src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecCSTab_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_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: CubicSplineTable
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCSTab_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          vfitab;
  78     __m128i          ifour       = _mm_set1_epi32(4);
  79     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  80     real             *vftab;
  81     __m256d          dummy_mask,cutoff_mask;
  82     __m128           tmpmask0,tmpmask1;
  83     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  84     __m256d          one     = _mm256_set1_pd(1.0);
  85     __m256d          two     = _mm256_set1_pd(2.0);
  86     x                = xx[0];
  87     f                = ff[0];
  88
  89     nri              = nlist->nri;
  90     iinr             = nlist->iinr;
  91     jindex           = nlist->jindex;
  92     jjnr             = nlist->jjnr;
  93     shiftidx         = nlist->shift;
  94     gid              = nlist->gid;
  95     shiftvec         = fr->shift_vec[0];
  96     fshift           = fr->fshift[0];
  97     facel            = _mm256_set1_pd(fr->epsfac);
  98     charge           = mdatoms->chargeA;
  99
 100     vftab            = kernel_data->table_elec->data;
 101     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 102
 103     /* Avoid stupid compiler warnings */
 104     jnrA = jnrB = jnrC = jnrD = 0;
 105     j_coord_offsetA = 0;
 106     j_coord_offsetB = 0;
 107     j_coord_offsetC = 0;
 108     j_coord_offsetD = 0;
 109
 110     outeriter        = 0;
 111     inneriter        = 0;
 112
 113     for(iidx=0;iidx<4*DIM;iidx++)
 114     {
 115         scratch[iidx] = 0.0;
 116     }
 117
 118     /* Start outer loop over neighborlists */
 119     for(iidx=0; iidx<nri; iidx++)
 120     {
 121         /* Load shift vector for this list */
 122         i_shift_offset   = DIM*shiftidx[iidx];
 123
 124         /* Load limits for loop over neighbors */
 125         j_index_start    = jindex[iidx];
 126         j_index_end      = jindex[iidx+1];
 127
 128         /* Get outer coordinate index */
 129         inr              = iinr[iidx];
 130         i_coord_offset   = DIM*inr;
 131
 132         /* Load i particle coords and add shift vector */
 133         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 134
 135         fix0             = _mm256_setzero_pd();
 136         fiy0             = _mm256_setzero_pd();
 137         fiz0             = _mm256_setzero_pd();
 138
 139         /* Load parameters for i particles */
 140         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 141
 142         /* Reset potential sums */
 143         velecsum         = _mm256_setzero_pd();
 144
 145         /* Start inner kernel loop */
 146         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 147         {
 148
 149             /* Get j neighbor index, and coordinate index */
 150             jnrA             = jjnr[jidx];
 151             jnrB             = jjnr[jidx+1];
 152             jnrC             = jjnr[jidx+2];
 153             jnrD             = jjnr[jidx+3];
 154             j_coord_offsetA  = DIM*jnrA;
 155             j_coord_offsetB  = DIM*jnrB;
 156             j_coord_offsetC  = DIM*jnrC;
 157             j_coord_offsetD  = DIM*jnrD;
 158
 159             /* load j atom coordinates */
 160             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 161                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 162                                                  &jx0,&jy0,&jz0);
 163
 164             /* Calculate displacement vector */
 165             dx00             = _mm256_sub_pd(ix0,jx0);
 166             dy00             = _mm256_sub_pd(iy0,jy0);
 167             dz00             = _mm256_sub_pd(iz0,jz0);
 168
 169             /* Calculate squared distance and things based on it */
 170             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 171
 172             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 173
 174             /* Load parameters for j particles */
 175             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 176                                                                  charge+jnrC+0,charge+jnrD+0);
 177
 178             /**************************
 179              * CALCULATE INTERACTIONS *
 180              **************************/
 181
 182             r00              = _mm256_mul_pd(rsq00,rinv00);
 183
 184             /* Compute parameters for interactions between i and j atoms */
 185             qq00             = _mm256_mul_pd(iq0,jq0);
 186
 187             /* Calculate table index by multiplying r with table scale and truncate to integer */
 188             rt               = _mm256_mul_pd(r00,vftabscale);
 189             vfitab           = _mm256_cvttpd_epi32(rt);
 190             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 191             vfitab           = _mm_slli_epi32(vfitab,2);
 192
 193             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 194             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 195             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 196             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 197             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 198             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 199             Heps             = _mm256_mul_pd(vfeps,H);
 200             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 201             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 202             velec            = _mm256_mul_pd(qq00,VV);
 203             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 204             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 205
 206             /* Update potential sum for this i atom from the interaction with this j atom. */
 207             velecsum         = _mm256_add_pd(velecsum,velec);
 208
 209             fscal            = felec;
 210
 211             /* Calculate temporary vectorial force */
 212             tx               = _mm256_mul_pd(fscal,dx00);
 213             ty               = _mm256_mul_pd(fscal,dy00);
 214             tz               = _mm256_mul_pd(fscal,dz00);
 215
 216             /* Update vectorial force */
 217             fix0             = _mm256_add_pd(fix0,tx);
 218             fiy0             = _mm256_add_pd(fiy0,ty);
 219             fiz0             = _mm256_add_pd(fiz0,tz);
 220
 221             fjptrA             = f+j_coord_offsetA;
 222             fjptrB             = f+j_coord_offsetB;
 223             fjptrC             = f+j_coord_offsetC;
 224             fjptrD             = f+j_coord_offsetD;
 225             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 226
 227             /* Inner loop uses 43 flops */
 228         }
 229
 230         if(jidx<j_index_end)
 231         {
 232
 233             /* Get j neighbor index, and coordinate index */
 234             jnrlistA         = jjnr[jidx];
 235             jnrlistB         = jjnr[jidx+1];
 236             jnrlistC         = jjnr[jidx+2];
 237             jnrlistD         = jjnr[jidx+3];
 238             /* Sign of each element will be negative for non-real atoms.
 239              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 240              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 241              */
 242             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 243
 244             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 245             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 246             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 247
 248             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 249             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 250             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 251             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 252             j_coord_offsetA  = DIM*jnrA;
 253             j_coord_offsetB  = DIM*jnrB;
 254             j_coord_offsetC  = DIM*jnrC;
 255             j_coord_offsetD  = DIM*jnrD;
 256
 257             /* load j atom coordinates */
 258             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 259                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 260                                                  &jx0,&jy0,&jz0);
 261
 262             /* Calculate displacement vector */
 263             dx00             = _mm256_sub_pd(ix0,jx0);
 264             dy00             = _mm256_sub_pd(iy0,jy0);
 265             dz00             = _mm256_sub_pd(iz0,jz0);
 266
 267             /* Calculate squared distance and things based on it */
 268             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 269
 270             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 271
 272             /* Load parameters for j particles */
 273             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 274                                                                  charge+jnrC+0,charge+jnrD+0);
 275
 276             /**************************
 277              * CALCULATE INTERACTIONS *
 278              **************************/
 279
 280             r00              = _mm256_mul_pd(rsq00,rinv00);
 281             r00              = _mm256_andnot_pd(dummy_mask,r00);
 282
 283             /* Compute parameters for interactions between i and j atoms */
 284             qq00             = _mm256_mul_pd(iq0,jq0);
 285
 286             /* Calculate table index by multiplying r with table scale and truncate to integer */
 287             rt               = _mm256_mul_pd(r00,vftabscale);
 288             vfitab           = _mm256_cvttpd_epi32(rt);
 289             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 290             vfitab           = _mm_slli_epi32(vfitab,2);
 291
 292             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 293             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 294             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 295             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 296             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 297             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 298             Heps             = _mm256_mul_pd(vfeps,H);
 299             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 300             VV               = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
 301             velec            = _mm256_mul_pd(qq00,VV);
 302             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 303             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velec            = _mm256_andnot_pd(dummy_mask,velec);
 307             velecsum         = _mm256_add_pd(velecsum,velec);
 308
 309             fscal            = felec;
 310
 311             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 312
 313             /* Calculate temporary vectorial force */
 314             tx               = _mm256_mul_pd(fscal,dx00);
 315             ty               = _mm256_mul_pd(fscal,dy00);
 316             tz               = _mm256_mul_pd(fscal,dz00);
 317
 318             /* Update vectorial force */
 319             fix0             = _mm256_add_pd(fix0,tx);
 320             fiy0             = _mm256_add_pd(fiy0,ty);
 321             fiz0             = _mm256_add_pd(fiz0,tz);
 322
 323             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 324             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 325             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 326             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 327             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 328
 329             /* Inner loop uses 44 flops */
 330         }
 331
 332         /* End of innermost loop */
 333
 334         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 335                                                  f+i_coord_offset,fshift+i_shift_offset);
 336
 337         ggid                        = gid[iidx];
 338         /* Update potential energies */
 339         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 340
 341         /* Increment number of inner iterations */
 342         inneriter                  += j_index_end - j_index_start;
 343
 344         /* Outer loop uses 8 flops */
 345     }
 346
 347     /* Increment number of outer iterations */
 348     outeriter        += nri;
 349
 350     /* Update outer/inner flops */
 351
 352     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*44);
 353 }
 354 /*
 355  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_double
 356  * Electrostatics interaction: CubicSplineTable
 357  * VdW interaction:            None
 358  * Geometry:                   Particle-Particle
 359  * Calculate force/pot:        Force
 360  */
 361 void
 362 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_double
 363                     (t_nblist * gmx_restrict                nlist,
 364                      rvec * gmx_restrict                    xx,
 365                      rvec * gmx_restrict                    ff,
 366                      t_forcerec * gmx_restrict              fr,
 367                      t_mdatoms * gmx_restrict               mdatoms,
 368                      nb_kernel_data_t * gmx_restrict        kernel_data,
 369                      t_nrnb * gmx_restrict                  nrnb)
 370 {
 371     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 372      * just 0 for non-waters.
 373      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 374      * jnr indices corresponding to data put in the four positions in the SIMD register.
 375      */
 376     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 377     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 378     int              jnrA,jnrB,jnrC,jnrD;
 379     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 380     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 381     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 382     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 383     real             rcutoff_scalar;
 384     real             *shiftvec,*fshift,*x,*f;
 385     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 386     real             scratch[4*DIM];
 387     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 388     real *           vdwioffsetptr0;
 389     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 390     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 391     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 392     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 393     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 394     real             *charge;
 395     __m128i          vfitab;
 396     __m128i          ifour       = _mm_set1_epi32(4);
 397     __m256d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 398     real             *vftab;
 399     __m256d          dummy_mask,cutoff_mask;
 400     __m128           tmpmask0,tmpmask1;
 401     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 402     __m256d          one     = _mm256_set1_pd(1.0);
 403     __m256d          two     = _mm256_set1_pd(2.0);
 404     x                = xx[0];
 405     f                = ff[0];
 406
 407     nri              = nlist->nri;
 408     iinr             = nlist->iinr;
 409     jindex           = nlist->jindex;
 410     jjnr             = nlist->jjnr;
 411     shiftidx         = nlist->shift;
 412     gid              = nlist->gid;
 413     shiftvec         = fr->shift_vec[0];
 414     fshift           = fr->fshift[0];
 415     facel            = _mm256_set1_pd(fr->epsfac);
 416     charge           = mdatoms->chargeA;
 417
 418     vftab            = kernel_data->table_elec->data;
 419     vftabscale       = _mm256_set1_pd(kernel_data->table_elec->scale);
 420
 421     /* Avoid stupid compiler warnings */
 422     jnrA = jnrB = jnrC = jnrD = 0;
 423     j_coord_offsetA = 0;
 424     j_coord_offsetB = 0;
 425     j_coord_offsetC = 0;
 426     j_coord_offsetD = 0;
 427
 428     outeriter        = 0;
 429     inneriter        = 0;
 430
 431     for(iidx=0;iidx<4*DIM;iidx++)
 432     {
 433         scratch[iidx] = 0.0;
 434     }
 435
 436     /* Start outer loop over neighborlists */
 437     for(iidx=0; iidx<nri; iidx++)
 438     {
 439         /* Load shift vector for this list */
 440         i_shift_offset   = DIM*shiftidx[iidx];
 441
 442         /* Load limits for loop over neighbors */
 443         j_index_start    = jindex[iidx];
 444         j_index_end      = jindex[iidx+1];
 445
 446         /* Get outer coordinate index */
 447         inr              = iinr[iidx];
 448         i_coord_offset   = DIM*inr;
 449
 450         /* Load i particle coords and add shift vector */
 451         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 452
 453         fix0             = _mm256_setzero_pd();
 454         fiy0             = _mm256_setzero_pd();
 455         fiz0             = _mm256_setzero_pd();
 456
 457         /* Load parameters for i particles */
 458         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 459
 460         /* Start inner kernel loop */
 461         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 462         {
 463
 464             /* Get j neighbor index, and coordinate index */
 465             jnrA             = jjnr[jidx];
 466             jnrB             = jjnr[jidx+1];
 467             jnrC             = jjnr[jidx+2];
 468             jnrD             = jjnr[jidx+3];
 469             j_coord_offsetA  = DIM*jnrA;
 470             j_coord_offsetB  = DIM*jnrB;
 471             j_coord_offsetC  = DIM*jnrC;
 472             j_coord_offsetD  = DIM*jnrD;
 473
 474             /* load j atom coordinates */
 475             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 476                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 477                                                  &jx0,&jy0,&jz0);
 478
 479             /* Calculate displacement vector */
 480             dx00             = _mm256_sub_pd(ix0,jx0);
 481             dy00             = _mm256_sub_pd(iy0,jy0);
 482             dz00             = _mm256_sub_pd(iz0,jz0);
 483
 484             /* Calculate squared distance and things based on it */
 485             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 486
 487             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 488
 489             /* Load parameters for j particles */
 490             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 491                                                                  charge+jnrC+0,charge+jnrD+0);
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             r00              = _mm256_mul_pd(rsq00,rinv00);
 498
 499             /* Compute parameters for interactions between i and j atoms */
 500             qq00             = _mm256_mul_pd(iq0,jq0);
 501
 502             /* Calculate table index by multiplying r with table scale and truncate to integer */
 503             rt               = _mm256_mul_pd(r00,vftabscale);
 504             vfitab           = _mm256_cvttpd_epi32(rt);
 505             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 506             vfitab           = _mm_slli_epi32(vfitab,2);
 507
 508             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 509             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 510             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 511             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 512             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 513             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 514             Heps             = _mm256_mul_pd(vfeps,H);
 515             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 516             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 517             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 518
 519             fscal            = felec;
 520
 521             /* Calculate temporary vectorial force */
 522             tx               = _mm256_mul_pd(fscal,dx00);
 523             ty               = _mm256_mul_pd(fscal,dy00);
 524             tz               = _mm256_mul_pd(fscal,dz00);
 525
 526             /* Update vectorial force */
 527             fix0             = _mm256_add_pd(fix0,tx);
 528             fiy0             = _mm256_add_pd(fiy0,ty);
 529             fiz0             = _mm256_add_pd(fiz0,tz);
 530
 531             fjptrA             = f+j_coord_offsetA;
 532             fjptrB             = f+j_coord_offsetB;
 533             fjptrC             = f+j_coord_offsetC;
 534             fjptrD             = f+j_coord_offsetD;
 535             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 536
 537             /* Inner loop uses 39 flops */
 538         }
 539
 540         if(jidx<j_index_end)
 541         {
 542
 543             /* Get j neighbor index, and coordinate index */
 544             jnrlistA         = jjnr[jidx];
 545             jnrlistB         = jjnr[jidx+1];
 546             jnrlistC         = jjnr[jidx+2];
 547             jnrlistD         = jjnr[jidx+3];
 548             /* Sign of each element will be negative for non-real atoms.
 549              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 550              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 551              */
 552             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 553
 554             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 555             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 556             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 557
 558             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 559             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 560             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 561             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 562             j_coord_offsetA  = DIM*jnrA;
 563             j_coord_offsetB  = DIM*jnrB;
 564             j_coord_offsetC  = DIM*jnrC;
 565             j_coord_offsetD  = DIM*jnrD;
 566
 567             /* load j atom coordinates */
 568             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 569                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 570                                                  &jx0,&jy0,&jz0);
 571
 572             /* Calculate displacement vector */
 573             dx00             = _mm256_sub_pd(ix0,jx0);
 574             dy00             = _mm256_sub_pd(iy0,jy0);
 575             dz00             = _mm256_sub_pd(iz0,jz0);
 576
 577             /* Calculate squared distance and things based on it */
 578             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 579
 580             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 581
 582             /* Load parameters for j particles */
 583             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 584                                                                  charge+jnrC+0,charge+jnrD+0);
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             r00              = _mm256_mul_pd(rsq00,rinv00);
 591             r00              = _mm256_andnot_pd(dummy_mask,r00);
 592
 593             /* Compute parameters for interactions between i and j atoms */
 594             qq00             = _mm256_mul_pd(iq0,jq0);
 595
 596             /* Calculate table index by multiplying r with table scale and truncate to integer */
 597             rt               = _mm256_mul_pd(r00,vftabscale);
 598             vfitab           = _mm256_cvttpd_epi32(rt);
 599             vfeps            = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
 600             vfitab           = _mm_slli_epi32(vfitab,2);
 601
 602             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 603             Y                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 604             F                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 605             G                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
 606             H                = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
 607             GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
 608             Heps             = _mm256_mul_pd(vfeps,H);
 609             Fp               = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
 610             FF               = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
 611             felec            = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
 612
 613             fscal            = felec;
 614
 615             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 616
 617             /* Calculate temporary vectorial force */
 618             tx               = _mm256_mul_pd(fscal,dx00);
 619             ty               = _mm256_mul_pd(fscal,dy00);
 620             tz               = _mm256_mul_pd(fscal,dz00);
 621
 622             /* Update vectorial force */
 623             fix0             = _mm256_add_pd(fix0,tx);
 624             fiy0             = _mm256_add_pd(fiy0,ty);
 625             fiz0             = _mm256_add_pd(fiz0,tz);
 626
 627             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 628             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 629             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 630             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 631             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 632
 633             /* Inner loop uses 40 flops */
 634         }
 635
 636         /* End of innermost loop */
 637
 638         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 639                                                  f+i_coord_offset,fshift+i_shift_offset);
 640
 641         /* Increment number of inner iterations */
 642         inneriter                  += j_index_end - j_index_start;
 643
 644         /* Outer loop uses 7 flops */
 645     }
 646
 647     /* Increment number of outer iterations */
 648     outeriter        += nri;
 649
 650     /* Update outer/inner flops */
 651
 652     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 653 }