src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecGB_VdwNone_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_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwNone_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     __m128i          gbitab;
  78     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
  79     __m256d          minushalf = _mm256_set1_pd(-0.5);
  80     real             *invsqrta,*dvda,*gbtab;
  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     facel            = _mm256_set1_pd(fr->epsfac);
 102     charge           = mdatoms->chargeA;
 103
 104     invsqrta         = fr->invsqrta;
 105     dvda             = fr->dvda;
 106     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 107     gbtab            = fr->gbtab.data;
 108     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 109
 110     /* Avoid stupid compiler warnings */
 111     jnrA = jnrB = jnrC = jnrD = 0;
 112     j_coord_offsetA = 0;
 113     j_coord_offsetB = 0;
 114     j_coord_offsetC = 0;
 115     j_coord_offsetD = 0;
 116
 117     outeriter        = 0;
 118     inneriter        = 0;
 119
 120     for(iidx=0;iidx<4*DIM;iidx++)
 121     {
 122         scratch[iidx] = 0.0;
 123     }
 124
 125     /* Start outer loop over neighborlists */
 126     for(iidx=0; iidx<nri; iidx++)
 127     {
 128         /* Load shift vector for this list */
 129         i_shift_offset   = DIM*shiftidx[iidx];
 130
 131         /* Load limits for loop over neighbors */
 132         j_index_start    = jindex[iidx];
 133         j_index_end      = jindex[iidx+1];
 134
 135         /* Get outer coordinate index */
 136         inr              = iinr[iidx];
 137         i_coord_offset   = DIM*inr;
 138
 139         /* Load i particle coords and add shift vector */
 140         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 141
 142         fix0             = _mm256_setzero_pd();
 143         fiy0             = _mm256_setzero_pd();
 144         fiz0             = _mm256_setzero_pd();
 145
 146         /* Load parameters for i particles */
 147         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 148         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 149
 150         /* Reset potential sums */
 151         velecsum         = _mm256_setzero_pd();
 152         vgbsum           = _mm256_setzero_pd();
 153         dvdasum          = _mm256_setzero_pd();
 154
 155         /* Start inner kernel loop */
 156         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 157         {
 158
 159             /* Get j neighbor index, and coordinate index */
 160             jnrA             = jjnr[jidx];
 161             jnrB             = jjnr[jidx+1];
 162             jnrC             = jjnr[jidx+2];
 163             jnrD             = jjnr[jidx+3];
 164             j_coord_offsetA  = DIM*jnrA;
 165             j_coord_offsetB  = DIM*jnrB;
 166             j_coord_offsetC  = DIM*jnrC;
 167             j_coord_offsetD  = DIM*jnrD;
 168
 169             /* load j atom coordinates */
 170             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 171                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 172                                                  &jx0,&jy0,&jz0);
 173
 174             /* Calculate displacement vector */
 175             dx00             = _mm256_sub_pd(ix0,jx0);
 176             dy00             = _mm256_sub_pd(iy0,jy0);
 177             dz00             = _mm256_sub_pd(iz0,jz0);
 178
 179             /* Calculate squared distance and things based on it */
 180             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 181
 182             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 183
 184             /* Load parameters for j particles */
 185             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 186                                                                  charge+jnrC+0,charge+jnrD+0);
 187             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 188                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 189
 190             /**************************
 191              * CALCULATE INTERACTIONS *
 192              **************************/
 193
 194             r00              = _mm256_mul_pd(rsq00,rinv00);
 195
 196             /* Compute parameters for interactions between i and j atoms */
 197             qq00             = _mm256_mul_pd(iq0,jq0);
 198
 199             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 200             isaprod          = _mm256_mul_pd(isai0,isaj0);
 201             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 202             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 203
 204             /* Calculate generalized born table index - this is a separate table from the normal one,
 205              * but we use the same procedure by multiplying r with scale and truncating to integer.
 206              */
 207             rt               = _mm256_mul_pd(r00,gbscale);
 208             gbitab           = _mm256_cvttpd_epi32(rt);
 209             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 210             gbitab           = _mm_slli_epi32(gbitab,2);
 211             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 212             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 213             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 214             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 215             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 216             Heps             = _mm256_mul_pd(gbeps,H);
 217             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 218             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 219             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 220
 221             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 222             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 223             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 224             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 225             fjptrA           = dvda+jnrA;
 226             fjptrB           = dvda+jnrB;
 227             fjptrC           = dvda+jnrC;
 228             fjptrD           = dvda+jnrD;
 229             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 230                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 231             velec            = _mm256_mul_pd(qq00,rinv00);
 232             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 233
 234             /* Update potential sum for this i atom from the interaction with this j atom. */
 235             velecsum         = _mm256_add_pd(velecsum,velec);
 236             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 237
 238             fscal            = felec;
 239
 240             /* Calculate temporary vectorial force */
 241             tx               = _mm256_mul_pd(fscal,dx00);
 242             ty               = _mm256_mul_pd(fscal,dy00);
 243             tz               = _mm256_mul_pd(fscal,dz00);
 244
 245             /* Update vectorial force */
 246             fix0             = _mm256_add_pd(fix0,tx);
 247             fiy0             = _mm256_add_pd(fiy0,ty);
 248             fiz0             = _mm256_add_pd(fiz0,tz);
 249
 250             fjptrA             = f+j_coord_offsetA;
 251             fjptrB             = f+j_coord_offsetB;
 252             fjptrC             = f+j_coord_offsetC;
 253             fjptrD             = f+j_coord_offsetD;
 254             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 255
 256             /* Inner loop uses 57 flops */
 257         }
 258
 259         if(jidx<j_index_end)
 260         {
 261
 262             /* Get j neighbor index, and coordinate index */
 263             jnrlistA         = jjnr[jidx];
 264             jnrlistB         = jjnr[jidx+1];
 265             jnrlistC         = jjnr[jidx+2];
 266             jnrlistD         = jjnr[jidx+3];
 267             /* Sign of each element will be negative for non-real atoms.
 268              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 269              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 270              */
 271             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 272
 273             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 274             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 275             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 276
 277             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 278             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 279             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 280             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 281             j_coord_offsetA  = DIM*jnrA;
 282             j_coord_offsetB  = DIM*jnrB;
 283             j_coord_offsetC  = DIM*jnrC;
 284             j_coord_offsetD  = DIM*jnrD;
 285
 286             /* load j atom coordinates */
 287             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 288                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 289                                                  &jx0,&jy0,&jz0);
 290
 291             /* Calculate displacement vector */
 292             dx00             = _mm256_sub_pd(ix0,jx0);
 293             dy00             = _mm256_sub_pd(iy0,jy0);
 294             dz00             = _mm256_sub_pd(iz0,jz0);
 295
 296             /* Calculate squared distance and things based on it */
 297             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 298
 299             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 300
 301             /* Load parameters for j particles */
 302             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 303                                                                  charge+jnrC+0,charge+jnrD+0);
 304             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 305                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             r00              = _mm256_mul_pd(rsq00,rinv00);
 312             r00              = _mm256_andnot_pd(dummy_mask,r00);
 313
 314             /* Compute parameters for interactions between i and j atoms */
 315             qq00             = _mm256_mul_pd(iq0,jq0);
 316
 317             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 318             isaprod          = _mm256_mul_pd(isai0,isaj0);
 319             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 320             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 321
 322             /* Calculate generalized born table index - this is a separate table from the normal one,
 323              * but we use the same procedure by multiplying r with scale and truncating to integer.
 324              */
 325             rt               = _mm256_mul_pd(r00,gbscale);
 326             gbitab           = _mm256_cvttpd_epi32(rt);
 327             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 328             gbitab           = _mm_slli_epi32(gbitab,2);
 329             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 330             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 331             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 332             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 333             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 334             Heps             = _mm256_mul_pd(gbeps,H);
 335             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 336             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 337             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 338
 339             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 340             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 341             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 342             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 343             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 344             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 345             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 346             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 347             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 348             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 349             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 350                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 351             velec            = _mm256_mul_pd(qq00,rinv00);
 352             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 353
 354             /* Update potential sum for this i atom from the interaction with this j atom. */
 355             velec            = _mm256_andnot_pd(dummy_mask,velec);
 356             velecsum         = _mm256_add_pd(velecsum,velec);
 357             vgb              = _mm256_andnot_pd(dummy_mask,vgb);
 358             vgbsum           = _mm256_add_pd(vgbsum,vgb);
 359
 360             fscal            = felec;
 361
 362             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 363
 364             /* Calculate temporary vectorial force */
 365             tx               = _mm256_mul_pd(fscal,dx00);
 366             ty               = _mm256_mul_pd(fscal,dy00);
 367             tz               = _mm256_mul_pd(fscal,dz00);
 368
 369             /* Update vectorial force */
 370             fix0             = _mm256_add_pd(fix0,tx);
 371             fiy0             = _mm256_add_pd(fiy0,ty);
 372             fiz0             = _mm256_add_pd(fiz0,tz);
 373
 374             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 375             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 376             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 377             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 378             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 379
 380             /* Inner loop uses 58 flops */
 381         }
 382
 383         /* End of innermost loop */
 384
 385         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 386                                                  f+i_coord_offset,fshift+i_shift_offset);
 387
 388         ggid                        = gid[iidx];
 389         /* Update potential energies */
 390         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 391         gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 392         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 393         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 394
 395         /* Increment number of inner iterations */
 396         inneriter                  += j_index_end - j_index_start;
 397
 398         /* Outer loop uses 9 flops */
 399     }
 400
 401     /* Increment number of outer iterations */
 402     outeriter        += nri;
 403
 404     /* Update outer/inner flops */
 405
 406     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
 407 }
 408 /*
 409  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
 410  * Electrostatics interaction: GeneralizedBorn
 411  * VdW interaction:            None
 412  * Geometry:                   Particle-Particle
 413  * Calculate force/pot:        Force
 414  */
 415 void
 416 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
 417                     (t_nblist * gmx_restrict                nlist,
 418                      rvec * gmx_restrict                    xx,
 419                      rvec * gmx_restrict                    ff,
 420                      t_forcerec * gmx_restrict              fr,
 421                      t_mdatoms * gmx_restrict               mdatoms,
 422                      nb_kernel_data_t * gmx_restrict        kernel_data,
 423                      t_nrnb * gmx_restrict                  nrnb)
 424 {
 425     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 426      * just 0 for non-waters.
 427      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 428      * jnr indices corresponding to data put in the four positions in the SIMD register.
 429      */
 430     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 431     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 432     int              jnrA,jnrB,jnrC,jnrD;
 433     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 434     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 435     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 436     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 437     real             rcutoff_scalar;
 438     real             *shiftvec,*fshift,*x,*f;
 439     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 440     real             scratch[4*DIM];
 441     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 442     real *           vdwioffsetptr0;
 443     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 444     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 445     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 446     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 447     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 448     real             *charge;
 449     __m128i          gbitab;
 450     __m256d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
 451     __m256d          minushalf = _mm256_set1_pd(-0.5);
 452     real             *invsqrta,*dvda,*gbtab;
 453     __m128i          vfitab;
 454     __m128i          ifour       = _mm_set1_epi32(4);
 455     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 456     real             *vftab;
 457     __m256d          dummy_mask,cutoff_mask;
 458     __m128           tmpmask0,tmpmask1;
 459     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 460     __m256d          one     = _mm256_set1_pd(1.0);
 461     __m256d          two     = _mm256_set1_pd(2.0);
 462     x                = xx[0];
 463     f                = ff[0];
 464
 465     nri              = nlist->nri;
 466     iinr             = nlist->iinr;
 467     jindex           = nlist->jindex;
 468     jjnr             = nlist->jjnr;
 469     shiftidx         = nlist->shift;
 470     gid              = nlist->gid;
 471     shiftvec         = fr->shift_vec[0];
 472     fshift           = fr->fshift[0];
 473     facel            = _mm256_set1_pd(fr->epsfac);
 474     charge           = mdatoms->chargeA;
 475
 476     invsqrta         = fr->invsqrta;
 477     dvda             = fr->dvda;
 478     gbtabscale       = _mm256_set1_pd(fr->gbtab.scale);
 479     gbtab            = fr->gbtab.data;
 480     gbinvepsdiff     = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 481
 482     /* Avoid stupid compiler warnings */
 483     jnrA = jnrB = jnrC = jnrD = 0;
 484     j_coord_offsetA = 0;
 485     j_coord_offsetB = 0;
 486     j_coord_offsetC = 0;
 487     j_coord_offsetD = 0;
 488
 489     outeriter        = 0;
 490     inneriter        = 0;
 491
 492     for(iidx=0;iidx<4*DIM;iidx++)
 493     {
 494         scratch[iidx] = 0.0;
 495     }
 496
 497     /* Start outer loop over neighborlists */
 498     for(iidx=0; iidx<nri; iidx++)
 499     {
 500         /* Load shift vector for this list */
 501         i_shift_offset   = DIM*shiftidx[iidx];
 502
 503         /* Load limits for loop over neighbors */
 504         j_index_start    = jindex[iidx];
 505         j_index_end      = jindex[iidx+1];
 506
 507         /* Get outer coordinate index */
 508         inr              = iinr[iidx];
 509         i_coord_offset   = DIM*inr;
 510
 511         /* Load i particle coords and add shift vector */
 512         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 513
 514         fix0             = _mm256_setzero_pd();
 515         fiy0             = _mm256_setzero_pd();
 516         fiz0             = _mm256_setzero_pd();
 517
 518         /* Load parameters for i particles */
 519         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 520         isai0            = _mm256_set1_pd(invsqrta[inr+0]);
 521
 522         dvdasum          = _mm256_setzero_pd();
 523
 524         /* Start inner kernel loop */
 525         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 526         {
 527
 528             /* Get j neighbor index, and coordinate index */
 529             jnrA             = jjnr[jidx];
 530             jnrB             = jjnr[jidx+1];
 531             jnrC             = jjnr[jidx+2];
 532             jnrD             = jjnr[jidx+3];
 533             j_coord_offsetA  = DIM*jnrA;
 534             j_coord_offsetB  = DIM*jnrB;
 535             j_coord_offsetC  = DIM*jnrC;
 536             j_coord_offsetD  = DIM*jnrD;
 537
 538             /* load j atom coordinates */
 539             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 540                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 541                                                  &jx0,&jy0,&jz0);
 542
 543             /* Calculate displacement vector */
 544             dx00             = _mm256_sub_pd(ix0,jx0);
 545             dy00             = _mm256_sub_pd(iy0,jy0);
 546             dz00             = _mm256_sub_pd(iz0,jz0);
 547
 548             /* Calculate squared distance and things based on it */
 549             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 550
 551             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 552
 553             /* Load parameters for j particles */
 554             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 555                                                                  charge+jnrC+0,charge+jnrD+0);
 556             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 557                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 558
 559             /**************************
 560              * CALCULATE INTERACTIONS *
 561              **************************/
 562
 563             r00              = _mm256_mul_pd(rsq00,rinv00);
 564
 565             /* Compute parameters for interactions between i and j atoms */
 566             qq00             = _mm256_mul_pd(iq0,jq0);
 567
 568             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 569             isaprod          = _mm256_mul_pd(isai0,isaj0);
 570             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 571             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 572
 573             /* Calculate generalized born table index - this is a separate table from the normal one,
 574              * but we use the same procedure by multiplying r with scale and truncating to integer.
 575              */
 576             rt               = _mm256_mul_pd(r00,gbscale);
 577             gbitab           = _mm256_cvttpd_epi32(rt);
 578             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 579             gbitab           = _mm_slli_epi32(gbitab,2);
 580             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 581             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 582             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 583             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 584             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 585             Heps             = _mm256_mul_pd(gbeps,H);
 586             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 587             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 588             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 589
 590             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 591             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 592             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 593             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 594             fjptrA           = dvda+jnrA;
 595             fjptrB           = dvda+jnrB;
 596             fjptrC           = dvda+jnrC;
 597             fjptrD           = dvda+jnrD;
 598             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 599                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 600             velec            = _mm256_mul_pd(qq00,rinv00);
 601             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 602
 603             fscal            = felec;
 604
 605             /* Calculate temporary vectorial force */
 606             tx               = _mm256_mul_pd(fscal,dx00);
 607             ty               = _mm256_mul_pd(fscal,dy00);
 608             tz               = _mm256_mul_pd(fscal,dz00);
 609
 610             /* Update vectorial force */
 611             fix0             = _mm256_add_pd(fix0,tx);
 612             fiy0             = _mm256_add_pd(fiy0,ty);
 613             fiz0             = _mm256_add_pd(fiz0,tz);
 614
 615             fjptrA             = f+j_coord_offsetA;
 616             fjptrB             = f+j_coord_offsetB;
 617             fjptrC             = f+j_coord_offsetC;
 618             fjptrD             = f+j_coord_offsetD;
 619             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 620
 621             /* Inner loop uses 55 flops */
 622         }
 623
 624         if(jidx<j_index_end)
 625         {
 626
 627             /* Get j neighbor index, and coordinate index */
 628             jnrlistA         = jjnr[jidx];
 629             jnrlistB         = jjnr[jidx+1];
 630             jnrlistC         = jjnr[jidx+2];
 631             jnrlistD         = jjnr[jidx+3];
 632             /* Sign of each element will be negative for non-real atoms.
 633              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 634              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 635              */
 636             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 637
 638             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 639             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 640             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 641
 642             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 643             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 644             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 645             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 646             j_coord_offsetA  = DIM*jnrA;
 647             j_coord_offsetB  = DIM*jnrB;
 648             j_coord_offsetC  = DIM*jnrC;
 649             j_coord_offsetD  = DIM*jnrD;
 650
 651             /* load j atom coordinates */
 652             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 653                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 654                                                  &jx0,&jy0,&jz0);
 655
 656             /* Calculate displacement vector */
 657             dx00             = _mm256_sub_pd(ix0,jx0);
 658             dy00             = _mm256_sub_pd(iy0,jy0);
 659             dz00             = _mm256_sub_pd(iz0,jz0);
 660
 661             /* Calculate squared distance and things based on it */
 662             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 663
 664             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 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             isaj0            = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
 670                                                                  invsqrta+jnrC+0,invsqrta+jnrD+0);
 671
 672             /**************************
 673              * CALCULATE INTERACTIONS *
 674              **************************/
 675
 676             r00              = _mm256_mul_pd(rsq00,rinv00);
 677             r00              = _mm256_andnot_pd(dummy_mask,r00);
 678
 679             /* Compute parameters for interactions between i and j atoms */
 680             qq00             = _mm256_mul_pd(iq0,jq0);
 681
 682             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 683             isaprod          = _mm256_mul_pd(isai0,isaj0);
 684             gbqqfactor       = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
 685             gbscale          = _mm256_mul_pd(isaprod,gbtabscale);
 686
 687             /* Calculate generalized born table index - this is a separate table from the normal one,
 688              * but we use the same procedure by multiplying r with scale and truncating to integer.
 689              */
 690             rt               = _mm256_mul_pd(r00,gbscale);
 691             gbitab           = _mm256_cvttpd_epi32(rt);
 692             gbeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 693             gbitab           = _mm_slli_epi32(gbitab,2);
 694             Y                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 695             F                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 696             G                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
 697             H                = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
 698             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 699             Heps             = _mm256_mul_pd(gbeps,H);
 700             Fp               = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
 701             VV               = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
 702             vgb              = _mm256_mul_pd(gbqqfactor,VV);
 703
 704             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 705             fgb              = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
 706             dvdatmp          = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
 707             dvdatmp          = _mm256_andnot_ps(dummy_mask,dvdatmp);
 708             dvdasum          = _mm256_add_pd(dvdasum,dvdatmp);
 709             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 710             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 711             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 712             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 713             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 714             gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
 715                                                  _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
 716             velec            = _mm256_mul_pd(qq00,rinv00);
 717             felec            = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm256_mul_pd(fscal,dx00);
 725             ty               = _mm256_mul_pd(fscal,dy00);
 726             tz               = _mm256_mul_pd(fscal,dz00);
 727
 728             /* Update vectorial force */
 729             fix0             = _mm256_add_pd(fix0,tx);
 730             fiy0             = _mm256_add_pd(fiy0,ty);
 731             fiz0             = _mm256_add_pd(fiz0,tz);
 732
 733             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 734             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 735             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 736             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 737             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 738
 739             /* Inner loop uses 56 flops */
 740         }
 741
 742         /* End of innermost loop */
 743
 744         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 745                                                  f+i_coord_offset,fshift+i_shift_offset);
 746
 747         dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
 748         gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
 749
 750         /* Increment number of inner iterations */
 751         inneriter                  += j_index_end - j_index_start;
 752
 753         /* Outer loop uses 7 flops */
 754     }
 755
 756     /* Increment number of outer iterations */
 757     outeriter        += nri;
 758
 759     /* Update outer/inner flops */
 760
 761     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);
 762 }