src/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecCSTab_VdwNone_GeomP1P1_avx_256_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_256_single 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_single.h"
  34 #include "kernelutil_x86_avx_256_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_VF_avx_256_single
  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_single
  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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
  62     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  63     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  64     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  65     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  66     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  67     real             rcutoff_scalar;
  68     real             *shiftvec,*fshift,*x,*f;
  69     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  70     real             scratch[4*DIM];
  71     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  72     real *           vdwioffsetptr0;
  73     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  74     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  75     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  77     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     __m256i          vfitab;
  80     __m128i          vfitab_lo,vfitab_hi;
  81     __m128i          ifour       = _mm_set1_epi32(4);
  82     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  83     real             *vftab;
  84     __m256           dummy_mask,cutoff_mask;
  85     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  86     __m256           one     = _mm256_set1_ps(1.0);
  87     __m256           two     = _mm256_set1_ps(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     facel            = _mm256_set1_ps(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101
 102     vftab            = kernel_data->table_elec->data;
 103     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 104
 105     /* Avoid stupid compiler warnings */
 106     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 107     j_coord_offsetA = 0;
 108     j_coord_offsetB = 0;
 109     j_coord_offsetC = 0;
 110     j_coord_offsetD = 0;
 111     j_coord_offsetE = 0;
 112     j_coord_offsetF = 0;
 113     j_coord_offsetG = 0;
 114     j_coord_offsetH = 0;
 115
 116     outeriter        = 0;
 117     inneriter        = 0;
 118
 119     for(iidx=0;iidx<4*DIM;iidx++)
 120     {
 121         scratch[iidx] = 0.0;
 122     }
 123
 124     /* Start outer loop over neighborlists */
 125     for(iidx=0; iidx<nri; iidx++)
 126     {
 127         /* Load shift vector for this list */
 128         i_shift_offset   = DIM*shiftidx[iidx];
 129
 130         /* Load limits for loop over neighbors */
 131         j_index_start    = jindex[iidx];
 132         j_index_end      = jindex[iidx+1];
 133
 134         /* Get outer coordinate index */
 135         inr              = iinr[iidx];
 136         i_coord_offset   = DIM*inr;
 137
 138         /* Load i particle coords and add shift vector */
 139         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 140
 141         fix0             = _mm256_setzero_ps();
 142         fiy0             = _mm256_setzero_ps();
 143         fiz0             = _mm256_setzero_ps();
 144
 145         /* Load parameters for i particles */
 146         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 147
 148         /* Reset potential sums */
 149         velecsum         = _mm256_setzero_ps();
 150
 151         /* Start inner kernel loop */
 152         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 153         {
 154
 155             /* Get j neighbor index, and coordinate index */
 156             jnrA             = jjnr[jidx];
 157             jnrB             = jjnr[jidx+1];
 158             jnrC             = jjnr[jidx+2];
 159             jnrD             = jjnr[jidx+3];
 160             jnrE             = jjnr[jidx+4];
 161             jnrF             = jjnr[jidx+5];
 162             jnrG             = jjnr[jidx+6];
 163             jnrH             = jjnr[jidx+7];
 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             j_coord_offsetE  = DIM*jnrE;
 169             j_coord_offsetF  = DIM*jnrF;
 170             j_coord_offsetG  = DIM*jnrG;
 171             j_coord_offsetH  = DIM*jnrH;
 172
 173             /* load j atom coordinates */
 174             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 175                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 176                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 177                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 178                                                  &jx0,&jy0,&jz0);
 179
 180             /* Calculate displacement vector */
 181             dx00             = _mm256_sub_ps(ix0,jx0);
 182             dy00             = _mm256_sub_ps(iy0,jy0);
 183             dz00             = _mm256_sub_ps(iz0,jz0);
 184
 185             /* Calculate squared distance and things based on it */
 186             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 187
 188             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 189
 190             /* Load parameters for j particles */
 191             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 192                                                                  charge+jnrC+0,charge+jnrD+0,
 193                                                                  charge+jnrE+0,charge+jnrF+0,
 194                                                                  charge+jnrG+0,charge+jnrH+0);
 195
 196             /**************************
 197              * CALCULATE INTERACTIONS *
 198              **************************/
 199
 200             r00              = _mm256_mul_ps(rsq00,rinv00);
 201
 202             /* Compute parameters for interactions between i and j atoms */
 203             qq00             = _mm256_mul_ps(iq0,jq0);
 204
 205             /* Calculate table index by multiplying r with table scale and truncate to integer */
 206             rt               = _mm256_mul_ps(r00,vftabscale);
 207             vfitab           = _mm256_cvttps_epi32(rt);
 208             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 209             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 210             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 211             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 212             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 213             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 214
 215             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 216             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 217                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 218             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 219                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 220             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 221                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 222             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 223                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 224             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 225             Heps             = _mm256_mul_ps(vfeps,H);
 226             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 227             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 228             velec            = _mm256_mul_ps(qq00,VV);
 229             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 230             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 231
 232             /* Update potential sum for this i atom from the interaction with this j atom. */
 233             velecsum         = _mm256_add_ps(velecsum,velec);
 234
 235             fscal            = felec;
 236
 237             /* Calculate temporary vectorial force */
 238             tx               = _mm256_mul_ps(fscal,dx00);
 239             ty               = _mm256_mul_ps(fscal,dy00);
 240             tz               = _mm256_mul_ps(fscal,dz00);
 241
 242             /* Update vectorial force */
 243             fix0             = _mm256_add_ps(fix0,tx);
 244             fiy0             = _mm256_add_ps(fiy0,ty);
 245             fiz0             = _mm256_add_ps(fiz0,tz);
 246
 247             fjptrA             = f+j_coord_offsetA;
 248             fjptrB             = f+j_coord_offsetB;
 249             fjptrC             = f+j_coord_offsetC;
 250             fjptrD             = f+j_coord_offsetD;
 251             fjptrE             = f+j_coord_offsetE;
 252             fjptrF             = f+j_coord_offsetF;
 253             fjptrG             = f+j_coord_offsetG;
 254             fjptrH             = f+j_coord_offsetH;
 255             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 256
 257             /* Inner loop uses 43 flops */
 258         }
 259
 260         if(jidx<j_index_end)
 261         {
 262
 263             /* Get j neighbor index, and coordinate index */
 264             jnrlistA         = jjnr[jidx];
 265             jnrlistB         = jjnr[jidx+1];
 266             jnrlistC         = jjnr[jidx+2];
 267             jnrlistD         = jjnr[jidx+3];
 268             jnrlistE         = jjnr[jidx+4];
 269             jnrlistF         = jjnr[jidx+5];
 270             jnrlistG         = jjnr[jidx+6];
 271             jnrlistH         = jjnr[jidx+7];
 272             /* Sign of each element will be negative for non-real atoms.
 273              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 274              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 275              */
 276             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 277                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 278
 279             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 280             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 281             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 282             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 283             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 284             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 285             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 286             jnrH       = (jnrlistH>=0) ? jnrlistH : 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             j_coord_offsetE  = DIM*jnrE;
 292             j_coord_offsetF  = DIM*jnrF;
 293             j_coord_offsetG  = DIM*jnrG;
 294             j_coord_offsetH  = DIM*jnrH;
 295
 296             /* load j atom coordinates */
 297             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 298                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 299                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 300                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 301                                                  &jx0,&jy0,&jz0);
 302
 303             /* Calculate displacement vector */
 304             dx00             = _mm256_sub_ps(ix0,jx0);
 305             dy00             = _mm256_sub_ps(iy0,jy0);
 306             dz00             = _mm256_sub_ps(iz0,jz0);
 307
 308             /* Calculate squared distance and things based on it */
 309             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 310
 311             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 312
 313             /* Load parameters for j particles */
 314             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 315                                                                  charge+jnrC+0,charge+jnrD+0,
 316                                                                  charge+jnrE+0,charge+jnrF+0,
 317                                                                  charge+jnrG+0,charge+jnrH+0);
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r00              = _mm256_mul_ps(rsq00,rinv00);
 324             r00              = _mm256_andnot_ps(dummy_mask,r00);
 325
 326             /* Compute parameters for interactions between i and j atoms */
 327             qq00             = _mm256_mul_ps(iq0,jq0);
 328
 329             /* Calculate table index by multiplying r with table scale and truncate to integer */
 330             rt               = _mm256_mul_ps(r00,vftabscale);
 331             vfitab           = _mm256_cvttps_epi32(rt);
 332             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 333             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 334             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 335             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 336             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 337             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 338
 339             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 340             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 341                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 342             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 343                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 344             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 345                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 346             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 347                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 348             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 349             Heps             = _mm256_mul_ps(vfeps,H);
 350             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 351             VV               = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
 352             velec            = _mm256_mul_ps(qq00,VV);
 353             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 354             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velec            = _mm256_andnot_ps(dummy_mask,velec);
 358             velecsum         = _mm256_add_ps(velecsum,velec);
 359
 360             fscal            = felec;
 361
 362             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 363
 364             /* Calculate temporary vectorial force */
 365             tx               = _mm256_mul_ps(fscal,dx00);
 366             ty               = _mm256_mul_ps(fscal,dy00);
 367             tz               = _mm256_mul_ps(fscal,dz00);
 368
 369             /* Update vectorial force */
 370             fix0             = _mm256_add_ps(fix0,tx);
 371             fiy0             = _mm256_add_ps(fiy0,ty);
 372             fiz0             = _mm256_add_ps(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             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 379             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 380             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 381             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 382             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 383
 384             /* Inner loop uses 44 flops */
 385         }
 386
 387         /* End of innermost loop */
 388
 389         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 390                                                  f+i_coord_offset,fshift+i_shift_offset);
 391
 392         ggid                        = gid[iidx];
 393         /* Update potential energies */
 394         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 395
 396         /* Increment number of inner iterations */
 397         inneriter                  += j_index_end - j_index_start;
 398
 399         /* Outer loop uses 8 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_VF,outeriter*8 + inneriter*44);
 408 }
 409 /*
 410  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 411  * Electrostatics interaction: CubicSplineTable
 412  * VdW interaction:            None
 413  * Geometry:                   Particle-Particle
 414  * Calculate force/pot:        Force
 415  */
 416 void
 417 nb_kernel_ElecCSTab_VdwNone_GeomP1P1_F_avx_256_single
 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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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              jnrE,jnrF,jnrG,jnrH;
 435     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 436     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 437     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 438     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 439     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 440     real             rcutoff_scalar;
 441     real             *shiftvec,*fshift,*x,*f;
 442     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 443     real             scratch[4*DIM];
 444     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 445     real *           vdwioffsetptr0;
 446     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 447     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 448     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 449     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 450     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 451     real             *charge;
 452     __m256i          vfitab;
 453     __m128i          vfitab_lo,vfitab_hi;
 454     __m128i          ifour       = _mm_set1_epi32(4);
 455     __m256           rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 456     real             *vftab;
 457     __m256           dummy_mask,cutoff_mask;
 458     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 459     __m256           one     = _mm256_set1_ps(1.0);
 460     __m256           two     = _mm256_set1_ps(2.0);
 461     x                = xx[0];
 462     f                = ff[0];
 463
 464     nri              = nlist->nri;
 465     iinr             = nlist->iinr;
 466     jindex           = nlist->jindex;
 467     jjnr             = nlist->jjnr;
 468     shiftidx         = nlist->shift;
 469     gid              = nlist->gid;
 470     shiftvec         = fr->shift_vec[0];
 471     fshift           = fr->fshift[0];
 472     facel            = _mm256_set1_ps(fr->epsfac);
 473     charge           = mdatoms->chargeA;
 474
 475     vftab            = kernel_data->table_elec->data;
 476     vftabscale       = _mm256_set1_ps(kernel_data->table_elec->scale);
 477
 478     /* Avoid stupid compiler warnings */
 479     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 480     j_coord_offsetA = 0;
 481     j_coord_offsetB = 0;
 482     j_coord_offsetC = 0;
 483     j_coord_offsetD = 0;
 484     j_coord_offsetE = 0;
 485     j_coord_offsetF = 0;
 486     j_coord_offsetG = 0;
 487     j_coord_offsetH = 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_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 513
 514         fix0             = _mm256_setzero_ps();
 515         fiy0             = _mm256_setzero_ps();
 516         fiz0             = _mm256_setzero_ps();
 517
 518         /* Load parameters for i particles */
 519         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 520
 521         /* Start inner kernel loop */
 522         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 523         {
 524
 525             /* Get j neighbor index, and coordinate index */
 526             jnrA             = jjnr[jidx];
 527             jnrB             = jjnr[jidx+1];
 528             jnrC             = jjnr[jidx+2];
 529             jnrD             = jjnr[jidx+3];
 530             jnrE             = jjnr[jidx+4];
 531             jnrF             = jjnr[jidx+5];
 532             jnrG             = jjnr[jidx+6];
 533             jnrH             = jjnr[jidx+7];
 534             j_coord_offsetA  = DIM*jnrA;
 535             j_coord_offsetB  = DIM*jnrB;
 536             j_coord_offsetC  = DIM*jnrC;
 537             j_coord_offsetD  = DIM*jnrD;
 538             j_coord_offsetE  = DIM*jnrE;
 539             j_coord_offsetF  = DIM*jnrF;
 540             j_coord_offsetG  = DIM*jnrG;
 541             j_coord_offsetH  = DIM*jnrH;
 542
 543             /* load j atom coordinates */
 544             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 545                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 546                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 547                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 548                                                  &jx0,&jy0,&jz0);
 549
 550             /* Calculate displacement vector */
 551             dx00             = _mm256_sub_ps(ix0,jx0);
 552             dy00             = _mm256_sub_ps(iy0,jy0);
 553             dz00             = _mm256_sub_ps(iz0,jz0);
 554
 555             /* Calculate squared distance and things based on it */
 556             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 557
 558             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 559
 560             /* Load parameters for j particles */
 561             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 562                                                                  charge+jnrC+0,charge+jnrD+0,
 563                                                                  charge+jnrE+0,charge+jnrF+0,
 564                                                                  charge+jnrG+0,charge+jnrH+0);
 565
 566             /**************************
 567              * CALCULATE INTERACTIONS *
 568              **************************/
 569
 570             r00              = _mm256_mul_ps(rsq00,rinv00);
 571
 572             /* Compute parameters for interactions between i and j atoms */
 573             qq00             = _mm256_mul_ps(iq0,jq0);
 574
 575             /* Calculate table index by multiplying r with table scale and truncate to integer */
 576             rt               = _mm256_mul_ps(r00,vftabscale);
 577             vfitab           = _mm256_cvttps_epi32(rt);
 578             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 579             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 580             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 581             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 582             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 583             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 584
 585             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 586             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 587                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 588             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 589                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 590             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 591                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 592             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 593                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 594             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 595             Heps             = _mm256_mul_ps(vfeps,H);
 596             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 597             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 598             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 599
 600             fscal            = felec;
 601
 602             /* Calculate temporary vectorial force */
 603             tx               = _mm256_mul_ps(fscal,dx00);
 604             ty               = _mm256_mul_ps(fscal,dy00);
 605             tz               = _mm256_mul_ps(fscal,dz00);
 606
 607             /* Update vectorial force */
 608             fix0             = _mm256_add_ps(fix0,tx);
 609             fiy0             = _mm256_add_ps(fiy0,ty);
 610             fiz0             = _mm256_add_ps(fiz0,tz);
 611
 612             fjptrA             = f+j_coord_offsetA;
 613             fjptrB             = f+j_coord_offsetB;
 614             fjptrC             = f+j_coord_offsetC;
 615             fjptrD             = f+j_coord_offsetD;
 616             fjptrE             = f+j_coord_offsetE;
 617             fjptrF             = f+j_coord_offsetF;
 618             fjptrG             = f+j_coord_offsetG;
 619             fjptrH             = f+j_coord_offsetH;
 620             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 621
 622             /* Inner loop uses 39 flops */
 623         }
 624
 625         if(jidx<j_index_end)
 626         {
 627
 628             /* Get j neighbor index, and coordinate index */
 629             jnrlistA         = jjnr[jidx];
 630             jnrlistB         = jjnr[jidx+1];
 631             jnrlistC         = jjnr[jidx+2];
 632             jnrlistD         = jjnr[jidx+3];
 633             jnrlistE         = jjnr[jidx+4];
 634             jnrlistF         = jjnr[jidx+5];
 635             jnrlistG         = jjnr[jidx+6];
 636             jnrlistH         = jjnr[jidx+7];
 637             /* Sign of each element will be negative for non-real atoms.
 638              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 639              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 640              */
 641             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 642                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 643
 644             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 645             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 646             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 647             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 648             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 649             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 650             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 651             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 652             j_coord_offsetA  = DIM*jnrA;
 653             j_coord_offsetB  = DIM*jnrB;
 654             j_coord_offsetC  = DIM*jnrC;
 655             j_coord_offsetD  = DIM*jnrD;
 656             j_coord_offsetE  = DIM*jnrE;
 657             j_coord_offsetF  = DIM*jnrF;
 658             j_coord_offsetG  = DIM*jnrG;
 659             j_coord_offsetH  = DIM*jnrH;
 660
 661             /* load j atom coordinates */
 662             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 663                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 664                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 665                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 666                                                  &jx0,&jy0,&jz0);
 667
 668             /* Calculate displacement vector */
 669             dx00             = _mm256_sub_ps(ix0,jx0);
 670             dy00             = _mm256_sub_ps(iy0,jy0);
 671             dz00             = _mm256_sub_ps(iz0,jz0);
 672
 673             /* Calculate squared distance and things based on it */
 674             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 675
 676             rinv00           = gmx_mm256_invsqrt_ps(rsq00);
 677
 678             /* Load parameters for j particles */
 679             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 680                                                                  charge+jnrC+0,charge+jnrD+0,
 681                                                                  charge+jnrE+0,charge+jnrF+0,
 682                                                                  charge+jnrG+0,charge+jnrH+0);
 683
 684             /**************************
 685              * CALCULATE INTERACTIONS *
 686              **************************/
 687
 688             r00              = _mm256_mul_ps(rsq00,rinv00);
 689             r00              = _mm256_andnot_ps(dummy_mask,r00);
 690
 691             /* Compute parameters for interactions between i and j atoms */
 692             qq00             = _mm256_mul_ps(iq0,jq0);
 693
 694             /* Calculate table index by multiplying r with table scale and truncate to integer */
 695             rt               = _mm256_mul_ps(r00,vftabscale);
 696             vfitab           = _mm256_cvttps_epi32(rt);
 697             vfeps            = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
 698             /*         AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
 699             vfitab_lo        = _mm256_extractf128_si256(vfitab,0x0);
 700             vfitab_hi        = _mm256_extractf128_si256(vfitab,0x1);
 701             vfitab_lo        = _mm_slli_epi32(vfitab_lo,2);
 702             vfitab_hi        = _mm_slli_epi32(vfitab_hi,2);
 703
 704             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 705             Y                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
 706                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
 707             F                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
 708                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
 709             G                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
 710                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
 711             H                = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
 712                                                   _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
 713             GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
 714             Heps             = _mm256_mul_ps(vfeps,H);
 715             Fp               = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
 716             FF               = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
 717             felec            = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_mul_ps(qq00,FF),_mm256_mul_ps(vftabscale,rinv00)));
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm256_mul_ps(fscal,dx00);
 725             ty               = _mm256_mul_ps(fscal,dy00);
 726             tz               = _mm256_mul_ps(fscal,dz00);
 727
 728             /* Update vectorial force */
 729             fix0             = _mm256_add_ps(fix0,tx);
 730             fiy0             = _mm256_add_ps(fiy0,ty);
 731             fiz0             = _mm256_add_ps(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             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 738             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 739             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 740             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 741             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 742
 743             /* Inner loop uses 40 flops */
 744         }
 745
 746         /* End of innermost loop */
 747
 748         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 749                                                  f+i_coord_offset,fshift+i_shift_offset);
 750
 751         /* Increment number of inner iterations */
 752         inneriter                  += j_index_end - j_index_start;
 753
 754         /* Outer loop uses 7 flops */
 755     }
 756
 757     /* Increment number of outer iterations */
 758     outeriter        += nri;
 759
 760     /* Update outer/inner flops */
 761
 762     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*40);
 763 }