src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_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_ElecEw_VdwLJ_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEw_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          ewitab;
  84     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  85     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  86     real             *ewtab;
  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     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 110     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 111     beta2            = _mm256_mul_pd(beta,beta);
 112     beta3            = _mm256_mul_pd(beta,beta2);
 113
 114     ewtab            = fr->ic->tabq_coul_FDV0;
 115     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 116     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 117
 118     /* Avoid stupid compiler warnings */
 119     jnrA = jnrB = jnrC = jnrD = 0;
 120     j_coord_offsetA = 0;
 121     j_coord_offsetB = 0;
 122     j_coord_offsetC = 0;
 123     j_coord_offsetD = 0;
 124
 125     outeriter        = 0;
 126     inneriter        = 0;
 127
 128     for(iidx=0;iidx<4*DIM;iidx++)
 129     {
 130         scratch[iidx] = 0.0;
 131     }
 132
 133     /* Start outer loop over neighborlists */
 134     for(iidx=0; iidx<nri; iidx++)
 135     {
 136         /* Load shift vector for this list */
 137         i_shift_offset   = DIM*shiftidx[iidx];
 138
 139         /* Load limits for loop over neighbors */
 140         j_index_start    = jindex[iidx];
 141         j_index_end      = jindex[iidx+1];
 142
 143         /* Get outer coordinate index */
 144         inr              = iinr[iidx];
 145         i_coord_offset   = DIM*inr;
 146
 147         /* Load i particle coords and add shift vector */
 148         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 149
 150         fix0             = _mm256_setzero_pd();
 151         fiy0             = _mm256_setzero_pd();
 152         fiz0             = _mm256_setzero_pd();
 153
 154         /* Load parameters for i particles */
 155         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 156         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 157
 158         /* Reset potential sums */
 159         velecsum         = _mm256_setzero_pd();
 160         vvdwsum          = _mm256_setzero_pd();
 161
 162         /* Start inner kernel loop */
 163         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 164         {
 165
 166             /* Get j neighbor index, and coordinate index */
 167             jnrA             = jjnr[jidx];
 168             jnrB             = jjnr[jidx+1];
 169             jnrC             = jjnr[jidx+2];
 170             jnrD             = jjnr[jidx+3];
 171             j_coord_offsetA  = DIM*jnrA;
 172             j_coord_offsetB  = DIM*jnrB;
 173             j_coord_offsetC  = DIM*jnrC;
 174             j_coord_offsetD  = DIM*jnrD;
 175
 176             /* load j atom coordinates */
 177             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 178                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 179                                                  &jx0,&jy0,&jz0);
 180
 181             /* Calculate displacement vector */
 182             dx00             = _mm256_sub_pd(ix0,jx0);
 183             dy00             = _mm256_sub_pd(iy0,jy0);
 184             dz00             = _mm256_sub_pd(iz0,jz0);
 185
 186             /* Calculate squared distance and things based on it */
 187             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 188
 189             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 190
 191             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 192
 193             /* Load parameters for j particles */
 194             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 195                                                                  charge+jnrC+0,charge+jnrD+0);
 196             vdwjidx0A        = 2*vdwtype[jnrA+0];
 197             vdwjidx0B        = 2*vdwtype[jnrB+0];
 198             vdwjidx0C        = 2*vdwtype[jnrC+0];
 199             vdwjidx0D        = 2*vdwtype[jnrD+0];
 200
 201             /**************************
 202              * CALCULATE INTERACTIONS *
 203              **************************/
 204
 205             r00              = _mm256_mul_pd(rsq00,rinv00);
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             qq00             = _mm256_mul_pd(iq0,jq0);
 209             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 210                                             vdwioffsetptr0+vdwjidx0B,
 211                                             vdwioffsetptr0+vdwjidx0C,
 212                                             vdwioffsetptr0+vdwjidx0D,
 213                                             &c6_00,&c12_00);
 214
 215             /* EWALD ELECTROSTATICS */
 216
 217             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 218             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 219             ewitab           = _mm256_cvttpd_epi32(ewrt);
 220             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 221             ewitab           = _mm_slli_epi32(ewitab,2);
 222             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 223             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 224             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 225             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 226             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 227             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 228             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 229             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 230             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 231
 232             /* LENNARD-JONES DISPERSION/REPULSION */
 233
 234             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 235             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 236             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 237             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 238             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 239
 240             /* Update potential sum for this i atom from the interaction with this j atom. */
 241             velecsum         = _mm256_add_pd(velecsum,velec);
 242             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 243
 244             fscal            = _mm256_add_pd(felec,fvdw);
 245
 246             /* Calculate temporary vectorial force */
 247             tx               = _mm256_mul_pd(fscal,dx00);
 248             ty               = _mm256_mul_pd(fscal,dy00);
 249             tz               = _mm256_mul_pd(fscal,dz00);
 250
 251             /* Update vectorial force */
 252             fix0             = _mm256_add_pd(fix0,tx);
 253             fiy0             = _mm256_add_pd(fiy0,ty);
 254             fiz0             = _mm256_add_pd(fiz0,tz);
 255
 256             fjptrA             = f+j_coord_offsetA;
 257             fjptrB             = f+j_coord_offsetB;
 258             fjptrC             = f+j_coord_offsetC;
 259             fjptrD             = f+j_coord_offsetD;
 260             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 261
 262             /* Inner loop uses 53 flops */
 263         }
 264
 265         if(jidx<j_index_end)
 266         {
 267
 268             /* Get j neighbor index, and coordinate index */
 269             jnrlistA         = jjnr[jidx];
 270             jnrlistB         = jjnr[jidx+1];
 271             jnrlistC         = jjnr[jidx+2];
 272             jnrlistD         = jjnr[jidx+3];
 273             /* Sign of each element will be negative for non-real atoms.
 274              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 275              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 276              */
 277             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 278
 279             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 280             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 281             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 282
 283             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 284             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 285             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 286             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 287             j_coord_offsetA  = DIM*jnrA;
 288             j_coord_offsetB  = DIM*jnrB;
 289             j_coord_offsetC  = DIM*jnrC;
 290             j_coord_offsetD  = DIM*jnrD;
 291
 292             /* load j atom coordinates */
 293             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 294                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 295                                                  &jx0,&jy0,&jz0);
 296
 297             /* Calculate displacement vector */
 298             dx00             = _mm256_sub_pd(ix0,jx0);
 299             dy00             = _mm256_sub_pd(iy0,jy0);
 300             dz00             = _mm256_sub_pd(iz0,jz0);
 301
 302             /* Calculate squared distance and things based on it */
 303             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 304
 305             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 306
 307             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 308
 309             /* Load parameters for j particles */
 310             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 311                                                                  charge+jnrC+0,charge+jnrD+0);
 312             vdwjidx0A        = 2*vdwtype[jnrA+0];
 313             vdwjidx0B        = 2*vdwtype[jnrB+0];
 314             vdwjidx0C        = 2*vdwtype[jnrC+0];
 315             vdwjidx0D        = 2*vdwtype[jnrD+0];
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             r00              = _mm256_mul_pd(rsq00,rinv00);
 322             r00              = _mm256_andnot_pd(dummy_mask,r00);
 323
 324             /* Compute parameters for interactions between i and j atoms */
 325             qq00             = _mm256_mul_pd(iq0,jq0);
 326             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 327                                             vdwioffsetptr0+vdwjidx0B,
 328                                             vdwioffsetptr0+vdwjidx0C,
 329                                             vdwioffsetptr0+vdwjidx0D,
 330                                             &c6_00,&c12_00);
 331
 332             /* EWALD ELECTROSTATICS */
 333
 334             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 335             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 336             ewitab           = _mm256_cvttpd_epi32(ewrt);
 337             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 338             ewitab           = _mm_slli_epi32(ewitab,2);
 339             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 340             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 341             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 342             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 343             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 344             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 345             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 346             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 347             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 348
 349             /* LENNARD-JONES DISPERSION/REPULSION */
 350
 351             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 352             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 353             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 354             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 355             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 356
 357             /* Update potential sum for this i atom from the interaction with this j atom. */
 358             velec            = _mm256_andnot_pd(dummy_mask,velec);
 359             velecsum         = _mm256_add_pd(velecsum,velec);
 360             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 361             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 362
 363             fscal            = _mm256_add_pd(felec,fvdw);
 364
 365             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 366
 367             /* Calculate temporary vectorial force */
 368             tx               = _mm256_mul_pd(fscal,dx00);
 369             ty               = _mm256_mul_pd(fscal,dy00);
 370             tz               = _mm256_mul_pd(fscal,dz00);
 371
 372             /* Update vectorial force */
 373             fix0             = _mm256_add_pd(fix0,tx);
 374             fiy0             = _mm256_add_pd(fiy0,ty);
 375             fiz0             = _mm256_add_pd(fiz0,tz);
 376
 377             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 378             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 379             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 380             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 381             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 382
 383             /* Inner loop uses 54 flops */
 384         }
 385
 386         /* End of innermost loop */
 387
 388         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 389                                                  f+i_coord_offset,fshift+i_shift_offset);
 390
 391         ggid                        = gid[iidx];
 392         /* Update potential energies */
 393         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 394         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 395
 396         /* Increment number of inner iterations */
 397         inneriter                  += j_index_end - j_index_start;
 398
 399         /* Outer loop uses 9 flops */
 400     }
 401
 402     /* Increment number of outer iterations */
 403     outeriter        += nri;
 404
 405     /* Update outer/inner flops */
 406
 407     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 408 }
 409 /*
 410  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 411  * Electrostatics interaction: Ewald
 412  * VdW interaction:            LennardJones
 413  * Geometry:                   Particle-Particle
 414  * Calculate force/pot:        Force
 415  */
 416 void
 417 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 418                     (t_nblist * gmx_restrict                nlist,
 419                      rvec * gmx_restrict                    xx,
 420                      rvec * gmx_restrict                    ff,
 421                      t_forcerec * gmx_restrict              fr,
 422                      t_mdatoms * gmx_restrict               mdatoms,
 423                      nb_kernel_data_t * gmx_restrict        kernel_data,
 424                      t_nrnb * gmx_restrict                  nrnb)
 425 {
 426     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 427      * just 0 for non-waters.
 428      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 429      * jnr indices corresponding to data put in the four positions in the SIMD register.
 430      */
 431     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 432     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 433     int              jnrA,jnrB,jnrC,jnrD;
 434     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 435     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 436     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 437     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 438     real             rcutoff_scalar;
 439     real             *shiftvec,*fshift,*x,*f;
 440     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 441     real             scratch[4*DIM];
 442     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 443     real *           vdwioffsetptr0;
 444     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 445     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 446     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 447     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 448     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 449     real             *charge;
 450     int              nvdwtype;
 451     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 452     int              *vdwtype;
 453     real             *vdwparam;
 454     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 455     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 456     __m128i          ewitab;
 457     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 458     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 459     real             *ewtab;
 460     __m256d          dummy_mask,cutoff_mask;
 461     __m128           tmpmask0,tmpmask1;
 462     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 463     __m256d          one     = _mm256_set1_pd(1.0);
 464     __m256d          two     = _mm256_set1_pd(2.0);
 465     x                = xx[0];
 466     f                = ff[0];
 467
 468     nri              = nlist->nri;
 469     iinr             = nlist->iinr;
 470     jindex           = nlist->jindex;
 471     jjnr             = nlist->jjnr;
 472     shiftidx         = nlist->shift;
 473     gid              = nlist->gid;
 474     shiftvec         = fr->shift_vec[0];
 475     fshift           = fr->fshift[0];
 476     facel            = _mm256_set1_pd(fr->epsfac);
 477     charge           = mdatoms->chargeA;
 478     nvdwtype         = fr->ntype;
 479     vdwparam         = fr->nbfp;
 480     vdwtype          = mdatoms->typeA;
 481
 482     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 483     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 484     beta2            = _mm256_mul_pd(beta,beta);
 485     beta3            = _mm256_mul_pd(beta,beta2);
 486
 487     ewtab            = fr->ic->tabq_coul_F;
 488     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 489     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 490
 491     /* Avoid stupid compiler warnings */
 492     jnrA = jnrB = jnrC = jnrD = 0;
 493     j_coord_offsetA = 0;
 494     j_coord_offsetB = 0;
 495     j_coord_offsetC = 0;
 496     j_coord_offsetD = 0;
 497
 498     outeriter        = 0;
 499     inneriter        = 0;
 500
 501     for(iidx=0;iidx<4*DIM;iidx++)
 502     {
 503         scratch[iidx] = 0.0;
 504     }
 505
 506     /* Start outer loop over neighborlists */
 507     for(iidx=0; iidx<nri; iidx++)
 508     {
 509         /* Load shift vector for this list */
 510         i_shift_offset   = DIM*shiftidx[iidx];
 511
 512         /* Load limits for loop over neighbors */
 513         j_index_start    = jindex[iidx];
 514         j_index_end      = jindex[iidx+1];
 515
 516         /* Get outer coordinate index */
 517         inr              = iinr[iidx];
 518         i_coord_offset   = DIM*inr;
 519
 520         /* Load i particle coords and add shift vector */
 521         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 522
 523         fix0             = _mm256_setzero_pd();
 524         fiy0             = _mm256_setzero_pd();
 525         fiz0             = _mm256_setzero_pd();
 526
 527         /* Load parameters for i particles */
 528         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 529         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 530
 531         /* Start inner kernel loop */
 532         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 533         {
 534
 535             /* Get j neighbor index, and coordinate index */
 536             jnrA             = jjnr[jidx];
 537             jnrB             = jjnr[jidx+1];
 538             jnrC             = jjnr[jidx+2];
 539             jnrD             = jjnr[jidx+3];
 540             j_coord_offsetA  = DIM*jnrA;
 541             j_coord_offsetB  = DIM*jnrB;
 542             j_coord_offsetC  = DIM*jnrC;
 543             j_coord_offsetD  = DIM*jnrD;
 544
 545             /* load j atom coordinates */
 546             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 547                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 548                                                  &jx0,&jy0,&jz0);
 549
 550             /* Calculate displacement vector */
 551             dx00             = _mm256_sub_pd(ix0,jx0);
 552             dy00             = _mm256_sub_pd(iy0,jy0);
 553             dz00             = _mm256_sub_pd(iz0,jz0);
 554
 555             /* Calculate squared distance and things based on it */
 556             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 557
 558             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 559
 560             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 561
 562             /* Load parameters for j particles */
 563             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 564                                                                  charge+jnrC+0,charge+jnrD+0);
 565             vdwjidx0A        = 2*vdwtype[jnrA+0];
 566             vdwjidx0B        = 2*vdwtype[jnrB+0];
 567             vdwjidx0C        = 2*vdwtype[jnrC+0];
 568             vdwjidx0D        = 2*vdwtype[jnrD+0];
 569
 570             /**************************
 571              * CALCULATE INTERACTIONS *
 572              **************************/
 573
 574             r00              = _mm256_mul_pd(rsq00,rinv00);
 575
 576             /* Compute parameters for interactions between i and j atoms */
 577             qq00             = _mm256_mul_pd(iq0,jq0);
 578             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 579                                             vdwioffsetptr0+vdwjidx0B,
 580                                             vdwioffsetptr0+vdwjidx0C,
 581                                             vdwioffsetptr0+vdwjidx0D,
 582                                             &c6_00,&c12_00);
 583
 584             /* EWALD ELECTROSTATICS */
 585
 586             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 587             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 588             ewitab           = _mm256_cvttpd_epi32(ewrt);
 589             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 590             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 591                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 592                                             &ewtabF,&ewtabFn);
 593             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 594             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 595
 596             /* LENNARD-JONES DISPERSION/REPULSION */
 597
 598             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 599             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 600
 601             fscal            = _mm256_add_pd(felec,fvdw);
 602
 603             /* Calculate temporary vectorial force */
 604             tx               = _mm256_mul_pd(fscal,dx00);
 605             ty               = _mm256_mul_pd(fscal,dy00);
 606             tz               = _mm256_mul_pd(fscal,dz00);
 607
 608             /* Update vectorial force */
 609             fix0             = _mm256_add_pd(fix0,tx);
 610             fiy0             = _mm256_add_pd(fiy0,ty);
 611             fiz0             = _mm256_add_pd(fiz0,tz);
 612
 613             fjptrA             = f+j_coord_offsetA;
 614             fjptrB             = f+j_coord_offsetB;
 615             fjptrC             = f+j_coord_offsetC;
 616             fjptrD             = f+j_coord_offsetD;
 617             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 618
 619             /* Inner loop uses 43 flops */
 620         }
 621
 622         if(jidx<j_index_end)
 623         {
 624
 625             /* Get j neighbor index, and coordinate index */
 626             jnrlistA         = jjnr[jidx];
 627             jnrlistB         = jjnr[jidx+1];
 628             jnrlistC         = jjnr[jidx+2];
 629             jnrlistD         = jjnr[jidx+3];
 630             /* Sign of each element will be negative for non-real atoms.
 631              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 632              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 633              */
 634             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 635
 636             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 637             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 638             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 639
 640             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 641             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 642             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 643             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 644             j_coord_offsetA  = DIM*jnrA;
 645             j_coord_offsetB  = DIM*jnrB;
 646             j_coord_offsetC  = DIM*jnrC;
 647             j_coord_offsetD  = DIM*jnrD;
 648
 649             /* load j atom coordinates */
 650             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 651                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 652                                                  &jx0,&jy0,&jz0);
 653
 654             /* Calculate displacement vector */
 655             dx00             = _mm256_sub_pd(ix0,jx0);
 656             dy00             = _mm256_sub_pd(iy0,jy0);
 657             dz00             = _mm256_sub_pd(iz0,jz0);
 658
 659             /* Calculate squared distance and things based on it */
 660             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 661
 662             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 663
 664             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 665
 666             /* Load parameters for j particles */
 667             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 668                                                                  charge+jnrC+0,charge+jnrD+0);
 669             vdwjidx0A        = 2*vdwtype[jnrA+0];
 670             vdwjidx0B        = 2*vdwtype[jnrB+0];
 671             vdwjidx0C        = 2*vdwtype[jnrC+0];
 672             vdwjidx0D        = 2*vdwtype[jnrD+0];
 673
 674             /**************************
 675              * CALCULATE INTERACTIONS *
 676              **************************/
 677
 678             r00              = _mm256_mul_pd(rsq00,rinv00);
 679             r00              = _mm256_andnot_pd(dummy_mask,r00);
 680
 681             /* Compute parameters for interactions between i and j atoms */
 682             qq00             = _mm256_mul_pd(iq0,jq0);
 683             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 684                                             vdwioffsetptr0+vdwjidx0B,
 685                                             vdwioffsetptr0+vdwjidx0C,
 686                                             vdwioffsetptr0+vdwjidx0D,
 687                                             &c6_00,&c12_00);
 688
 689             /* EWALD ELECTROSTATICS */
 690
 691             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 692             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 693             ewitab           = _mm256_cvttpd_epi32(ewrt);
 694             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 695             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 696                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 697                                             &ewtabF,&ewtabFn);
 698             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 699             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 700
 701             /* LENNARD-JONES DISPERSION/REPULSION */
 702
 703             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 704             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 705
 706             fscal            = _mm256_add_pd(felec,fvdw);
 707
 708             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 709
 710             /* Calculate temporary vectorial force */
 711             tx               = _mm256_mul_pd(fscal,dx00);
 712             ty               = _mm256_mul_pd(fscal,dy00);
 713             tz               = _mm256_mul_pd(fscal,dz00);
 714
 715             /* Update vectorial force */
 716             fix0             = _mm256_add_pd(fix0,tx);
 717             fiy0             = _mm256_add_pd(fiy0,ty);
 718             fiz0             = _mm256_add_pd(fiz0,tz);
 719
 720             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 721             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 722             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 723             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 724             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 725
 726             /* Inner loop uses 44 flops */
 727         }
 728
 729         /* End of innermost loop */
 730
 731         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 732                                                  f+i_coord_offset,fshift+i_shift_offset);
 733
 734         /* Increment number of inner iterations */
 735         inneriter                  += j_index_end - j_index_start;
 736
 737         /* Outer loop uses 7 flops */
 738     }
 739
 740     /* Increment number of outer iterations */
 741     outeriter        += nri;
 742
 743     /* Update outer/inner flops */
 744
 745     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*44);
 746 }