src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwCSTab_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: None
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     int              nvdwtype;
  75     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  76     int              *vdwtype;
  77     real             *vdwparam;
  78     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  79     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  80     __m128i          vfitab;
  81     __m128i          ifour       = _mm_set1_epi32(4);
  82     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  83     real             *vftab;
  84     __m128           dummy_mask,cutoff_mask;
  85     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  86     __m128           one     = _mm_set1_ps(1.0);
  87     __m128           two     = _mm_set1_ps(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     nvdwtype         = fr->ntype;
 100     vdwparam         = fr->nbfp;
 101     vdwtype          = mdatoms->typeA;
 102
 103     vftab            = kernel_data->table_vdw->data;
 104     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 105
 106     /* Avoid stupid compiler warnings */
 107     jnrA = jnrB = jnrC = jnrD = 0;
 108     j_coord_offsetA = 0;
 109     j_coord_offsetB = 0;
 110     j_coord_offsetC = 0;
 111     j_coord_offsetD = 0;
 112
 113     outeriter        = 0;
 114     inneriter        = 0;
 115
 116     for(iidx=0;iidx<4*DIM;iidx++)
 117     {
 118         scratch[iidx] = 0.0;
 119     }
 120
 121     /* Start outer loop over neighborlists */
 122     for(iidx=0; iidx<nri; iidx++)
 123     {
 124         /* Load shift vector for this list */
 125         i_shift_offset   = DIM*shiftidx[iidx];
 126
 127         /* Load limits for loop over neighbors */
 128         j_index_start    = jindex[iidx];
 129         j_index_end      = jindex[iidx+1];
 130
 131         /* Get outer coordinate index */
 132         inr              = iinr[iidx];
 133         i_coord_offset   = DIM*inr;
 134
 135         /* Load i particle coords and add shift vector */
 136         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 137
 138         fix0             = _mm_setzero_ps();
 139         fiy0             = _mm_setzero_ps();
 140         fiz0             = _mm_setzero_ps();
 141
 142         /* Load parameters for i particles */
 143         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145         /* Reset potential sums */
 146         vvdwsum          = _mm_setzero_ps();
 147
 148         /* Start inner kernel loop */
 149         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 150         {
 151
 152             /* Get j neighbor index, and coordinate index */
 153             jnrA             = jjnr[jidx];
 154             jnrB             = jjnr[jidx+1];
 155             jnrC             = jjnr[jidx+2];
 156             jnrD             = jjnr[jidx+3];
 157             j_coord_offsetA  = DIM*jnrA;
 158             j_coord_offsetB  = DIM*jnrB;
 159             j_coord_offsetC  = DIM*jnrC;
 160             j_coord_offsetD  = DIM*jnrD;
 161
 162             /* load j atom coordinates */
 163             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 164                                               x+j_coord_offsetC,x+j_coord_offsetD,
 165                                               &jx0,&jy0,&jz0);
 166
 167             /* Calculate displacement vector */
 168             dx00             = _mm_sub_ps(ix0,jx0);
 169             dy00             = _mm_sub_ps(iy0,jy0);
 170             dz00             = _mm_sub_ps(iz0,jz0);
 171
 172             /* Calculate squared distance and things based on it */
 173             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 174
 175             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 176
 177             /* Load parameters for j particles */
 178             vdwjidx0A        = 2*vdwtype[jnrA+0];
 179             vdwjidx0B        = 2*vdwtype[jnrB+0];
 180             vdwjidx0C        = 2*vdwtype[jnrC+0];
 181             vdwjidx0D        = 2*vdwtype[jnrD+0];
 182
 183             /**************************
 184              * CALCULATE INTERACTIONS *
 185              **************************/
 186
 187             r00              = _mm_mul_ps(rsq00,rinv00);
 188
 189             /* Compute parameters for interactions between i and j atoms */
 190             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 191                                          vdwparam+vdwioffset0+vdwjidx0B,
 192                                          vdwparam+vdwioffset0+vdwjidx0C,
 193                                          vdwparam+vdwioffset0+vdwjidx0D,
 194                                          &c6_00,&c12_00);
 195
 196             /* Calculate table index by multiplying r with table scale and truncate to integer */
 197             rt               = _mm_mul_ps(r00,vftabscale);
 198             vfitab           = _mm_cvttps_epi32(rt);
 199 #ifdef __XOP__
 200             vfeps            = _mm_frcz_ps(rt);
 201 #else
 202             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 203 #endif
 204             twovfeps         = _mm_add_ps(vfeps,vfeps);
 205             vfitab           = _mm_slli_epi32(vfitab,3);
 206
 207             /* CUBIC SPLINE TABLE DISPERSION */
 208             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 209             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 210             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 211             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 212             _MM_TRANSPOSE4_PS(Y,F,G,H);
 213             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 214             VV               = _mm_macc_ps(vfeps,Fp,Y);
 215             vvdw6            = _mm_mul_ps(c6_00,VV);
 216             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 217             fvdw6            = _mm_mul_ps(c6_00,FF);
 218
 219             /* CUBIC SPLINE TABLE REPULSION */
 220             vfitab           = _mm_add_epi32(vfitab,ifour);
 221             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 222             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 223             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 224             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 225             _MM_TRANSPOSE4_PS(Y,F,G,H);
 226             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 227             VV               = _mm_macc_ps(vfeps,Fp,Y);
 228             vvdw12           = _mm_mul_ps(c12_00,VV);
 229             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 230             fvdw12           = _mm_mul_ps(c12_00,FF);
 231             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 232             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 233
 234             /* Update potential sum for this i atom from the interaction with this j atom. */
 235             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 236
 237             fscal            = fvdw;
 238
 239              /* Update vectorial force */
 240             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 241             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 242             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 243
 244             fjptrA             = f+j_coord_offsetA;
 245             fjptrB             = f+j_coord_offsetB;
 246             fjptrC             = f+j_coord_offsetC;
 247             fjptrD             = f+j_coord_offsetD;
 248             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 249                                                    _mm_mul_ps(dx00,fscal),
 250                                                    _mm_mul_ps(dy00,fscal),
 251                                                    _mm_mul_ps(dz00,fscal));
 252
 253             /* Inner loop uses 59 flops */
 254         }
 255
 256         if(jidx<j_index_end)
 257         {
 258
 259             /* Get j neighbor index, and coordinate index */
 260             jnrlistA         = jjnr[jidx];
 261             jnrlistB         = jjnr[jidx+1];
 262             jnrlistC         = jjnr[jidx+2];
 263             jnrlistD         = jjnr[jidx+3];
 264             /* Sign of each element will be negative for non-real atoms.
 265              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 266              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 267              */
 268             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 269             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 270             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 271             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 272             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 273             j_coord_offsetA  = DIM*jnrA;
 274             j_coord_offsetB  = DIM*jnrB;
 275             j_coord_offsetC  = DIM*jnrC;
 276             j_coord_offsetD  = DIM*jnrD;
 277
 278             /* load j atom coordinates */
 279             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 280                                               x+j_coord_offsetC,x+j_coord_offsetD,
 281                                               &jx0,&jy0,&jz0);
 282
 283             /* Calculate displacement vector */
 284             dx00             = _mm_sub_ps(ix0,jx0);
 285             dy00             = _mm_sub_ps(iy0,jy0);
 286             dz00             = _mm_sub_ps(iz0,jz0);
 287
 288             /* Calculate squared distance and things based on it */
 289             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 290
 291             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 292
 293             /* Load parameters for j particles */
 294             vdwjidx0A        = 2*vdwtype[jnrA+0];
 295             vdwjidx0B        = 2*vdwtype[jnrB+0];
 296             vdwjidx0C        = 2*vdwtype[jnrC+0];
 297             vdwjidx0D        = 2*vdwtype[jnrD+0];
 298
 299             /**************************
 300              * CALCULATE INTERACTIONS *
 301              **************************/
 302
 303             r00              = _mm_mul_ps(rsq00,rinv00);
 304             r00              = _mm_andnot_ps(dummy_mask,r00);
 305
 306             /* Compute parameters for interactions between i and j atoms */
 307             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 308                                          vdwparam+vdwioffset0+vdwjidx0B,
 309                                          vdwparam+vdwioffset0+vdwjidx0C,
 310                                          vdwparam+vdwioffset0+vdwjidx0D,
 311                                          &c6_00,&c12_00);
 312
 313             /* Calculate table index by multiplying r with table scale and truncate to integer */
 314             rt               = _mm_mul_ps(r00,vftabscale);
 315             vfitab           = _mm_cvttps_epi32(rt);
 316 #ifdef __XOP__
 317             vfeps            = _mm_frcz_ps(rt);
 318 #else
 319             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 320 #endif
 321             twovfeps         = _mm_add_ps(vfeps,vfeps);
 322             vfitab           = _mm_slli_epi32(vfitab,3);
 323
 324             /* CUBIC SPLINE TABLE DISPERSION */
 325             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 326             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 327             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 328             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 329             _MM_TRANSPOSE4_PS(Y,F,G,H);
 330             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 331             VV               = _mm_macc_ps(vfeps,Fp,Y);
 332             vvdw6            = _mm_mul_ps(c6_00,VV);
 333             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 334             fvdw6            = _mm_mul_ps(c6_00,FF);
 335
 336             /* CUBIC SPLINE TABLE REPULSION */
 337             vfitab           = _mm_add_epi32(vfitab,ifour);
 338             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 339             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 340             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 341             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 342             _MM_TRANSPOSE4_PS(Y,F,G,H);
 343             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 344             VV               = _mm_macc_ps(vfeps,Fp,Y);
 345             vvdw12           = _mm_mul_ps(c12_00,VV);
 346             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 347             fvdw12           = _mm_mul_ps(c12_00,FF);
 348             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 349             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 350
 351             /* Update potential sum for this i atom from the interaction with this j atom. */
 352             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 353             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 354
 355             fscal            = fvdw;
 356
 357             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 358
 359              /* Update vectorial force */
 360             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 361             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 362             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 363
 364             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 365             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 366             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 367             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 368             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 369                                                    _mm_mul_ps(dx00,fscal),
 370                                                    _mm_mul_ps(dy00,fscal),
 371                                                    _mm_mul_ps(dz00,fscal));
 372
 373             /* Inner loop uses 60 flops */
 374         }
 375
 376         /* End of innermost loop */
 377
 378         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 379                                               f+i_coord_offset,fshift+i_shift_offset);
 380
 381         ggid                        = gid[iidx];
 382         /* Update potential energies */
 383         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 384
 385         /* Increment number of inner iterations */
 386         inneriter                  += j_index_end - j_index_start;
 387
 388         /* Outer loop uses 7 flops */
 389     }
 390
 391     /* Increment number of outer iterations */
 392     outeriter        += nri;
 393
 394     /* Update outer/inner flops */
 395
 396     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 397 }
 398 /*
 399  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 400  * Electrostatics interaction: None
 401  * VdW interaction:            CubicSplineTable
 402  * Geometry:                   Particle-Particle
 403  * Calculate force/pot:        Force
 404  */
 405 void
 406 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 407                     (t_nblist * gmx_restrict                nlist,
 408                      rvec * gmx_restrict                    xx,
 409                      rvec * gmx_restrict                    ff,
 410                      t_forcerec * gmx_restrict              fr,
 411                      t_mdatoms * gmx_restrict               mdatoms,
 412                      nb_kernel_data_t * gmx_restrict        kernel_data,
 413                      t_nrnb * gmx_restrict                  nrnb)
 414 {
 415     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 416      * just 0 for non-waters.
 417      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 418      * jnr indices corresponding to data put in the four positions in the SIMD register.
 419      */
 420     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 421     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 422     int              jnrA,jnrB,jnrC,jnrD;
 423     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 424     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 425     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 426     real             rcutoff_scalar;
 427     real             *shiftvec,*fshift,*x,*f;
 428     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 429     real             scratch[4*DIM];
 430     __m128           fscal,rcutoff,rcutoff2,jidxall;
 431     int              vdwioffset0;
 432     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 433     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 434     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 435     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 436     int              nvdwtype;
 437     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 438     int              *vdwtype;
 439     real             *vdwparam;
 440     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 441     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 442     __m128i          vfitab;
 443     __m128i          ifour       = _mm_set1_epi32(4);
 444     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 445     real             *vftab;
 446     __m128           dummy_mask,cutoff_mask;
 447     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 448     __m128           one     = _mm_set1_ps(1.0);
 449     __m128           two     = _mm_set1_ps(2.0);
 450     x                = xx[0];
 451     f                = ff[0];
 452
 453     nri              = nlist->nri;
 454     iinr             = nlist->iinr;
 455     jindex           = nlist->jindex;
 456     jjnr             = nlist->jjnr;
 457     shiftidx         = nlist->shift;
 458     gid              = nlist->gid;
 459     shiftvec         = fr->shift_vec[0];
 460     fshift           = fr->fshift[0];
 461     nvdwtype         = fr->ntype;
 462     vdwparam         = fr->nbfp;
 463     vdwtype          = mdatoms->typeA;
 464
 465     vftab            = kernel_data->table_vdw->data;
 466     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 467
 468     /* Avoid stupid compiler warnings */
 469     jnrA = jnrB = jnrC = jnrD = 0;
 470     j_coord_offsetA = 0;
 471     j_coord_offsetB = 0;
 472     j_coord_offsetC = 0;
 473     j_coord_offsetD = 0;
 474
 475     outeriter        = 0;
 476     inneriter        = 0;
 477
 478     for(iidx=0;iidx<4*DIM;iidx++)
 479     {
 480         scratch[iidx] = 0.0;
 481     }
 482
 483     /* Start outer loop over neighborlists */
 484     for(iidx=0; iidx<nri; iidx++)
 485     {
 486         /* Load shift vector for this list */
 487         i_shift_offset   = DIM*shiftidx[iidx];
 488
 489         /* Load limits for loop over neighbors */
 490         j_index_start    = jindex[iidx];
 491         j_index_end      = jindex[iidx+1];
 492
 493         /* Get outer coordinate index */
 494         inr              = iinr[iidx];
 495         i_coord_offset   = DIM*inr;
 496
 497         /* Load i particle coords and add shift vector */
 498         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 499
 500         fix0             = _mm_setzero_ps();
 501         fiy0             = _mm_setzero_ps();
 502         fiz0             = _mm_setzero_ps();
 503
 504         /* Load parameters for i particles */
 505         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 506
 507         /* Start inner kernel loop */
 508         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 509         {
 510
 511             /* Get j neighbor index, and coordinate index */
 512             jnrA             = jjnr[jidx];
 513             jnrB             = jjnr[jidx+1];
 514             jnrC             = jjnr[jidx+2];
 515             jnrD             = jjnr[jidx+3];
 516             j_coord_offsetA  = DIM*jnrA;
 517             j_coord_offsetB  = DIM*jnrB;
 518             j_coord_offsetC  = DIM*jnrC;
 519             j_coord_offsetD  = DIM*jnrD;
 520
 521             /* load j atom coordinates */
 522             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 523                                               x+j_coord_offsetC,x+j_coord_offsetD,
 524                                               &jx0,&jy0,&jz0);
 525
 526             /* Calculate displacement vector */
 527             dx00             = _mm_sub_ps(ix0,jx0);
 528             dy00             = _mm_sub_ps(iy0,jy0);
 529             dz00             = _mm_sub_ps(iz0,jz0);
 530
 531             /* Calculate squared distance and things based on it */
 532             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 533
 534             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 535
 536             /* Load parameters for j particles */
 537             vdwjidx0A        = 2*vdwtype[jnrA+0];
 538             vdwjidx0B        = 2*vdwtype[jnrB+0];
 539             vdwjidx0C        = 2*vdwtype[jnrC+0];
 540             vdwjidx0D        = 2*vdwtype[jnrD+0];
 541
 542             /**************************
 543              * CALCULATE INTERACTIONS *
 544              **************************/
 545
 546             r00              = _mm_mul_ps(rsq00,rinv00);
 547
 548             /* Compute parameters for interactions between i and j atoms */
 549             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 550                                          vdwparam+vdwioffset0+vdwjidx0B,
 551                                          vdwparam+vdwioffset0+vdwjidx0C,
 552                                          vdwparam+vdwioffset0+vdwjidx0D,
 553                                          &c6_00,&c12_00);
 554
 555             /* Calculate table index by multiplying r with table scale and truncate to integer */
 556             rt               = _mm_mul_ps(r00,vftabscale);
 557             vfitab           = _mm_cvttps_epi32(rt);
 558 #ifdef __XOP__
 559             vfeps            = _mm_frcz_ps(rt);
 560 #else
 561             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 562 #endif
 563             twovfeps         = _mm_add_ps(vfeps,vfeps);
 564             vfitab           = _mm_slli_epi32(vfitab,3);
 565
 566             /* CUBIC SPLINE TABLE DISPERSION */
 567             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 568             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 569             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 570             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 571             _MM_TRANSPOSE4_PS(Y,F,G,H);
 572             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 573             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 574             fvdw6            = _mm_mul_ps(c6_00,FF);
 575
 576             /* CUBIC SPLINE TABLE REPULSION */
 577             vfitab           = _mm_add_epi32(vfitab,ifour);
 578             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 579             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 580             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 581             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 582             _MM_TRANSPOSE4_PS(Y,F,G,H);
 583             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 584             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 585             fvdw12           = _mm_mul_ps(c12_00,FF);
 586             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 587
 588             fscal            = fvdw;
 589
 590              /* Update vectorial force */
 591             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 592             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 593             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 594
 595             fjptrA             = f+j_coord_offsetA;
 596             fjptrB             = f+j_coord_offsetB;
 597             fjptrC             = f+j_coord_offsetC;
 598             fjptrD             = f+j_coord_offsetD;
 599             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 600                                                    _mm_mul_ps(dx00,fscal),
 601                                                    _mm_mul_ps(dy00,fscal),
 602                                                    _mm_mul_ps(dz00,fscal));
 603
 604             /* Inner loop uses 51 flops */
 605         }
 606
 607         if(jidx<j_index_end)
 608         {
 609
 610             /* Get j neighbor index, and coordinate index */
 611             jnrlistA         = jjnr[jidx];
 612             jnrlistB         = jjnr[jidx+1];
 613             jnrlistC         = jjnr[jidx+2];
 614             jnrlistD         = jjnr[jidx+3];
 615             /* Sign of each element will be negative for non-real atoms.
 616              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 617              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 618              */
 619             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 620             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 621             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 622             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 623             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 624             j_coord_offsetA  = DIM*jnrA;
 625             j_coord_offsetB  = DIM*jnrB;
 626             j_coord_offsetC  = DIM*jnrC;
 627             j_coord_offsetD  = DIM*jnrD;
 628
 629             /* load j atom coordinates */
 630             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 631                                               x+j_coord_offsetC,x+j_coord_offsetD,
 632                                               &jx0,&jy0,&jz0);
 633
 634             /* Calculate displacement vector */
 635             dx00             = _mm_sub_ps(ix0,jx0);
 636             dy00             = _mm_sub_ps(iy0,jy0);
 637             dz00             = _mm_sub_ps(iz0,jz0);
 638
 639             /* Calculate squared distance and things based on it */
 640             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 641
 642             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 643
 644             /* Load parameters for j particles */
 645             vdwjidx0A        = 2*vdwtype[jnrA+0];
 646             vdwjidx0B        = 2*vdwtype[jnrB+0];
 647             vdwjidx0C        = 2*vdwtype[jnrC+0];
 648             vdwjidx0D        = 2*vdwtype[jnrD+0];
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             r00              = _mm_mul_ps(rsq00,rinv00);
 655             r00              = _mm_andnot_ps(dummy_mask,r00);
 656
 657             /* Compute parameters for interactions between i and j atoms */
 658             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 659                                          vdwparam+vdwioffset0+vdwjidx0B,
 660                                          vdwparam+vdwioffset0+vdwjidx0C,
 661                                          vdwparam+vdwioffset0+vdwjidx0D,
 662                                          &c6_00,&c12_00);
 663
 664             /* Calculate table index by multiplying r with table scale and truncate to integer */
 665             rt               = _mm_mul_ps(r00,vftabscale);
 666             vfitab           = _mm_cvttps_epi32(rt);
 667 #ifdef __XOP__
 668             vfeps            = _mm_frcz_ps(rt);
 669 #else
 670             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 671 #endif
 672             twovfeps         = _mm_add_ps(vfeps,vfeps);
 673             vfitab           = _mm_slli_epi32(vfitab,3);
 674
 675             /* CUBIC SPLINE TABLE DISPERSION */
 676             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 677             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 678             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 679             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 680             _MM_TRANSPOSE4_PS(Y,F,G,H);
 681             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 682             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 683             fvdw6            = _mm_mul_ps(c6_00,FF);
 684
 685             /* CUBIC SPLINE TABLE REPULSION */
 686             vfitab           = _mm_add_epi32(vfitab,ifour);
 687             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 688             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 689             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 690             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 691             _MM_TRANSPOSE4_PS(Y,F,G,H);
 692             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 693             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 694             fvdw12           = _mm_mul_ps(c12_00,FF);
 695             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 696
 697             fscal            = fvdw;
 698
 699             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 700
 701              /* Update vectorial force */
 702             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 703             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 704             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 705
 706             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 707             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 708             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 709             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 710             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 711                                                    _mm_mul_ps(dx00,fscal),
 712                                                    _mm_mul_ps(dy00,fscal),
 713                                                    _mm_mul_ps(dz00,fscal));
 714
 715             /* Inner loop uses 52 flops */
 716         }
 717
 718         /* End of innermost loop */
 719
 720         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 721                                               f+i_coord_offset,fshift+i_shift_offset);
 722
 723         /* Increment number of inner iterations */
 724         inneriter                  += j_index_end - j_index_start;
 725
 726         /* Outer loop uses 6 flops */
 727     }
 728
 729     /* Increment number of outer iterations */
 730     outeriter        += nri;
 731
 732     /* Update outer/inner flops */
 733
 734     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*52);
 735 }