src/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecNone_VdwCSTab_GeomP1P1_avx_256_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_double 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_256_double.h"
  34 #include "kernelutil_x86_avx_256_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_double
  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_256_double
  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, 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              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  63     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  64     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  65     real             rcutoff_scalar;
  66     real             *shiftvec,*fshift,*x,*f;
  67     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  68     real             scratch[4*DIM];
  69     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  70     real *           vdwioffsetptr0;
  71     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     int              nvdwtype;
  76     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  77     int              *vdwtype;
  78     real             *vdwparam;
  79     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  80     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  81     __m128i          vfitab;
  82     __m128i          ifour       = _mm_set1_epi32(4);
  83     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  84     real             *vftab;
  85     __m256d          dummy_mask,cutoff_mask;
  86     __m128           tmpmask0,tmpmask1;
  87     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  88     __m256d          one     = _mm256_set1_pd(1.0);
  89     __m256d          two     = _mm256_set1_pd(2.0);
  90     x                = xx[0];
  91     f                = ff[0];
  92
  93     nri              = nlist->nri;
  94     iinr             = nlist->iinr;
  95     jindex           = nlist->jindex;
  96     jjnr             = nlist->jjnr;
  97     shiftidx         = nlist->shift;
  98     gid              = nlist->gid;
  99     shiftvec         = fr->shift_vec[0];
 100     fshift           = fr->fshift[0];
 101     nvdwtype         = fr->ntype;
 102     vdwparam         = fr->nbfp;
 103     vdwtype          = mdatoms->typeA;
 104
 105     vftab            = kernel_data->table_vdw->data;
 106     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 107
 108     /* Avoid stupid compiler warnings */
 109     jnrA = jnrB = jnrC = jnrD = 0;
 110     j_coord_offsetA = 0;
 111     j_coord_offsetB = 0;
 112     j_coord_offsetC = 0;
 113     j_coord_offsetD = 0;
 114
 115     outeriter        = 0;
 116     inneriter        = 0;
 117
 118     for(iidx=0;iidx<4*DIM;iidx++)
 119     {
 120         scratch[iidx] = 0.0;
 121     }
 122
 123     /* Start outer loop over neighborlists */
 124     for(iidx=0; iidx<nri; iidx++)
 125     {
 126         /* Load shift vector for this list */
 127         i_shift_offset   = DIM*shiftidx[iidx];
 128
 129         /* Load limits for loop over neighbors */
 130         j_index_start    = jindex[iidx];
 131         j_index_end      = jindex[iidx+1];
 132
 133         /* Get outer coordinate index */
 134         inr              = iinr[iidx];
 135         i_coord_offset   = DIM*inr;
 136
 137         /* Load i particle coords and add shift vector */
 138         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 139
 140         fix0             = _mm256_setzero_pd();
 141         fiy0             = _mm256_setzero_pd();
 142         fiz0             = _mm256_setzero_pd();
 143
 144         /* Load parameters for i particles */
 145         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 146
 147         /* Reset potential sums */
 148         vvdwsum          = _mm256_setzero_pd();
 149
 150         /* Start inner kernel loop */
 151         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 152         {
 153
 154             /* Get j neighbor index, and coordinate index */
 155             jnrA             = jjnr[jidx];
 156             jnrB             = jjnr[jidx+1];
 157             jnrC             = jjnr[jidx+2];
 158             jnrD             = jjnr[jidx+3];
 159             j_coord_offsetA  = DIM*jnrA;
 160             j_coord_offsetB  = DIM*jnrB;
 161             j_coord_offsetC  = DIM*jnrC;
 162             j_coord_offsetD  = DIM*jnrD;
 163
 164             /* load j atom coordinates */
 165             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 166                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 167                                                  &jx0,&jy0,&jz0);
 168
 169             /* Calculate displacement vector */
 170             dx00             = _mm256_sub_pd(ix0,jx0);
 171             dy00             = _mm256_sub_pd(iy0,jy0);
 172             dz00             = _mm256_sub_pd(iz0,jz0);
 173
 174             /* Calculate squared distance and things based on it */
 175             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 176
 177             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 178
 179             /* Load parameters for j particles */
 180             vdwjidx0A        = 2*vdwtype[jnrA+0];
 181             vdwjidx0B        = 2*vdwtype[jnrB+0];
 182             vdwjidx0C        = 2*vdwtype[jnrC+0];
 183             vdwjidx0D        = 2*vdwtype[jnrD+0];
 184
 185             /**************************
 186              * CALCULATE INTERACTIONS *
 187              **************************/
 188
 189             r00              = _mm256_mul_pd(rsq00,rinv00);
 190
 191             /* Compute parameters for interactions between i and j atoms */
 192             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 193                                             vdwioffsetptr0+vdwjidx0B,
 194                                             vdwioffsetptr0+vdwjidx0C,
 195                                             vdwioffsetptr0+vdwjidx0D,
 196                                             &c6_00,&c12_00);
 197
 198             /* Calculate table index by multiplying r with table scale and truncate to integer */
 199             rt               = _mm256_mul_pd(r00,vftabscale);
 200             vfitab           = _mm256_cvttpd_epi32(rt);
 201             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 202             vfitab           = _mm_slli_epi32(vfitab,3);
 203
 204             /* CUBIC SPLINE TABLE DISPERSION */
 205             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 206             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 207             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 208             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 209             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 210             Heps             = _mm256_mul_pd(vfeps,H);
 211             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 212             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 213             vvdw6            = _mm256_mul_pd(c6_00,VV);
 214             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 215             fvdw6            = _mm256_mul_pd(c6_00,FF);
 216
 217             /* CUBIC SPLINE TABLE REPULSION */
 218             vfitab           = _mm_add_epi32(vfitab,ifour);
 219             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 220             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 221             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 222             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 223             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 224             Heps             = _mm256_mul_pd(vfeps,H);
 225             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 226             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 227             vvdw12           = _mm256_mul_pd(c12_00,VV);
 228             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 229             fvdw12           = _mm256_mul_pd(c12_00,FF);
 230             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 231             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 232
 233             /* Update potential sum for this i atom from the interaction with this j atom. */
 234             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 235
 236             fscal            = fvdw;
 237
 238             /* Calculate temporary vectorial force */
 239             tx               = _mm256_mul_pd(fscal,dx00);
 240             ty               = _mm256_mul_pd(fscal,dy00);
 241             tz               = _mm256_mul_pd(fscal,dz00);
 242
 243             /* Update vectorial force */
 244             fix0             = _mm256_add_pd(fix0,tx);
 245             fiy0             = _mm256_add_pd(fiy0,ty);
 246             fiz0             = _mm256_add_pd(fiz0,tz);
 247
 248             fjptrA             = f+j_coord_offsetA;
 249             fjptrB             = f+j_coord_offsetB;
 250             fjptrC             = f+j_coord_offsetC;
 251             fjptrD             = f+j_coord_offsetD;
 252             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 253
 254             /* Inner loop uses 56 flops */
 255         }
 256
 257         if(jidx<j_index_end)
 258         {
 259
 260             /* Get j neighbor index, and coordinate index */
 261             jnrlistA         = jjnr[jidx];
 262             jnrlistB         = jjnr[jidx+1];
 263             jnrlistC         = jjnr[jidx+2];
 264             jnrlistD         = jjnr[jidx+3];
 265             /* Sign of each element will be negative for non-real atoms.
 266              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 267              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 268              */
 269             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 270
 271             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 272             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 273             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 274
 275             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 276             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 277             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 278             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 279             j_coord_offsetA  = DIM*jnrA;
 280             j_coord_offsetB  = DIM*jnrB;
 281             j_coord_offsetC  = DIM*jnrC;
 282             j_coord_offsetD  = DIM*jnrD;
 283
 284             /* load j atom coordinates */
 285             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 286                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 287                                                  &jx0,&jy0,&jz0);
 288
 289             /* Calculate displacement vector */
 290             dx00             = _mm256_sub_pd(ix0,jx0);
 291             dy00             = _mm256_sub_pd(iy0,jy0);
 292             dz00             = _mm256_sub_pd(iz0,jz0);
 293
 294             /* Calculate squared distance and things based on it */
 295             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 296
 297             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 298
 299             /* Load parameters for j particles */
 300             vdwjidx0A        = 2*vdwtype[jnrA+0];
 301             vdwjidx0B        = 2*vdwtype[jnrB+0];
 302             vdwjidx0C        = 2*vdwtype[jnrC+0];
 303             vdwjidx0D        = 2*vdwtype[jnrD+0];
 304
 305             /**************************
 306              * CALCULATE INTERACTIONS *
 307              **************************/
 308
 309             r00              = _mm256_mul_pd(rsq00,rinv00);
 310             r00              = _mm256_andnot_pd(dummy_mask,r00);
 311
 312             /* Compute parameters for interactions between i and j atoms */
 313             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 314                                             vdwioffsetptr0+vdwjidx0B,
 315                                             vdwioffsetptr0+vdwjidx0C,
 316                                             vdwioffsetptr0+vdwjidx0D,
 317                                             &c6_00,&c12_00);
 318
 319             /* Calculate table index by multiplying r with table scale and truncate to integer */
 320             rt               = _mm256_mul_pd(r00,vftabscale);
 321             vfitab           = _mm256_cvttpd_epi32(rt);
 322             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 323             vfitab           = _mm_slli_epi32(vfitab,3);
 324
 325             /* CUBIC SPLINE TABLE DISPERSION */
 326             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 327             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 328             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 329             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 330             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 331             Heps             = _mm256_mul_pd(vfeps,H);
 332             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 333             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 334             vvdw6            = _mm256_mul_pd(c6_00,VV);
 335             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 336             fvdw6            = _mm256_mul_pd(c6_00,FF);
 337
 338             /* CUBIC SPLINE TABLE REPULSION */
 339             vfitab           = _mm_add_epi32(vfitab,ifour);
 340             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 341             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 342             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 343             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 344             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 345             Heps             = _mm256_mul_pd(vfeps,H);
 346             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 347             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 348             vvdw12           = _mm256_mul_pd(c12_00,VV);
 349             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 350             fvdw12           = _mm256_mul_pd(c12_00,FF);
 351             vvdw             = _mm256_add_pd(vvdw12,vvdw6);
 352             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 353
 354             /* Update potential sum for this i atom from the interaction with this j atom. */
 355             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 356             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 357
 358             fscal            = fvdw;
 359
 360             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 361
 362             /* Calculate temporary vectorial force */
 363             tx               = _mm256_mul_pd(fscal,dx00);
 364             ty               = _mm256_mul_pd(fscal,dy00);
 365             tz               = _mm256_mul_pd(fscal,dz00);
 366
 367             /* Update vectorial force */
 368             fix0             = _mm256_add_pd(fix0,tx);
 369             fiy0             = _mm256_add_pd(fiy0,ty);
 370             fiz0             = _mm256_add_pd(fiz0,tz);
 371
 372             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 373             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 374             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 375             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 376             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 377
 378             /* Inner loop uses 57 flops */
 379         }
 380
 381         /* End of innermost loop */
 382
 383         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 384                                                  f+i_coord_offset,fshift+i_shift_offset);
 385
 386         ggid                        = gid[iidx];
 387         /* Update potential energies */
 388         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 389
 390         /* Increment number of inner iterations */
 391         inneriter                  += j_index_end - j_index_start;
 392
 393         /* Outer loop uses 7 flops */
 394     }
 395
 396     /* Increment number of outer iterations */
 397     outeriter        += nri;
 398
 399     /* Update outer/inner flops */
 400
 401     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*57);
 402 }
 403 /*
 404  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_double
 405  * Electrostatics interaction: None
 406  * VdW interaction:            CubicSplineTable
 407  * Geometry:                   Particle-Particle
 408  * Calculate force/pot:        Force
 409  */
 410 void
 411 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_double
 412                     (t_nblist * gmx_restrict                nlist,
 413                      rvec * gmx_restrict                    xx,
 414                      rvec * gmx_restrict                    ff,
 415                      t_forcerec * gmx_restrict              fr,
 416                      t_mdatoms * gmx_restrict               mdatoms,
 417                      nb_kernel_data_t * gmx_restrict        kernel_data,
 418                      t_nrnb * gmx_restrict                  nrnb)
 419 {
 420     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 421      * just 0 for non-waters.
 422      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 423      * jnr indices corresponding to data put in the four positions in the SIMD register.
 424      */
 425     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 426     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 427     int              jnrA,jnrB,jnrC,jnrD;
 428     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 429     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 430     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 431     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 432     real             rcutoff_scalar;
 433     real             *shiftvec,*fshift,*x,*f;
 434     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 435     real             scratch[4*DIM];
 436     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 437     real *           vdwioffsetptr0;
 438     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 439     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 440     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 441     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 442     int              nvdwtype;
 443     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 444     int              *vdwtype;
 445     real             *vdwparam;
 446     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 447     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 448     __m128i          vfitab;
 449     __m128i          ifour       = _mm_set1_epi32(4);
 450     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 451     real             *vftab;
 452     __m256d          dummy_mask,cutoff_mask;
 453     __m128           tmpmask0,tmpmask1;
 454     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 455     __m256d          one     = _mm256_set1_pd(1.0);
 456     __m256d          two     = _mm256_set1_pd(2.0);
 457     x                = xx[0];
 458     f                = ff[0];
 459
 460     nri              = nlist->nri;
 461     iinr             = nlist->iinr;
 462     jindex           = nlist->jindex;
 463     jjnr             = nlist->jjnr;
 464     shiftidx         = nlist->shift;
 465     gid              = nlist->gid;
 466     shiftvec         = fr->shift_vec[0];
 467     fshift           = fr->fshift[0];
 468     nvdwtype         = fr->ntype;
 469     vdwparam         = fr->nbfp;
 470     vdwtype          = mdatoms->typeA;
 471
 472     vftab            = kernel_data->table_vdw->data;
 473     vftabscale       = _mm256_set1_pd(kernel_data->table_vdw->scale);
 474
 475     /* Avoid stupid compiler warnings */
 476     jnrA = jnrB = jnrC = jnrD = 0;
 477     j_coord_offsetA = 0;
 478     j_coord_offsetB = 0;
 479     j_coord_offsetC = 0;
 480     j_coord_offsetD = 0;
 481
 482     outeriter        = 0;
 483     inneriter        = 0;
 484
 485     for(iidx=0;iidx<4*DIM;iidx++)
 486     {
 487         scratch[iidx] = 0.0;
 488     }
 489
 490     /* Start outer loop over neighborlists */
 491     for(iidx=0; iidx<nri; iidx++)
 492     {
 493         /* Load shift vector for this list */
 494         i_shift_offset   = DIM*shiftidx[iidx];
 495
 496         /* Load limits for loop over neighbors */
 497         j_index_start    = jindex[iidx];
 498         j_index_end      = jindex[iidx+1];
 499
 500         /* Get outer coordinate index */
 501         inr              = iinr[iidx];
 502         i_coord_offset   = DIM*inr;
 503
 504         /* Load i particle coords and add shift vector */
 505         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 506
 507         fix0             = _mm256_setzero_pd();
 508         fiy0             = _mm256_setzero_pd();
 509         fiz0             = _mm256_setzero_pd();
 510
 511         /* Load parameters for i particles */
 512         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 513
 514         /* Start inner kernel loop */
 515         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 516         {
 517
 518             /* Get j neighbor index, and coordinate index */
 519             jnrA             = jjnr[jidx];
 520             jnrB             = jjnr[jidx+1];
 521             jnrC             = jjnr[jidx+2];
 522             jnrD             = jjnr[jidx+3];
 523             j_coord_offsetA  = DIM*jnrA;
 524             j_coord_offsetB  = DIM*jnrB;
 525             j_coord_offsetC  = DIM*jnrC;
 526             j_coord_offsetD  = DIM*jnrD;
 527
 528             /* load j atom coordinates */
 529             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 530                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 531                                                  &jx0,&jy0,&jz0);
 532
 533             /* Calculate displacement vector */
 534             dx00             = _mm256_sub_pd(ix0,jx0);
 535             dy00             = _mm256_sub_pd(iy0,jy0);
 536             dz00             = _mm256_sub_pd(iz0,jz0);
 537
 538             /* Calculate squared distance and things based on it */
 539             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 540
 541             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 542
 543             /* Load parameters for j particles */
 544             vdwjidx0A        = 2*vdwtype[jnrA+0];
 545             vdwjidx0B        = 2*vdwtype[jnrB+0];
 546             vdwjidx0C        = 2*vdwtype[jnrC+0];
 547             vdwjidx0D        = 2*vdwtype[jnrD+0];
 548
 549             /**************************
 550              * CALCULATE INTERACTIONS *
 551              **************************/
 552
 553             r00              = _mm256_mul_pd(rsq00,rinv00);
 554
 555             /* Compute parameters for interactions between i and j atoms */
 556             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 557                                             vdwioffsetptr0+vdwjidx0B,
 558                                             vdwioffsetptr0+vdwjidx0C,
 559                                             vdwioffsetptr0+vdwjidx0D,
 560                                             &c6_00,&c12_00);
 561
 562             /* Calculate table index by multiplying r with table scale and truncate to integer */
 563             rt               = _mm256_mul_pd(r00,vftabscale);
 564             vfitab           = _mm256_cvttpd_epi32(rt);
 565             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 566             vfitab           = _mm_slli_epi32(vfitab,3);
 567
 568             /* CUBIC SPLINE TABLE DISPERSION */
 569             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 570             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 571             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 572             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 573             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 574             Heps             = _mm256_mul_pd(vfeps,H);
 575             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 576             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 577             fvdw6            = _mm256_mul_pd(c6_00,FF);
 578
 579             /* CUBIC SPLINE TABLE REPULSION */
 580             vfitab           = _mm_add_epi32(vfitab,ifour);
 581             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 582             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 583             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 584             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 585             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 586             Heps             = _mm256_mul_pd(vfeps,H);
 587             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 588             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 589             fvdw12           = _mm256_mul_pd(c12_00,FF);
 590             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 591
 592             fscal            = fvdw;
 593
 594             /* Calculate temporary vectorial force */
 595             tx               = _mm256_mul_pd(fscal,dx00);
 596             ty               = _mm256_mul_pd(fscal,dy00);
 597             tz               = _mm256_mul_pd(fscal,dz00);
 598
 599             /* Update vectorial force */
 600             fix0             = _mm256_add_pd(fix0,tx);
 601             fiy0             = _mm256_add_pd(fiy0,ty);
 602             fiz0             = _mm256_add_pd(fiz0,tz);
 603
 604             fjptrA             = f+j_coord_offsetA;
 605             fjptrB             = f+j_coord_offsetB;
 606             fjptrC             = f+j_coord_offsetC;
 607             fjptrD             = f+j_coord_offsetD;
 608             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 609
 610             /* Inner loop uses 48 flops */
 611         }
 612
 613         if(jidx<j_index_end)
 614         {
 615
 616             /* Get j neighbor index, and coordinate index */
 617             jnrlistA         = jjnr[jidx];
 618             jnrlistB         = jjnr[jidx+1];
 619             jnrlistC         = jjnr[jidx+2];
 620             jnrlistD         = jjnr[jidx+3];
 621             /* Sign of each element will be negative for non-real atoms.
 622              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 623              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 624              */
 625             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 626
 627             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 628             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 629             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 630
 631             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 632             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 633             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 634             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 635             j_coord_offsetA  = DIM*jnrA;
 636             j_coord_offsetB  = DIM*jnrB;
 637             j_coord_offsetC  = DIM*jnrC;
 638             j_coord_offsetD  = DIM*jnrD;
 639
 640             /* load j atom coordinates */
 641             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 642                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 643                                                  &jx0,&jy0,&jz0);
 644
 645             /* Calculate displacement vector */
 646             dx00             = _mm256_sub_pd(ix0,jx0);
 647             dy00             = _mm256_sub_pd(iy0,jy0);
 648             dz00             = _mm256_sub_pd(iz0,jz0);
 649
 650             /* Calculate squared distance and things based on it */
 651             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 652
 653             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 654
 655             /* Load parameters for j particles */
 656             vdwjidx0A        = 2*vdwtype[jnrA+0];
 657             vdwjidx0B        = 2*vdwtype[jnrB+0];
 658             vdwjidx0C        = 2*vdwtype[jnrC+0];
 659             vdwjidx0D        = 2*vdwtype[jnrD+0];
 660
 661             /**************************
 662              * CALCULATE INTERACTIONS *
 663              **************************/
 664
 665             r00              = _mm256_mul_pd(rsq00,rinv00);
 666             r00              = _mm256_andnot_pd(dummy_mask,r00);
 667
 668             /* Compute parameters for interactions between i and j atoms */
 669             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 670                                             vdwioffsetptr0+vdwjidx0B,
 671                                             vdwioffsetptr0+vdwjidx0C,
 672                                             vdwioffsetptr0+vdwjidx0D,
 673                                             &c6_00,&c12_00);
 674
 675             /* Calculate table index by multiplying r with table scale and truncate to integer */
 676             rt               = _mm256_mul_pd(r00,vftabscale);
 677             vfitab           = _mm256_cvttpd_epi32(rt);
 678             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 679             vfitab           = _mm_slli_epi32(vfitab,3);
 680
 681             /* CUBIC SPLINE TABLE DISPERSION */
 682             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 683             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 684             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 685             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 686             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 687             Heps             = _mm256_mul_pd(vfeps,H);
 688             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 689             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 690             fvdw6            = _mm256_mul_pd(c6_00,FF);
 691
 692             /* CUBIC SPLINE TABLE REPULSION */
 693             vfitab           = _mm_add_epi32(vfitab,ifour);
 694             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 695             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 696             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 697             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 698             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 699             Heps             = _mm256_mul_pd(vfeps,H);
 700             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 701             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 702             fvdw12           = _mm256_mul_pd(c12_00,FF);
 703             fvdw             = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
 704
 705             fscal            = fvdw;
 706
 707             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 708
 709             /* Calculate temporary vectorial force */
 710             tx               = _mm256_mul_pd(fscal,dx00);
 711             ty               = _mm256_mul_pd(fscal,dy00);
 712             tz               = _mm256_mul_pd(fscal,dz00);
 713
 714             /* Update vectorial force */
 715             fix0             = _mm256_add_pd(fix0,tx);
 716             fiy0             = _mm256_add_pd(fiy0,ty);
 717             fiz0             = _mm256_add_pd(fiz0,tz);
 718
 719             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 720             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 721             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 722             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 723             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 724
 725             /* Inner loop uses 49 flops */
 726         }
 727
 728         /* End of innermost loop */
 729
 730         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 731                                                  f+i_coord_offset,fshift+i_shift_offset);
 732
 733         /* Increment number of inner iterations */
 734         inneriter                  += j_index_end - j_index_start;
 735
 736         /* Outer loop uses 6 flops */
 737     }
 738
 739     /* Increment number of outer iterations */
 740     outeriter        += nri;
 741
 742     /* Update outer/inner flops */
 743
 744     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*49);
 745 }