src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCSTab_VdwLJ_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_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: CubicSplineTable
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCSTab_VdwLJ_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     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  76     real             *charge;
  77     int              nvdwtype;
  78     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  79     int              *vdwtype;
  80     real             *vdwparam;
  81     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  82     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  83     __m128i          vfitab;
  84     __m128i          ifour       = _mm_set1_epi32(4);
  85     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  86     real             *vftab;
  87     __m256d          dummy_mask,cutoff_mask;
  88     __m128           tmpmask0,tmpmask1;
  89     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  90     __m256d          one     = _mm256_set1_pd(1.0);
  91     __m256d          two     = _mm256_set1_pd(2.0);
  92     x                = xx[0];
  93     f                = ff[0];
  94
  95     nri              = nlist->nri;
  96     iinr             = nlist->iinr;
  97     jindex           = nlist->jindex;
  98     jjnr             = nlist->jjnr;
  99     shiftidx         = nlist->shift;
 100     gid              = nlist->gid;
 101     shiftvec         = fr->shift_vec[0];
 102     fshift           = fr->fshift[0];
 103     facel            = _mm256_set1_pd(fr->epsfac);
 104     charge           = mdatoms->chargeA;
 105     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     vftab            = kernel_data->table_elec->data;
 110     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 111
 112     /* Avoid stupid compiler warnings */
 113     jnrA = jnrB = jnrC = jnrD = 0;
 114     j_coord_offsetA = 0;
 115     j_coord_offsetB = 0;
 116     j_coord_offsetC = 0;
 117     j_coord_offsetD = 0;
 118
 119     outeriter        = 0;
 120     inneriter        = 0;
 121
 122     for(iidx=0;iidx<4*DIM;iidx++)
 123     {
 124         scratch[iidx] = 0.0;
 125     }
 126
 127     /* Start outer loop over neighborlists */
 128     for(iidx=0; iidx<nri; iidx++)
 129     {
 130         /* Load shift vector for this list */
 131         i_shift_offset   = DIM*shiftidx[iidx];
 132
 133         /* Load limits for loop over neighbors */
 134         j_index_start    = jindex[iidx];
 135         j_index_end      = jindex[iidx+1];
 136
 137         /* Get outer coordinate index */
 138         inr              = iinr[iidx];
 139         i_coord_offset   = DIM*inr;
 140
 141         /* Load i particle coords and add shift vector */
 142         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 143
 144         fix0             = _mm256_setzero_pd();
 145         fiy0             = _mm256_setzero_pd();
 146         fiz0             = _mm256_setzero_pd();
 147
 148         /* Load parameters for i particles */
 149         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 150         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 151
 152         /* Reset potential sums */
 153         velecsum         = _mm256_setzero_pd();
 154         vvdwsum          = _mm256_setzero_pd();
 155
 156         /* Start inner kernel loop */
 157         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 158         {
 159
 160             /* Get j neighbor index, and coordinate index */
 161             jnrA             = jjnr[jidx];
 162             jnrB             = jjnr[jidx+1];
 163             jnrC             = jjnr[jidx+2];
 164             jnrD             = jjnr[jidx+3];
 165             j_coord_offsetA  = DIM*jnrA;
 166             j_coord_offsetB  = DIM*jnrB;
 167             j_coord_offsetC  = DIM*jnrC;
 168             j_coord_offsetD  = DIM*jnrD;
 169
 170             /* load j atom coordinates */
 171             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 172                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 173                                                  &jx0,&jy0,&jz0);
 174
 175             /* Calculate displacement vector */
 176             dx00             = _mm256_sub_pd(ix0,jx0);
 177             dy00             = _mm256_sub_pd(iy0,jy0);
 178             dz00             = _mm256_sub_pd(iz0,jz0);
 179
 180             /* Calculate squared distance and things based on it */
 181             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 182
 183             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 184
 185             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 186
 187             /* Load parameters for j particles */
 188             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 189                                                                  charge+jnrC+0,charge+jnrD+0);
 190             vdwjidx0A        = 2*vdwtype[jnrA+0];
 191             vdwjidx0B        = 2*vdwtype[jnrB+0];
 192             vdwjidx0C        = 2*vdwtype[jnrC+0];
 193             vdwjidx0D        = 2*vdwtype[jnrD+0];
 194
 195             /**************************
 196              * CALCULATE INTERACTIONS *
 197              **************************/
 198
 199             r00              = _mm256_mul_pd(rsq00,rinv00);
 200
 201             /* Compute parameters for interactions between i and j atoms */
 202             qq00             = _mm256_mul_pd(iq0,jq0);
 203             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 204                                             vdwioffsetptr0+vdwjidx0B,
 205                                             vdwioffsetptr0+vdwjidx0C,
 206                                             vdwioffsetptr0+vdwjidx0D,
 207                                             &c6_00,&c12_00);
 208
 209             /* Calculate table index by multiplying r with table scale and truncate to integer */
 210             rt               = _mm256_mul_pd(r00,vftabscale);
 211             vfitab           = _mm256_cvttpd_epi32(rt);
 212             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 213             vfitab           = _mm_slli_epi32(vfitab,2);
 214
 215             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 216             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 217             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 218             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 219             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 220             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 221             Heps             = _mm256_mul_pd(vfeps,H);
 222             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 223             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 224             velec            = _mm256_mul_pd(qq00,VV);
 225             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 226             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 227
 228             /* LENNARD-JONES DISPERSION/REPULSION */
 229
 230             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 231             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 232             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 233             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 234             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 235
 236             /* Update potential sum for this i atom from the interaction with this j atom. */
 237             velecsum         = _mm256_add_pd(velecsum,velec);
 238             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 239
 240             fscal            = _mm256_add_pd(felec,fvdw);
 241
 242             /* Calculate temporary vectorial force */
 243             tx               = _mm256_mul_pd(fscal,dx00);
 244             ty               = _mm256_mul_pd(fscal,dy00);
 245             tz               = _mm256_mul_pd(fscal,dz00);
 246
 247             /* Update vectorial force */
 248             fix0             = _mm256_add_pd(fix0,tx);
 249             fiy0             = _mm256_add_pd(fiy0,ty);
 250             fiz0             = _mm256_add_pd(fiz0,tz);
 251
 252             fjptrA             = f+j_coord_offsetA;
 253             fjptrB             = f+j_coord_offsetB;
 254             fjptrC             = f+j_coord_offsetC;
 255             fjptrD             = f+j_coord_offsetD;
 256             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 257
 258             /* Inner loop uses 56 flops */
 259         }
 260
 261         if(jidx<j_index_end)
 262         {
 263
 264             /* Get j neighbor index, and coordinate index */
 265             jnrlistA         = jjnr[jidx];
 266             jnrlistB         = jjnr[jidx+1];
 267             jnrlistC         = jjnr[jidx+2];
 268             jnrlistD         = jjnr[jidx+3];
 269             /* Sign of each element will be negative for non-real atoms.
 270              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 271              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 272              */
 273             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 274
 275             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 276             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 277             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 278
 279             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 280             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 281             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 282             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 283             j_coord_offsetA  = DIM*jnrA;
 284             j_coord_offsetB  = DIM*jnrB;
 285             j_coord_offsetC  = DIM*jnrC;
 286             j_coord_offsetD  = DIM*jnrD;
 287
 288             /* load j atom coordinates */
 289             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 290                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 291                                                  &jx0,&jy0,&jz0);
 292
 293             /* Calculate displacement vector */
 294             dx00             = _mm256_sub_pd(ix0,jx0);
 295             dy00             = _mm256_sub_pd(iy0,jy0);
 296             dz00             = _mm256_sub_pd(iz0,jz0);
 297
 298             /* Calculate squared distance and things based on it */
 299             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 300
 301             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 302
 303             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 304
 305             /* Load parameters for j particles */
 306             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 307                                                                  charge+jnrC+0,charge+jnrD+0);
 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              = _mm256_mul_pd(rsq00,rinv00);
 318             r00              = _mm256_andnot_pd(dummy_mask,r00);
 319
 320             /* Compute parameters for interactions between i and j atoms */
 321             qq00             = _mm256_mul_pd(iq0,jq0);
 322             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 323                                             vdwioffsetptr0+vdwjidx0B,
 324                                             vdwioffsetptr0+vdwjidx0C,
 325                                             vdwioffsetptr0+vdwjidx0D,
 326                                             &c6_00,&c12_00);
 327
 328             /* Calculate table index by multiplying r with table scale and truncate to integer */
 329             rt               = _mm256_mul_pd(r00,vftabscale);
 330             vfitab           = _mm256_cvttpd_epi32(rt);
 331             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 332             vfitab           = _mm_slli_epi32(vfitab,2);
 333
 334             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 335             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 336             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 337             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 338             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 339             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 340             Heps             = _mm256_mul_pd(vfeps,H);
 341             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 342             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 343             velec            = _mm256_mul_pd(qq00,VV);
 344             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 345             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 346
 347             /* LENNARD-JONES DISPERSION/REPULSION */
 348
 349             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 350             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 351             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 352             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 353             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velec            = _mm256_andnot_pd(dummy_mask,velec);
 357             velecsum         = _mm256_add_pd(velecsum,velec);
 358             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 359             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 360
 361             fscal            = _mm256_add_pd(felec,fvdw);
 362
 363             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 364
 365             /* Calculate temporary vectorial force */
 366             tx               = _mm256_mul_pd(fscal,dx00);
 367             ty               = _mm256_mul_pd(fscal,dy00);
 368             tz               = _mm256_mul_pd(fscal,dz00);
 369
 370             /* Update vectorial force */
 371             fix0             = _mm256_add_pd(fix0,tx);
 372             fiy0             = _mm256_add_pd(fiy0,ty);
 373             fiz0             = _mm256_add_pd(fiz0,tz);
 374
 375             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 376             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 377             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 378             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 379             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 380
 381             /* Inner loop uses 57 flops */
 382         }
 383
 384         /* End of innermost loop */
 385
 386         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 387                                                  f+i_coord_offset,fshift+i_shift_offset);
 388
 389         ggid                        = gid[iidx];
 390         /* Update potential energies */
 391         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 392         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 393
 394         /* Increment number of inner iterations */
 395         inneriter                  += j_index_end - j_index_start;
 396
 397         /* Outer loop uses 9 flops */
 398     }
 399
 400     /* Increment number of outer iterations */
 401     outeriter        += nri;
 402
 403     /* Update outer/inner flops */
 404
 405     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*57);
 406 }
 407 /*
 408  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 409  * Electrostatics interaction: CubicSplineTable
 410  * VdW interaction:            LennardJones
 411  * Geometry:                   Particle-Particle
 412  * Calculate force/pot:        Force
 413  */
 414 void
 415 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_256_double
 416                     (t_nblist * gmx_restrict                nlist,
 417                      rvec * gmx_restrict                    xx,
 418                      rvec * gmx_restrict                    ff,
 419                      t_forcerec * gmx_restrict              fr,
 420                      t_mdatoms * gmx_restrict               mdatoms,
 421                      nb_kernel_data_t * gmx_restrict        kernel_data,
 422                      t_nrnb * gmx_restrict                  nrnb)
 423 {
 424     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 425      * just 0 for non-waters.
 426      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 427      * jnr indices corresponding to data put in the four positions in the SIMD register.
 428      */
 429     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 430     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 431     int              jnrA,jnrB,jnrC,jnrD;
 432     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 433     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 434     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 435     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 436     real             rcutoff_scalar;
 437     real             *shiftvec,*fshift,*x,*f;
 438     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 439     real             scratch[4*DIM];
 440     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 441     real *           vdwioffsetptr0;
 442     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 443     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 444     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 445     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 446     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 447     real             *charge;
 448     int              nvdwtype;
 449     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 450     int              *vdwtype;
 451     real             *vdwparam;
 452     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 453     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 454     __m128i          vfitab;
 455     __m128i          ifour       = _mm_set1_epi32(4);
 456     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 457     real             *vftab;
 458     __m256d          dummy_mask,cutoff_mask;
 459     __m128           tmpmask0,tmpmask1;
 460     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 461     __m256d          one     = _mm256_set1_pd(1.0);
 462     __m256d          two     = _mm256_set1_pd(2.0);
 463     x                = xx[0];
 464     f                = ff[0];
 465
 466     nri              = nlist->nri;
 467     iinr             = nlist->iinr;
 468     jindex           = nlist->jindex;
 469     jjnr             = nlist->jjnr;
 470     shiftidx         = nlist->shift;
 471     gid              = nlist->gid;
 472     shiftvec         = fr->shift_vec[0];
 473     fshift           = fr->fshift[0];
 474     facel            = _mm256_set1_pd(fr->epsfac);
 475     charge           = mdatoms->chargeA;
 476     nvdwtype         = fr->ntype;
 477     vdwparam         = fr->nbfp;
 478     vdwtype          = mdatoms->typeA;
 479
 480     vftab            = kernel_data->table_elec->data;
 481     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 482
 483     /* Avoid stupid compiler warnings */
 484     jnrA = jnrB = jnrC = jnrD = 0;
 485     j_coord_offsetA = 0;
 486     j_coord_offsetB = 0;
 487     j_coord_offsetC = 0;
 488     j_coord_offsetD = 0;
 489
 490     outeriter        = 0;
 491     inneriter        = 0;
 492
 493     for(iidx=0;iidx<4*DIM;iidx++)
 494     {
 495         scratch[iidx] = 0.0;
 496     }
 497
 498     /* Start outer loop over neighborlists */
 499     for(iidx=0; iidx<nri; iidx++)
 500     {
 501         /* Load shift vector for this list */
 502         i_shift_offset   = DIM*shiftidx[iidx];
 503
 504         /* Load limits for loop over neighbors */
 505         j_index_start    = jindex[iidx];
 506         j_index_end      = jindex[iidx+1];
 507
 508         /* Get outer coordinate index */
 509         inr              = iinr[iidx];
 510         i_coord_offset   = DIM*inr;
 511
 512         /* Load i particle coords and add shift vector */
 513         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 514
 515         fix0             = _mm256_setzero_pd();
 516         fiy0             = _mm256_setzero_pd();
 517         fiz0             = _mm256_setzero_pd();
 518
 519         /* Load parameters for i particles */
 520         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 521         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 522
 523         /* Start inner kernel loop */
 524         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 525         {
 526
 527             /* Get j neighbor index, and coordinate index */
 528             jnrA             = jjnr[jidx];
 529             jnrB             = jjnr[jidx+1];
 530             jnrC             = jjnr[jidx+2];
 531             jnrD             = jjnr[jidx+3];
 532             j_coord_offsetA  = DIM*jnrA;
 533             j_coord_offsetB  = DIM*jnrB;
 534             j_coord_offsetC  = DIM*jnrC;
 535             j_coord_offsetD  = DIM*jnrD;
 536
 537             /* load j atom coordinates */
 538             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 539                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 540                                                  &jx0,&jy0,&jz0);
 541
 542             /* Calculate displacement vector */
 543             dx00             = _mm256_sub_pd(ix0,jx0);
 544             dy00             = _mm256_sub_pd(iy0,jy0);
 545             dz00             = _mm256_sub_pd(iz0,jz0);
 546
 547             /* Calculate squared distance and things based on it */
 548             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 549
 550             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 551
 552             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 553
 554             /* Load parameters for j particles */
 555             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 556                                                                  charge+jnrC+0,charge+jnrD+0);
 557             vdwjidx0A        = 2*vdwtype[jnrA+0];
 558             vdwjidx0B        = 2*vdwtype[jnrB+0];
 559             vdwjidx0C        = 2*vdwtype[jnrC+0];
 560             vdwjidx0D        = 2*vdwtype[jnrD+0];
 561
 562             /**************************
 563              * CALCULATE INTERACTIONS *
 564              **************************/
 565
 566             r00              = _mm256_mul_pd(rsq00,rinv00);
 567
 568             /* Compute parameters for interactions between i and j atoms */
 569             qq00             = _mm256_mul_pd(iq0,jq0);
 570             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 571                                             vdwioffsetptr0+vdwjidx0B,
 572                                             vdwioffsetptr0+vdwjidx0C,
 573                                             vdwioffsetptr0+vdwjidx0D,
 574                                             &c6_00,&c12_00);
 575
 576             /* Calculate table index by multiplying r with table scale and truncate to integer */
 577             rt               = _mm256_mul_pd(r00,vftabscale);
 578             vfitab           = _mm256_cvttpd_epi32(rt);
 579             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 580             vfitab           = _mm_slli_epi32(vfitab,2);
 581
 582             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 583             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 584             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 585             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 586             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 587             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 588             Heps             = _mm256_mul_pd(vfeps,H);
 589             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 590             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 591             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 592
 593             /* LENNARD-JONES DISPERSION/REPULSION */
 594
 595             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 596             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 597
 598             fscal            = _mm256_add_pd(felec,fvdw);
 599
 600             /* Calculate temporary vectorial force */
 601             tx               = _mm256_mul_pd(fscal,dx00);
 602             ty               = _mm256_mul_pd(fscal,dy00);
 603             tz               = _mm256_mul_pd(fscal,dz00);
 604
 605             /* Update vectorial force */
 606             fix0             = _mm256_add_pd(fix0,tx);
 607             fiy0             = _mm256_add_pd(fiy0,ty);
 608             fiz0             = _mm256_add_pd(fiz0,tz);
 609
 610             fjptrA             = f+j_coord_offsetA;
 611             fjptrB             = f+j_coord_offsetB;
 612             fjptrC             = f+j_coord_offsetC;
 613             fjptrD             = f+j_coord_offsetD;
 614             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 615
 616             /* Inner loop uses 47 flops */
 617         }
 618
 619         if(jidx<j_index_end)
 620         {
 621
 622             /* Get j neighbor index, and coordinate index */
 623             jnrlistA         = jjnr[jidx];
 624             jnrlistB         = jjnr[jidx+1];
 625             jnrlistC         = jjnr[jidx+2];
 626             jnrlistD         = jjnr[jidx+3];
 627             /* Sign of each element will be negative for non-real atoms.
 628              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 629              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 630              */
 631             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 632
 633             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 634             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 635             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 636
 637             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 638             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 639             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 640             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 641             j_coord_offsetA  = DIM*jnrA;
 642             j_coord_offsetB  = DIM*jnrB;
 643             j_coord_offsetC  = DIM*jnrC;
 644             j_coord_offsetD  = DIM*jnrD;
 645
 646             /* load j atom coordinates */
 647             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 648                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 649                                                  &jx0,&jy0,&jz0);
 650
 651             /* Calculate displacement vector */
 652             dx00             = _mm256_sub_pd(ix0,jx0);
 653             dy00             = _mm256_sub_pd(iy0,jy0);
 654             dz00             = _mm256_sub_pd(iz0,jz0);
 655
 656             /* Calculate squared distance and things based on it */
 657             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 658
 659             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 660
 661             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 662
 663             /* Load parameters for j particles */
 664             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 665                                                                  charge+jnrC+0,charge+jnrD+0);
 666             vdwjidx0A        = 2*vdwtype[jnrA+0];
 667             vdwjidx0B        = 2*vdwtype[jnrB+0];
 668             vdwjidx0C        = 2*vdwtype[jnrC+0];
 669             vdwjidx0D        = 2*vdwtype[jnrD+0];
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             r00              = _mm256_mul_pd(rsq00,rinv00);
 676             r00              = _mm256_andnot_pd(dummy_mask,r00);
 677
 678             /* Compute parameters for interactions between i and j atoms */
 679             qq00             = _mm256_mul_pd(iq0,jq0);
 680             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 681                                             vdwioffsetptr0+vdwjidx0B,
 682                                             vdwioffsetptr0+vdwjidx0C,
 683                                             vdwioffsetptr0+vdwjidx0D,
 684                                             &c6_00,&c12_00);
 685
 686             /* Calculate table index by multiplying r with table scale and truncate to integer */
 687             rt               = _mm256_mul_pd(r00,vftabscale);
 688             vfitab           = _mm256_cvttpd_epi32(rt);
 689             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 690             vfitab           = _mm_slli_epi32(vfitab,2);
 691
 692             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 693             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 694             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 695             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 696             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 697             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 698             Heps             = _mm256_mul_pd(vfeps,H);
 699             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 700             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 701             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 702
 703             /* LENNARD-JONES DISPERSION/REPULSION */
 704
 705             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 706             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 707
 708             fscal            = _mm256_add_pd(felec,fvdw);
 709
 710             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 711
 712             /* Calculate temporary vectorial force */
 713             tx               = _mm256_mul_pd(fscal,dx00);
 714             ty               = _mm256_mul_pd(fscal,dy00);
 715             tz               = _mm256_mul_pd(fscal,dz00);
 716
 717             /* Update vectorial force */
 718             fix0             = _mm256_add_pd(fix0,tx);
 719             fiy0             = _mm256_add_pd(fiy0,ty);
 720             fiz0             = _mm256_add_pd(fiz0,tz);
 721
 722             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 723             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 724             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 725             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 726             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 727
 728             /* Inner loop uses 48 flops */
 729         }
 730
 731         /* End of innermost loop */
 732
 733         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 734                                                  f+i_coord_offset,fshift+i_shift_offset);
 735
 736         /* Increment number of inner iterations */
 737         inneriter                  += j_index_end - j_index_start;
 738
 739         /* Outer loop uses 7 flops */
 740     }
 741
 742     /* Increment number of outer iterations */
 743     outeriter        += nri;
 744
 745     /* Update outer/inner flops */
 746
 747     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*48);
 748 }