vfeps);
 219             vfitab           = _mm_slli_epi32(vfitab,3);
 220
 221             /* CUBIC SPLINE TABLE DISPERSION */
 222             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 223             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 224             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 225             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 226             _MM_TRANSPOSE4_PS(Y,F,G,H);
 227             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 228             VV               = _mm_macc_ps(vfeps,Fp,Y);
 229             vvdw6            = _mm_mul_ps(c6_00,VV);
 230             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 231             fvdw6            = _mm_mul_ps(c6_00,FF);
 232
 233             /* CUBIC SPLINE TABLE REPULSION */
 234             vfitab           = _mm_add_epi32(vfitab,ifour);
 235             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 236             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 237             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 238             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 239             _MM_TRANSPOSE4_PS(Y,F,G,H);
 240             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 241             VV               = _mm_macc_ps(vfeps,Fp,Y);
 242             vvdw12           = _mm_mul_ps(c12_00,VV);
 243             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 244             fvdw12           = _mm_mul_ps(c12_00,FF);
 245             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 246             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 250
 251             fscal            = fvdw;
 252
 253              /* Update vectorial force */
 254             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 255             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 256             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 257
 258             fjptrA             = f+j_coord_offsetA;
 259             fjptrB             = f+j_coord_offsetB;
 260             fjptrC             = f+j_coord_offsetC;
 261             fjptrD             = f+j_coord_offsetD;
 262             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 263                                                    _mm_mul_ps(dx00,fscal),
 264                                                    _mm_mul_ps(dy00,fscal),
 265                                                    _mm_mul_ps(dz00,fscal));
 266
 267             /* Inner loop uses 59 flops */
 268         }
 269
 270         if(jidx<j_index_end)
 271         {
 272
 273             /* Get j neighbor index, and coordinate index */
 274             jnrlistA         = jjnr[jidx];
 275             jnrlistB         = jjnr[jidx+1];
 276             jnrlistC         = jjnr[jidx+2];
 277             jnrlistD         = jjnr[jidx+3];
 278             /* Sign of each element will be negative for non-real atoms.
 279              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 280              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 281              */
 282             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 283             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 284             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 285             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 286             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 287             j_coord_offsetA  = DIM*jnrA;
 288             j_coord_offsetB  = DIM*jnrB;
 289             j_coord_offsetC  = DIM*jnrC;
 290             j_coord_offsetD  = DIM*jnrD;
 291
 292             /* load j atom coordinates */
 293             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 294                                               x+j_coord_offsetC,x+j_coord_offsetD,
 295                                               &jx0,&jy0,&jz0);
 296
 297             /* Calculate displacement vector */
 298             dx00             = _mm_sub_ps(ix0,jx0);
 299             dy00             = _mm_sub_ps(iy0,jy0);
 300             dz00             = _mm_sub_ps(iz0,jz0);
 301
 302             /* Calculate squared distance and things based on it */
 303             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 304
 305             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 306
 307             /* Load parameters for j particles */
 308             vdwjidx0A        = 2*vdwtype[jnrA+0];
 309             vdwjidx0B        = 2*vdwtype[jnrB+0];
 310             vdwjidx0C        = 2*vdwtype[jnrC+0];
 311             vdwjidx0D        = 2*vdwtype[jnrD+0];
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             r00              = _mm_mul_ps(rsq00,rinv00);
 318             r00              = _mm_andnot_ps(dummy_mask,r00);
 319
 320             /* Compute parameters for interactions between i and j atoms */
 321             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 322                                          vdwparam+vdwioffset0+vdwjidx0B,
 323                                          vdwparam+vdwioffset0+vdwjidx0C,
 324                                          vdwparam+vdwioffset0+vdwjidx0D,
 325                                          &c6_00,&c12_00);
 326
 327             /* Calculate table index by multiplying r with table scale and truncate to integer */
 328             rt               = _mm_mul_ps(r00,vftabscale);
 329             vfitab           = _mm_cvttps_epi32(rt);
 330 #ifdef __XOP__
 331             vfeps            = _mm_frcz_ps(rt);
 332 #else
 333             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 334 #endif
 335             twovfeps         = _mm_add_ps(vfeps,vfeps);
 336             vfitab           = _mm_slli_epi32(vfitab,3);
 337
 338             /* CUBIC SPLINE TABLE DISPERSION */
 339             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 340             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 341             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 342             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 343             _MM_TRANSPOSE4_PS(Y,F,G,H);
 344             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 345             VV               = _mm_macc_ps(vfeps,Fp,Y);
 346             vvdw6            = _mm_mul_ps(c6_00,VV);
 347             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 348             fvdw6            = _mm_mul_ps(c6_00,FF);
 349
 350             /* CUBIC SPLINE TABLE REPULSION */
 351             vfitab           = _mm_add_epi32(vfitab,ifour);
 352             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 353             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 354             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 355             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 356             _MM_TRANSPOSE4_PS(Y,F,G,H);
 357             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 358             VV               = _mm_macc_ps(vfeps,Fp,Y);
 359             vvdw12           = _mm_mul_ps(c12_00,VV);
 360             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 361             fvdw12           = _mm_mul_ps(c12_00,FF);
 362             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 363             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 364
 365             /* Update potential sum for this i atom from the interaction with this j atom. */
 366             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 367             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 368
 369             fscal            = fvdw;
 370
 371             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 372
 373              /* Update vectorial force */
 374             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 375             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 376             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 377
 378             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 379             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 380             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 381             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 382             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 383                                                    _mm_mul_ps(dx00,fscal),
 384                                                    _mm_mul_ps(dy00,fscal),
 385                                                    _mm_mul_ps(dz00,fscal));
 386
 387             /* Inner loop uses 60 flops */
 388         }
 389
 390         /* End of innermost loop */
 391
 392         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 393                                               f+i_coord_offset,fshift+i_shift_offset);
 394
 395         ggid                        = gid[iidx];
 396         /* Update potential energies */
 397         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 398
 399         /* Increment number of inner iterations */
 400         inneriter                  += j_index_end - j_index_start;
 401
 402         /* Outer loop uses 7 flops */
 403     }
 404
 405     /* Increment number of outer iterations */
 406     outeriter        += nri;
 407
 408     /* Update outer/inner flops */
 409
 410     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 411 }
 412 /*
 413  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 414  * Electrostatics interaction: None
 415  * VdW interaction:            CubicSplineTable
 416  * Geometry:                   Particle-Particle
 417  * Calculate force/pot:        Force
 418  */
 419 void
 420 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 421                     (t_nblist                    * gmx_restrict       nlist,
 422                      rvec                        * gmx_restrict          xx,
 423                      rvec                        * gmx_restrict          ff,
 424                      t_forcerec                  * gmx_restrict          fr,
 425                      t_mdatoms                   * gmx_restrict     mdatoms,
 426                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 427                      t_nrnb                      * gmx_restrict        nrnb)
 428 {
 429     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 430      * just 0 for non-waters.
 431      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 432      * jnr indices corresponding to data put in the four positions in the SIMD register.
 433      */
 434     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 435     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 436     int              jnrA,jnrB,jnrC,jnrD;
 437     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 438     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 439     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 440     real             rcutoff_scalar;
 441     real             *shiftvec,*fshift,*x,*f;
 442     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 443     real             scratch[4*DIM];
 444     __m128           fscal,rcutoff,rcutoff2,jidxall;
 445     int              vdwioffset0;
 446     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 447     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 448     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 449     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 450     int              nvdwtype;
 451     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 452     int              *vdwtype;
 453     real             *vdwparam;
 454     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 455     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 456     __m128i          vfitab;
 457     __m128i          ifour       = _mm_set1_epi32(4);
 458     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 459     real             *vftab;
 460     __m128           dummy_mask,cutoff_mask;
 461     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 462     __m128           one     = _mm_set1_ps(1.0);
 463     __m128           two     = _mm_set1_ps(2.0);
 464     x                = xx[0];
 465     f                = ff[0];
 466
 467     nri              = nlist->nri;
 468     iinr             = nlist->iinr;
 469     jindex           = nlist->jindex;
 470     jjnr             = nlist->jjnr;
 471     shiftidx         = nlist->shift;
 472     gid              = nlist->gid;
 473     shiftvec         = fr->shift_vec[0];
 474     fshift           = fr->fshift[0];
 475     nvdwtype         = fr->ntype;
 476     vdwparam         = fr->nbfp;
 477     vdwtype          = mdatoms->typeA;
 478
 479     vftab            = kernel_data->table_vdw->data;
 480     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 481
 482     /* Avoid stupid compiler warnings */
 483     jnrA = jnrB = jnrC = jnrD = 0;
 484     j_coord_offsetA = 0;
 485     j_coord_offsetB = 0;
 486     j_coord_offsetC = 0;
 487     j_coord_offsetD = 0;
 488
 489     outeriter        = 0;
 490     inneriter        = 0;
 491
 492     for(iidx=0;iidx<4*DIM;iidx++)
 493     {
 494         scratch[iidx] = 0.0;
 495     }
 496
 497     /* Start outer loop over neighborlists */
 498     for(iidx=0; iidx<nri; iidx++)
 499     {
 500         /* Load shift vector for this list */
 501         i_shift_offset   = DIM*shiftidx[iidx];
 502
 503         /* Load limits for loop over neighbors */
 504         j_index_start    = jindex[iidx];
 505         j_index_end      = jindex[iidx+1];
 506
 507         /* Get outer coordinate index */
 508         inr              = iinr[iidx];
 509         i_coord_offset   = DIM*inr;
 510
 511         /* Load i particle coords and add shift vector */
 512         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 513
 514         fix0             = _mm_setzero_ps();
 515         fiy0             = _mm_setzero_ps();
 516         fiz0             = _mm_setzero_ps();
 517
 518         /* Load parameters for i particles */
 519         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 520
 521         /* Start inner kernel loop */
 522         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 523         {
 524
 525             /* Get j neighbor index, and coordinate index */
 526             jnrA             = jjnr[jidx];
 527             jnrB             = jjnr[jidx+1];
 528             jnrC             = jjnr[jidx+2];
 529             jnrD             = jjnr[jidx+3];
 530             j_coord_offsetA  = DIM*jnrA;
 531             j_coord_offsetB  = DIM*jnrB;
 532             j_coord_offsetC  = DIM*jnrC;
 533             j_coord_offsetD  = DIM*jnrD;
 534
 535             /* load j atom coordinates */
 536             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 537                                               x+j_coord_offsetC,x+j_coord_offsetD,
 538                                               &jx0,&jy0,&jz0);
 539
 540             /* Calculate displacement vector */
 541             dx00             = _mm_sub_ps(ix0,jx0);
 542             dy00             = _mm_sub_ps(iy0,jy0);
 543             dz00             = _mm_sub_ps(iz0,jz0);
 544
 545             /* Calculate squared distance and things based on it */
 546             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 547
 548             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 549
 550             /* Load parameters for j particles */
 551             vdwjidx0A        = 2*vdwtype[jnrA+0];
 552             vdwjidx0B        = 2*vdwtype[jnrB+0];
 553             vdwjidx0C        = 2*vdwtype[jnrC+0];
 554             vdwjidx0D        = 2*vdwtype[jnrD+0];
 555
 556             /**************************
 557              * CALCULATE INTERACTIONS *
 558              **************************/
 559
 560             r00              = _mm_mul_ps(rsq00,rinv00);
 561
 562             /* Compute parameters for interactions between i and j atoms */
 563             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 564                                          vdwparam+vdwioffset0+vdwjidx0B,
 565                                          vdwparam+vdwioffset0+vdwjidx0C,
 566                                          vdwparam+vdwioffset0+vdwjidx0D,
 567                                          &c6_00,&c12_00);
 568
 569             /* Calculate table index by multiplying r with table scale and truncate to integer */
 570             rt               = _mm_mul_ps(r00,vftabscale);
 571             vfitab           = _mm_cvttps_epi32(rt);
 572 #ifdef __XOP__
 573             vfeps            = _mm_frcz_ps(rt);
 574 #else
 575             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 576 #endif
 577             twovfeps         = _mm_add_ps(vfeps,vfeps);
 578             vfitab           = _mm_slli_epi32(vfitab,3);
 579
 580             /* CUBIC SPLINE TABLE DISPERSION */
 581             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 582             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 583             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 584             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 585             _MM_TRANSPOSE4_PS(Y,F,G,H);
 586             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 587             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 588             fvdw6            = _mm_mul_ps(c6_00,FF);
 589
 590             /* CUBIC SPLINE TABLE REPULSION */
 591             vfitab           = _mm_add_epi32(vfitab,ifour);
 592             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 593             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 594             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 595             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 596             _MM_TRANSPOSE4_PS(Y,F,G,H);
 597             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 598             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 599             fvdw12           = _mm_mul_ps(c12_00,FF);
 600             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 601
 602             fscal            = fvdw;
 603
 604              /* Update vectorial force */
 605             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 606             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 607             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 608
 609             fjptrA             = f+j_coord_offsetA;
 610             fjptrB             = f+j_coord_offsetB;
 611             fjptrC             = f+j_coord_offsetC;
 612             fjptrD             = f+j_coord_offsetD;
 613             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 614                                                    _mm_mul_ps(dx00,fscal),
 615                                                    _mm_mul_ps(dy00,fscal),
 616                                                    _mm_mul_ps(dz00,fscal));
 617
 618             /* Inner loop uses 51 flops */
 619         }
 620
 621         if(jidx<j_index_end)
 622         {
 623
 624             /* Get j neighbor index, and coordinate index */
 625             jnrlistA         = jjnr[jidx];
 626             jnrlistB         = jjnr[jidx+1];
 627             jnrlistC         = jjnr[jidx+2];
 628             jnrlistD         = jjnr[jidx+3];
 629             /* Sign of each element will be negative for non-real atoms.
 630              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 631              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 632              */
 633             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 634             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 635             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 636             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 637             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 638             j_coord_offsetA  = DIM*jnrA;
 639             j_coord_offsetB  = DIM*jnrB;
 640             j_coord_offsetC  = DIM*jnrC;
 641             j_coord_offsetD  = DIM*jnrD;
 642
 643             /* load j atom coordinates */
 644             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 645                                               x+j_coord_offsetC,x+j_coord_offsetD,
 646                                               &jx0,&jy0,&jz0);
 647
 648             /* Calculate displacement vector */
 649             dx00             = _mm_sub_ps(ix0,jx0);
 650             dy00             = _mm_sub_ps(iy0,jy0);
 651             dz00             = _mm_sub_ps(iz0,jz0);
 652
 653             /* Calculate squared distance and things based on it */
 654             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 655
 656             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 657
 658             /* Load parameters for j particles */
 659             vdwjidx0A        = 2*vdwtype[jnrA+0];
 660             vdwjidx0B        = 2*vdwtype[jnrB+0];
 661             vdwjidx0C        = 2*vdwtype[jnrC+0];
 662             vdwjidx0D        = 2*vdwtype[jnrD+0];
 663
 664             /**************************
 665              * CALCULATE INTERACTIONS *
 666              **************************/
 667
 668             r00              = _mm_mul_ps(rsq00,rinv00);
 669             r00              = _mm_andnot_ps(dummy_mask,r00);
 670
 671             /* Compute parameters for interactions between i and j atoms */
 672             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 673                                          vdwparam+vdwioffset0+vdwjidx0B,
 674                                          vdwparam+vdwioffset0+vdwjidx0C,
 675                                          vdwparam+vdwioffset0+vdwjidx0D,
 676                                          &c6_00,&c12_00);
 677
 678             /* Calculate table index by multiplying r with table scale and truncate to integer */
 679             rt               = _mm_mul_ps(r00,vftabscale);
 680             vfitab           = _mm_cvttps_epi32(rt);
 681 #ifdef __XOP__
 682             vfeps            = _mm_frcz_ps(rt);
 683 #else
 684             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 685 #endif
 686             twovfeps         = _mm_add_ps(vfeps,vfeps);
 687             vfitab           = _mm_slli_epi32(vfitab,3);
 688
 689             /* CUBIC SPLINE TABLE DISPERSION */
 690             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 691             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 692             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 693             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 694             _MM_TRANSPOSE4_PS(Y,F,G,H);
 695             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 696             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 697             fvdw6            = _mm_mul_ps(c6_00,FF);
 698
 699             /* CUBIC SPLINE TABLE REPULSION */
 700             vfitab           = _mm_add_epi32(vfitab,ifour);
 701             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 702             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 703             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 704             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 705             _MM_TRANSPOSE4_PS(Y,F,G,H);
 706             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 707             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 708             fvdw12           = _mm_mul_ps(c12_00,FF);
 709             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 710
 711             fscal            = fvdw;
 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             /* Inner loop uses 52 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*52);
 749 }